AuroraMiddleware/SPIRV-Tools
1
0
Fork 0
mirror of https://github.com/KhronosGroup/SPIRV-Tools synced 2024-11-29 14:31:04 +00:00
Code Issues Packages Projects Releases Wiki Activity
5b371918b9
SPIRV-Tools/test/opt/local_ssa_elim_test.cpp

4266 lines
133 KiB
C++
Raw Normal View History

Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// Copyright (c) 2017 Valve Corporation
// Copyright (c) 2017 LunarG Inc.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
Cleanup includes. (#1795) This Cl cleans up the include paths to be relative to the top level directory. Various include-what-you-use fixes have been added.
2018-08-03 19:06:09 +00:00
#include <memory>
#include <string>
#include "test/opt/pass_fixture.h"
#include "test/opt/pass_utils.h"
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
Move tests into namespaces (#1689) This CL moves the test into namespaces based on their directories.
2018-07-11 13:24:49 +00:00
namespace spvtools {
namespace opt {
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
namespace {
using LocalSSAElimTest = PassTest<::testing::Test>;
TEST_F(LocalSSAElimTest, ForLoop) {
// #version 140
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// in vec4 BC;
// out float fo;
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// void main()
// {
// float f = 0.0;
// for (int i=0; i<4; i++) {
// f = f + BC[i];
// }
// fo = f;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpName %main "main"
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpName %f "f"
OpName %i "i"
OpName %BC "BC"
OpName %fo "fo"
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%void = OpTypeVoid
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
)";
const std::string before =
R"(%main = OpFunction %void None %8
%22 = OpLabel
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
OpBranch %23
%23 = OpLabel
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%27 = OpLoad %int %i
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%30 = OpLoad %float %f
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %25
%25 = OpLabel
%35 = OpLoad %int %i
%36 = OpIAdd %int %35 %int_1
OpStore %i %36
OpBranch %23
%24 = OpLabel
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
const std::string after =
R"(%main = OpFunction %void None %8
%22 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %23
%23 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%39 = OpPhi %float %float_0 %22 %34 %25
%38 = OpPhi %int %int_0 %22 %36 %25
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%28 = OpSLessThan %bool %38 %int_4
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranchConditional %28 %29 %24
%29 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%32 = OpAccessChain %_ptr_Input_float %BC %38
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%33 = OpLoad %float %32
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%34 = OpFAdd %float %39 %33
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %f %34
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %25
%25 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%36 = OpIAdd %int %38 %int_1
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %i %36
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %23
%24 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
OpStore %fo %39
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + before, predefs + after, true,
true);
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
}
Fix SSA rewrite for nested loops. From the test case, the slice of the CFG that is interesting for the bug is 25 | v 30 | v 31<-+ | | v | 34--+ 1. In block 25, we have a Phi candidate for %f with arguments %47 = Phi[%float_0, %0]. This merges %float_0 and a yet unknown argument from the external loop backedge. 2. We are now processing block 34: i. The load %35 = OpLoad %f triggers a Phi candidate to be placed in block 31. ii. The Phi candidate %50 = Phi needs two arguments. The one coming from block 30 is %47. But the one coming from block 34 (which we are now processing and have marked sealed), finds %50 itself as the reaching def for %f. 3. This wrongfully marks %50 as a copy-of Phi, which ultimately makes both %47 and %50 copy-of Phis that get eliminated.
2018-04-05 00:15:48 +00:00
TEST_F(LocalSSAElimTest, NestedForLoop) {
// #version 450
//
// layout (location=0) in mat4 BC;
// layout (location=0) out float fo;
//
// void main()
// {
// float f = 0.0;
// for (int i=0; i<4; i++)
// for (int j=0; j<4; j++)
// f = f + BC[i][j];
// fo = f;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 450
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %j "j"
OpName %BC "BC"
OpName %fo "fo"
OpDecorate %BC Location 0
OpDecorate %fo Location 0
%void = OpTypeVoid
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%mat4v4float = OpTypeMatrix %v4float 4
%_ptr_Input_mat4v4float = OpTypePointer Input %mat4v4float
%BC = OpVariable %_ptr_Input_mat4v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
)";
const std::string before =
Validate nested constructs (#3068) * Validate that if a construct contains a header and it's merge is reachable, the construct also contains the merge * updated block merging to not merge into the continue * update inlining to mark the original block of a single block loop as the continue * updated some tests * remove dead code * rename kBlockTypeHeader to kBlockTypeSelection for clarity
2019-11-27 21:45:57 +00:00
R"(
; CHECK: = OpFunction
; CHECK-NEXT: [[entry:%\w+]] = OpLabel
; CHECK: [[outer_header:%\w+]] = OpLabel
; CHECK-NEXT: [[outer_f:%\w+]] = OpPhi %float %float_0 [[entry]] [[inner_f:%\w+]] [[outer_be:%\w+]]
; CHECK-NEXT: [[i:%\w+]] = OpPhi %int %int_0 [[entry]] [[i_next:%\w+]] [[outer_be]]
; CHECK-NEXT: OpSLessThan {{%\w+}} [[i]]
; CHECK: [[inner_pre_header:%\w+]] = OpLabel
; CHECK: [[inner_header:%\w+]] = OpLabel
; CHECK-NEXT: [[inner_f]] = OpPhi %float [[outer_f]] [[inner_pre_header]] [[f_next:%\w+]] [[inner_be:%\w+]]
; CHECK-NEXT: [[j:%\w+]] = OpPhi %int %int_0 [[inner_pre_header]] [[j_next:%\w+]] [[inner_be]]
; CHECK: [[inner_be]] = OpLabel
; CHECK: [[f_next]] = OpFAdd %float [[inner_f]]
; CHECK: [[j_next]] = OpIAdd %int [[j]] %int_1
; CHECK: [[outer_be]] = OpLabel
; CHECK: [[i_next]] = OpIAdd
; CHECK: OpStore %fo [[outer_f]]
%main = OpFunction %void None %9
Fix SSA rewrite for nested loops. From the test case, the slice of the CFG that is interesting for the bug is 25 | v 30 | v 31<-+ | | v | 34--+ 1. In block 25, we have a Phi candidate for %f with arguments %47 = Phi[%float_0, %0]. This merges %float_0 and a yet unknown argument from the external loop backedge. 2. We are now processing block 34: i. The load %35 = OpLoad %f triggers a Phi candidate to be placed in block 31. ii. The Phi candidate %50 = Phi needs two arguments. The one coming from block 30 is %47. But the one coming from block 34 (which we are now processing and have marked sealed), finds %50 itself as the reaching def for %f. 3. This wrongfully marks %50 as a copy-of Phi, which ultimately makes both %47 and %50 copy-of Phis that get eliminated.
2018-04-05 00:15:48 +00:00
%24 = OpLabel
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
%j = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
OpBranch %25
%25 = OpLabel
%26 = OpLoad %int %i
%27 = OpSLessThan %bool %26 %int_4
OpLoopMerge %28 %29 None
OpBranchConditional %27 %30 %28
%30 = OpLabel
OpStore %j %int_0
OpBranch %31
%31 = OpLabel
%32 = OpLoad %int %j
%33 = OpSLessThan %bool %32 %int_4
Validate nested constructs (#3068) * Validate that if a construct contains a header and it's merge is reachable, the construct also contains the merge * updated block merging to not merge into the continue * update inlining to mark the original block of a single block loop as the continue * updated some tests * remove dead code * rename kBlockTypeHeader to kBlockTypeSelection for clarity
2019-11-27 21:45:57 +00:00
OpLoopMerge %50 %34 None
OpBranchConditional %33 %34 %50
Fix SSA rewrite for nested loops. From the test case, the slice of the CFG that is interesting for the bug is 25 | v 30 | v 31<-+ | | v | 34--+ 1. In block 25, we have a Phi candidate for %f with arguments %47 = Phi[%float_0, %0]. This merges %float_0 and a yet unknown argument from the external loop backedge. 2. We are now processing block 34: i. The load %35 = OpLoad %f triggers a Phi candidate to be placed in block 31. ii. The Phi candidate %50 = Phi needs two arguments. The one coming from block 30 is %47. But the one coming from block 34 (which we are now processing and have marked sealed), finds %50 itself as the reaching def for %f. 3. This wrongfully marks %50 as a copy-of Phi, which ultimately makes both %47 and %50 copy-of Phis that get eliminated.
2018-04-05 00:15:48 +00:00
%34 = OpLabel
%35 = OpLoad %float %f
%36 = OpLoad %int %i
%37 = OpLoad %int %j
%38 = OpAccessChain %_ptr_Input_float %BC %36 %37
%39 = OpLoad %float %38
%40 = OpFAdd %float %35 %39
OpStore %f %40
%41 = OpLoad %int %j
%42 = OpIAdd %int %41 %int_1
OpStore %j %42
OpBranch %31
Validate nested constructs (#3068) * Validate that if a construct contains a header and it's merge is reachable, the construct also contains the merge * updated block merging to not merge into the continue * update inlining to mark the original block of a single block loop as the continue * updated some tests * remove dead code * rename kBlockTypeHeader to kBlockTypeSelection for clarity
2019-11-27 21:45:57 +00:00
%50 = OpLabel
OpBranch %29
Fix SSA rewrite for nested loops. From the test case, the slice of the CFG that is interesting for the bug is 25 | v 30 | v 31<-+ | | v | 34--+ 1. In block 25, we have a Phi candidate for %f with arguments %47 = Phi[%float_0, %0]. This merges %float_0 and a yet unknown argument from the external loop backedge. 2. We are now processing block 34: i. The load %35 = OpLoad %f triggers a Phi candidate to be placed in block 31. ii. The Phi candidate %50 = Phi needs two arguments. The one coming from block 30 is %47. But the one coming from block 34 (which we are now processing and have marked sealed), finds %50 itself as the reaching def for %f. 3. This wrongfully marks %50 as a copy-of Phi, which ultimately makes both %47 and %50 copy-of Phis that get eliminated.
2018-04-05 00:15:48 +00:00
%29 = OpLabel
%43 = OpLoad %int %i
%44 = OpIAdd %int %43 %int_1
OpStore %i %44
OpBranch %25
%28 = OpLabel
%45 = OpLoad %float %f
OpStore %fo %45
OpReturn
OpFunctionEnd
)";
Validate nested constructs (#3068) * Validate that if a construct contains a header and it's merge is reachable, the construct also contains the merge * updated block merging to not merge into the continue * update inlining to mark the original block of a single block loop as the continue * updated some tests * remove dead code * rename kBlockTypeHeader to kBlockTypeSelection for clarity
2019-11-27 21:45:57 +00:00
SinglePassRunAndMatch<SSARewritePass>(predefs + before, true);
Fix SSA rewrite for nested loops. From the test case, the slice of the CFG that is interesting for the bug is 25 | v 30 | v 31<-+ | | v | 34--+ 1. In block 25, we have a Phi candidate for %f with arguments %47 = Phi[%float_0, %0]. This merges %float_0 and a yet unknown argument from the external loop backedge. 2. We are now processing block 34: i. The load %35 = OpLoad %f triggers a Phi candidate to be placed in block 31. ii. The Phi candidate %50 = Phi needs two arguments. The one coming from block 30 is %47. But the one coming from block 34 (which we are now processing and have marked sealed), finds %50 itself as the reaching def for %f. 3. This wrongfully marks %50 as a copy-of Phi, which ultimately makes both %47 and %50 copy-of Phis that get eliminated.
2018-04-05 00:15:48 +00:00
}
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
TEST_F(LocalSSAElimTest, ForLoopWithContinue) {
// #version 140
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// in vec4 BC;
// out float fo;
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// void main()
// {
// float f = 0.0;
// for (int i=0; i<4; i++) {
// float t = BC[i];
// if (t < 0.0)
// continue;
// f = f + t;
// }
// fo = f;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
)";
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
const std::string names =
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
R"(OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %t "t"
OpName %BC "BC"
OpName %fo "fo"
)";
const std::string predefs2 =
R"(%void = OpTypeVoid
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
)";
const std::string before =
R"(%main = OpFunction %void None %9
%23 = OpLabel
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
OpStore %f %float_0
OpStore %i %int_0
OpBranch %24
%24 = OpLabel
OpLoopMerge %25 %26 None
OpBranch %27
%27 = OpLabel
%28 = OpLoad %int %i
%29 = OpSLessThan %bool %28 %int_4
OpBranchConditional %29 %30 %25
%30 = OpLabel
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
OpStore %t %33
%34 = OpLoad %float %t
%35 = OpFOrdLessThan %bool %34 %float_0
OpSelectionMerge %36 None
OpBranchConditional %35 %37 %36
%37 = OpLabel
OpBranch %26
%36 = OpLabel
%38 = OpLoad %float %f
%39 = OpLoad %float %t
%40 = OpFAdd %float %38 %39
OpStore %f %40
OpBranch %26
%26 = OpLabel
%41 = OpLoad %int %i
%42 = OpIAdd %int %41 %int_1
OpStore %i %42
OpBranch %24
%25 = OpLabel
%43 = OpLoad %float %f
OpStore %fo %43
OpReturn
OpFunctionEnd
)";
const std::string after =
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
R"(%main = OpFunction %void None %9
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%23 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
OpStore %f %float_0
OpStore %i %int_0
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %24
%24 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%45 = OpPhi %float %float_0 %23 %47 %26
%44 = OpPhi %int %int_0 %23 %42 %26
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpLoopMerge %25 %26 None
OpBranch %27
%27 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%29 = OpSLessThan %bool %44 %int_4
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranchConditional %29 %30 %25
%30 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%32 = OpAccessChain %_ptr_Input_float %BC %44
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%33 = OpLoad %float %32
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %t %33
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%35 = OpFOrdLessThan %bool %33 %float_0
OpSelectionMerge %36 None
OpBranchConditional %35 %37 %36
%37 = OpLabel
OpBranch %26
%36 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%40 = OpFAdd %float %45 %33
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %f %40
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %26
%26 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%47 = OpPhi %float %45 %37 %40 %36
%42 = OpIAdd %int %44 %int_1
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %i %42
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %24
%25 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
OpStore %fo %45
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + names + predefs2 + before,
predefs + names + predefs2 + after,
true, true);
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
}
TEST_F(LocalSSAElimTest, ForLoopWithBreak) {
// #version 140
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// in vec4 BC;
// out float fo;
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// void main()
// {
// float f = 0.0;
// for (int i=0; i<4; i++) {
// float t = f + BC[i];
// if (t > 1.0)
// break;
// f = t;
// }
// fo = f;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpName %main "main"
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpName %f "f"
OpName %i "i"
OpName %t "t"
OpName %BC "BC"
OpName %fo "fo"
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%void = OpTypeVoid
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%float_1 = OpConstant %float 1
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
)";
const std::string before =
R"(%main = OpFunction %void None %9
%24 = OpLabel
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
OpStore %f %float_0
OpStore %i %int_0
OpBranch %25
%25 = OpLabel
OpLoopMerge %26 %27 None
OpBranch %28
%28 = OpLabel
%29 = OpLoad %int %i
%30 = OpSLessThan %bool %29 %int_4
OpBranchConditional %30 %31 %26
%31 = OpLabel
%32 = OpLoad %float %f
%33 = OpLoad %int %i
%34 = OpAccessChain %_ptr_Input_float %BC %33
%35 = OpLoad %float %34
%36 = OpFAdd %float %32 %35
OpStore %t %36
%37 = OpLoad %float %t
%38 = OpFOrdGreaterThan %bool %37 %float_1
OpSelectionMerge %39 None
OpBranchConditional %38 %40 %39
%40 = OpLabel
OpBranch %26
%39 = OpLabel
%41 = OpLoad %float %t
OpStore %f %41
OpBranch %27
%27 = OpLabel
%42 = OpLoad %int %i
%43 = OpIAdd %int %42 %int_1
OpStore %i %43
OpBranch %25
%26 = OpLabel
%44 = OpLoad %float %f
OpStore %fo %44
OpReturn
OpFunctionEnd
)";
const std::string after =
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
R"(%main = OpFunction %void None %9
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%24 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
OpStore %f %float_0
OpStore %i %int_0
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %25
%25 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%46 = OpPhi %float %float_0 %24 %36 %27
%45 = OpPhi %int %int_0 %24 %43 %27
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpLoopMerge %26 %27 None
OpBranch %28
%28 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%30 = OpSLessThan %bool %45 %int_4
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranchConditional %30 %31 %26
%31 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%34 = OpAccessChain %_ptr_Input_float %BC %45
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%35 = OpLoad %float %34
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%36 = OpFAdd %float %46 %35
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %t %36
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%38 = OpFOrdGreaterThan %bool %36 %float_1
OpSelectionMerge %39 None
OpBranchConditional %38 %40 %39
%40 = OpLabel
OpBranch %26
%39 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %f %36
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %27
%27 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%43 = OpIAdd %int %45 %int_1
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %i %43
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %25
%26 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
OpStore %fo %46
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + before, predefs + after, true,
true);
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
}
TEST_F(LocalSSAElimTest, SwapProblem) {
// #version 140
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// in float fe;
// out float fo;
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// void main()
// {
// float f1 = 0.0;
// float f2 = 1.0;
// int ie = int(fe);
// for (int i=0; i<ie; i++) {
// float t = f1;
// f1 = f2;
// f2 = t;
// }
// fo = f1;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %fe %fo
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpName %main "main"
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpName %f1 "f1"
OpName %f2 "f2"
OpName %ie "ie"
OpName %fe "fe"
OpName %i "i"
OpName %t "t"
OpName %fo "fo"
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%void = OpTypeVoid
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%11 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%float_1 = OpConstant %float 1
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%_ptr_Input_float = OpTypePointer Input %float
%fe = OpVariable %_ptr_Input_float Input
%int_0 = OpConstant %int 0
%bool = OpTypeBool
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
)";
const std::string before =
R"(%main = OpFunction %void None %11
%23 = OpLabel
%f1 = OpVariable %_ptr_Function_float Function
%f2 = OpVariable %_ptr_Function_float Function
%ie = OpVariable %_ptr_Function_int Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
OpStore %f1 %float_0
OpStore %f2 %float_1
%24 = OpLoad %float %fe
%25 = OpConvertFToS %int %24
OpStore %ie %25
OpStore %i %int_0
OpBranch %26
%26 = OpLabel
OpLoopMerge %27 %28 None
OpBranch %29
%29 = OpLabel
%30 = OpLoad %int %i
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
%31 = OpLoad %int %ie
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%32 = OpSLessThan %bool %30 %31
OpBranchConditional %32 %33 %27
%33 = OpLabel
%34 = OpLoad %float %f1
OpStore %t %34
%35 = OpLoad %float %f2
OpStore %f1 %35
%36 = OpLoad %float %t
OpStore %f2 %36
OpBranch %28
%28 = OpLabel
%37 = OpLoad %int %i
%38 = OpIAdd %int %37 %int_1
OpStore %i %38
OpBranch %26
%27 = OpLabel
%39 = OpLoad %float %f1
OpStore %fo %39
OpReturn
OpFunctionEnd
)";
const std::string after =
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
R"(%main = OpFunction %void None %11
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%23 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%f1 = OpVariable %_ptr_Function_float Function
%f2 = OpVariable %_ptr_Function_float Function
%ie = OpVariable %_ptr_Function_int Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
OpStore %f1 %float_0
OpStore %f2 %float_1
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%24 = OpLoad %float %fe
%25 = OpConvertFToS %int %24
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %ie %25
OpStore %i %int_0
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %26
%26 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%43 = OpPhi %float %float_1 %23 %42 %28
%42 = OpPhi %float %float_0 %23 %43 %28
%40 = OpPhi %int %int_0 %23 %38 %28
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpLoopMerge %27 %28 None
OpBranch %29
%29 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%32 = OpSLessThan %bool %40 %25
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranchConditional %32 %33 %27
%33 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
OpStore %t %42
OpStore %f1 %43
OpStore %f2 %42
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %28
%28 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%38 = OpIAdd %int %40 %int_1
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %i %38
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %26
%27 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
OpStore %fo %42
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + before, predefs + after, true,
true);
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
}
TEST_F(LocalSSAElimTest, LostCopyProblem) {
// #version 140
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// in vec4 BC;
// out float fo;
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// void main()
// {
// float f = 0.0;
// float t;
// for (int i=0; i<4; i++) {
// t = f;
// f = f + BC[i];
// if (f > 1.0)
// break;
// }
// fo = t;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpName %main "main"
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpName %f "f"
OpName %i "i"
OpName %t "t"
OpName %BC "BC"
OpName %fo "fo"
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%void = OpTypeVoid
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%float_1 = OpConstant %float 1
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
)";
const std::string before =
R"(%main = OpFunction %void None %9
%24 = OpLabel
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
OpStore %f %float_0
OpStore %i %int_0
OpBranch %25
%25 = OpLabel
OpLoopMerge %26 %27 None
OpBranch %28
%28 = OpLabel
%29 = OpLoad %int %i
%30 = OpSLessThan %bool %29 %int_4
OpBranchConditional %30 %31 %26
%31 = OpLabel
%32 = OpLoad %float %f
OpStore %t %32
%33 = OpLoad %float %f
%34 = OpLoad %int %i
%35 = OpAccessChain %_ptr_Input_float %BC %34
%36 = OpLoad %float %35
%37 = OpFAdd %float %33 %36
OpStore %f %37
%38 = OpLoad %float %f
%39 = OpFOrdGreaterThan %bool %38 %float_1
OpSelectionMerge %40 None
OpBranchConditional %39 %41 %40
%41 = OpLabel
OpBranch %26
%40 = OpLabel
OpBranch %27
%27 = OpLabel
%42 = OpLoad %int %i
%43 = OpIAdd %int %42 %int_1
OpStore %i %43
OpBranch %25
%26 = OpLabel
%44 = OpLoad %float %t
OpStore %fo %44
OpReturn
OpFunctionEnd
)";
const std::string after =
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
R"(%49 = OpUndef %float
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%main = OpFunction %void None %9
%24 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
OpStore %f %float_0
OpStore %i %int_0
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %25
%25 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%46 = OpPhi %float %float_0 %24 %37 %27
%45 = OpPhi %int %int_0 %24 %43 %27
%48 = OpPhi %float %49 %24 %46 %27
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpLoopMerge %26 %27 None
OpBranch %28
%28 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%30 = OpSLessThan %bool %45 %int_4
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranchConditional %30 %31 %26
%31 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
OpStore %t %46
%35 = OpAccessChain %_ptr_Input_float %BC %45
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%36 = OpLoad %float %35
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%37 = OpFAdd %float %46 %36
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %f %37
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%39 = OpFOrdGreaterThan %bool %37 %float_1
OpSelectionMerge %40 None
OpBranchConditional %39 %41 %40
%41 = OpLabel
OpBranch %26
%40 = OpLabel
OpBranch %27
%27 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%43 = OpIAdd %int %45 %int_1
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %i %43
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %25
%26 = OpLabel
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%47 = OpPhi %float %48 %28 %46 %41
OpStore %fo %47
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + before, predefs + after, true,
true);
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
}
TEST_F(LocalSSAElimTest, IfThenElse) {
// #version 140
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// in vec4 BaseColor;
// in float f;
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// void main()
// {
// vec4 v;
// if (f >= 0)
// v = BaseColor * 0.5;
// else
// v = BaseColor + vec4(1.0,1.0,1.0,1.0);
// gl_FragColor = v;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %f %BaseColor %gl_FragColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpName %main "main"
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpName %f "f"
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %gl_FragColor "gl_FragColor"
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%void = OpTypeVoid
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Input_float = OpTypePointer Input %float
%f = OpVariable %_ptr_Input_float Input
%float_0 = OpConstant %float 0
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%float_0_5 = OpConstant %float 0.5
%float_1 = OpConstant %float 1
%18 = OpConstantComposite %v4float %float_1 %float_1 %float_1 %float_1
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
)";
const std::string before =
R"(%main = OpFunction %void None %8
%20 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%21 = OpLoad %float %f
%22 = OpFOrdGreaterThanEqual %bool %21 %float_0
OpSelectionMerge %23 None
OpBranchConditional %22 %24 %25
%24 = OpLabel
%26 = OpLoad %v4float %BaseColor
%27 = OpVectorTimesScalar %v4float %26 %float_0_5
OpStore %v %27
OpBranch %23
%25 = OpLabel
%28 = OpLoad %v4float %BaseColor
%29 = OpFAdd %v4float %28 %18
OpStore %v %29
OpBranch %23
%23 = OpLabel
%30 = OpLoad %v4float %v
OpStore %gl_FragColor %30
OpReturn
OpFunctionEnd
)";
const std::string after =
R"(%main = OpFunction %void None %8
%20 = OpLabel
Opt: Fix HasLoads to not report decoration as load.
2017-11-06 18:25:24 +00:00
%v = OpVariable %_ptr_Function_v4float Function
%21 = OpLoad %float %f
%22 = OpFOrdGreaterThanEqual %bool %21 %float_0
OpSelectionMerge %23 None
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
OpBranchConditional %22 %24 %25
Opt: Fix HasLoads to not report decoration as load.
2017-11-06 18:25:24 +00:00
%24 = OpLabel
%26 = OpLoad %v4float %BaseColor
%27 = OpVectorTimesScalar %v4float %26 %float_0_5
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
OpStore %v %27
Opt: Fix HasLoads to not report decoration as load.
2017-11-06 18:25:24 +00:00
OpBranch %23
%25 = OpLabel
%28 = OpLoad %v4float %BaseColor
%29 = OpFAdd %v4float %28 %18
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
OpStore %v %29
Opt: Fix HasLoads to not report decoration as load.
2017-11-06 18:25:24 +00:00
OpBranch %23
%23 = OpLabel
%31 = OpPhi %v4float %27 %24 %29 %25
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %gl_FragColor %31
Opt: Fix HasLoads to not report decoration as load.
2017-11-06 18:25:24 +00:00
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + before, predefs + after, true,
true);
Opt: Fix HasLoads to not report decoration as load.
2017-11-06 18:25:24 +00:00
}
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
TEST_F(LocalSSAElimTest, IfThen) {
// #version 140
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// in vec4 BaseColor;
// in float f;
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// void main()
// {
// vec4 v = BaseColor;
// if (f <= 0)
// v = v * 0.5;
// gl_FragColor = v;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BaseColor %f %gl_FragColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpName %main "main"
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %f "f"
OpName %gl_FragColor "gl_FragColor"
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%void = OpTypeVoid
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%f = OpVariable %_ptr_Input_float Input
%float_0 = OpConstant %float 0
%bool = OpTypeBool
%float_0_5 = OpConstant %float 0.5
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
)";
const std::string before =
R"(%main = OpFunction %void None %8
%18 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%19 = OpLoad %v4float %BaseColor
OpStore %v %19
%20 = OpLoad %float %f
%21 = OpFOrdLessThanEqual %bool %20 %float_0
OpSelectionMerge %22 None
OpBranchConditional %21 %23 %22
%23 = OpLabel
%24 = OpLoad %v4float %v
%25 = OpVectorTimesScalar %v4float %24 %float_0_5
OpStore %v %25
OpBranch %22
%22 = OpLabel
%26 = OpLoad %v4float %v
OpStore %gl_FragColor %26
OpReturn
OpFunctionEnd
)";
const std::string after =
R"(%main = OpFunction %void None %8
%18 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%v = OpVariable %_ptr_Function_v4float Function
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%19 = OpLoad %v4float %BaseColor
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %19
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%20 = OpLoad %float %f
%21 = OpFOrdLessThanEqual %bool %20 %float_0
OpSelectionMerge %22 None
OpBranchConditional %21 %23 %22
%23 = OpLabel
%25 = OpVectorTimesScalar %v4float %19 %float_0_5
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %25
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %22
%22 = OpLabel
%27 = OpPhi %v4float %19 %18 %25 %23
OpStore %gl_FragColor %27
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + before, predefs + after, true,
true);
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
}
TEST_F(LocalSSAElimTest, Switch) {
// #version 140
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// in vec4 BaseColor;
// in float f;
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// void main()
// {
// vec4 v = BaseColor;
// int i = int(f);
// switch (i) {
// case 0:
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
// v = v * 0.25;
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// break;
// case 1:
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
// v = v * 0.625;
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// break;
// case 2:
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
// v = v * 0.75;
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// break;
// default:
// break;
// }
// gl_FragColor = v;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BaseColor %f %gl_FragColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpName %main "main"
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %i "i"
OpName %f "f"
OpName %gl_FragColor "gl_FragColor"
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%void = OpTypeVoid
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%_ptr_Input_float = OpTypePointer Input %float
%f = OpVariable %_ptr_Input_float Input
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%float_0_25 = OpConstant %float 0.25
%float_0_625 = OpConstant %float 0.625
%float_0_75 = OpConstant %float 0.75
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
)";
const std::string before =
R"(%main = OpFunction %void None %9
%21 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%i = OpVariable %_ptr_Function_int Function
%22 = OpLoad %v4float %BaseColor
OpStore %v %22
%23 = OpLoad %float %f
%24 = OpConvertFToS %int %23
OpStore %i %24
%25 = OpLoad %int %i
OpSelectionMerge %26 None
OpSwitch %25 %27 0 %28 1 %29 2 %30
%27 = OpLabel
OpBranch %26
%28 = OpLabel
%31 = OpLoad %v4float %v
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%32 = OpVectorTimesScalar %v4float %31 %float_0_25
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpStore %v %32
OpBranch %26
%29 = OpLabel
%33 = OpLoad %v4float %v
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%34 = OpVectorTimesScalar %v4float %33 %float_0_625
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpStore %v %34
OpBranch %26
%30 = OpLabel
%35 = OpLoad %v4float %v
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%36 = OpVectorTimesScalar %v4float %35 %float_0_75
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpStore %v %36
OpBranch %26
%26 = OpLabel
%37 = OpLoad %v4float %v
OpStore %gl_FragColor %37
OpReturn
OpFunctionEnd
)";
const std::string after =
R"(%main = OpFunction %void None %9
%21 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%v = OpVariable %_ptr_Function_v4float Function
%i = OpVariable %_ptr_Function_int Function
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%22 = OpLoad %v4float %BaseColor
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %22
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%23 = OpLoad %float %f
%24 = OpConvertFToS %int %23
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %i %24
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpSelectionMerge %26 None
OpSwitch %24 %27 0 %28 1 %29 2 %30
%27 = OpLabel
OpBranch %26
%28 = OpLabel
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%32 = OpVectorTimesScalar %v4float %22 %float_0_25
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %32
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %26
%29 = OpLabel
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%34 = OpVectorTimesScalar %v4float %22 %float_0_625
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %34
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %26
%30 = OpLabel
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%36 = OpVectorTimesScalar %v4float %22 %float_0_75
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %36
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %26
%26 = OpLabel
%38 = OpPhi %v4float %22 %27 %32 %28 %34 %29 %36 %30
OpStore %gl_FragColor %38
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + before, predefs + after, true,
true);
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
}
TEST_F(LocalSSAElimTest, SwitchWithFallThrough) {
// #version 140
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// in vec4 BaseColor;
// in float f;
Re-format source tree - NFC. Re-formatted the source tree with the command: $ /usr/bin/clang-format -style=file -i \ $(find include source tools test utils -name '*.cpp' -or -name '*.h') This required a fix to source/val/decoration.h. It was not including spirv.h, which broke builds when the #include headers were re-ordered by clang-format.
2017-11-27 15:16:41 +00:00
//
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// void main()
// {
// vec4 v = BaseColor;
// int i = int(f);
// switch (i) {
// case 0:
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
// v = v * 0.25;
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// break;
// case 1:
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
// v = v + 0.25;
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// case 2:
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
// v = v * 0.75;
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// break;
// default:
// break;
// }
// gl_FragColor = v;
// }
const std::string predefs =
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BaseColor %f %gl_FragColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpName %main "main"
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %i "i"
OpName %f "f"
OpName %gl_FragColor "gl_FragColor"
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%void = OpTypeVoid
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%_ptr_Input_float = OpTypePointer Input %float
%f = OpVariable %_ptr_Input_float Input
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%float_0_25 = OpConstant %float 0.25
%float_0_75 = OpConstant %float 0.75
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
)";
const std::string before =
R"(%main = OpFunction %void None %9
%20 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function
%i = OpVariable %_ptr_Function_int Function
%21 = OpLoad %v4float %BaseColor
OpStore %v %21
%22 = OpLoad %float %f
%23 = OpConvertFToS %int %22
OpStore %i %23
%24 = OpLoad %int %i
OpSelectionMerge %25 None
OpSwitch %24 %26 0 %27 1 %28 2 %29
%26 = OpLabel
OpBranch %25
%27 = OpLabel
%30 = OpLoad %v4float %v
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%31 = OpVectorTimesScalar %v4float %30 %float_0_25
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpStore %v %31
OpBranch %25
%28 = OpLabel
%32 = OpLoad %v4float %v
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%33 = OpCompositeConstruct %v4float %float_0_25 %float_0_25 %float_0_25 %float_0_25
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%34 = OpFAdd %v4float %32 %33
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
OpStore %v %34
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %29
%29 = OpLabel
%35 = OpLoad %v4float %v
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%36 = OpVectorTimesScalar %v4float %35 %float_0_75
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
OpStore %v %36
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %25
%25 = OpLabel
%37 = OpLoad %v4float %v
OpStore %gl_FragColor %37
OpReturn
OpFunctionEnd
)";
const std::string after =
R"(%main = OpFunction %void None %9
%20 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%v = OpVariable %_ptr_Function_v4float Function
%i = OpVariable %_ptr_Function_int Function
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%21 = OpLoad %v4float %BaseColor
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %21
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%22 = OpLoad %float %f
%23 = OpConvertFToS %int %22
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %i %23
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpSelectionMerge %25 None
OpSwitch %23 %26 0 %27 1 %28 2 %29
%26 = OpLabel
OpBranch %25
%27 = OpLabel
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%31 = OpVectorTimesScalar %v4float %21 %float_0_25
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %31
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %25
%28 = OpLabel
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%33 = OpCompositeConstruct %v4float %float_0_25 %float_0_25 %float_0_25 %float_0_25
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
%34 = OpFAdd %v4float %21 %33
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %34
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %29
%29 = OpLabel
%38 = OpPhi %v4float %21 %20 %34 %28
hex_float: Use max_digits10 for the float precision CPPreference.com has this description of digits10: “The value of std::numeric_limits<T>::digits10 is the number of base-10 digits that can be represented by the type T without change, that is, any number with this many significant decimal digits can be converted to a value of type T and back to decimal form, without change due to rounding or overflow.” This means that any number with this many digits can be represented accurately in the corresponding type. A change in any digit in a number after that may or may not cause it a different bitwise representation. Therefore this isn’t necessarily enough precision to accurately represent the value in text. Instead we need max_digits10 which has the following description: “The value of std::numeric_limits<T>::max_digits10 is the number of base-10 digits that are necessary to uniquely represent all distinct values of the type T, such as necessary for serialization/deserialization to text.” The patch includes a test case in hex_float_test which tries to do a round-robin conversion of a number that requires more than 6 decimal places to be accurately represented. This would fail without the patch. Sadly this also breaks a bunch of other tests. Some of the tests in hex_float_test use ldexp and then compare it with a value which is not the same as the one returned by ldexp but instead is the value rounded to 6 decimals. Others use values that are not evenly representable as a binary floating fraction but then happened to generate the same value when rounded to 6 decimals. Where the actual value didn’t seem to matter these have been changed with different values that can be represented as a binary fraction.
2018-03-30 23:35:45 +00:00
%36 = OpVectorTimesScalar %v4float %38 %float_0_75
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %36
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
OpBranch %25
%25 = OpLabel
%39 = OpPhi %v4float %21 %26 %31 %27 %36 %29
OpStore %gl_FragColor %39
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + before, predefs + after, true,
true);
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
}
elim-multi-store: only patch loop header phis that we created There can already be OpPhi instructions in a loop header that are unrelated to the optimization. We should not be patching those. Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/826
2017-09-20 15:07:55 +00:00
TEST_F(LocalSSAElimTest, DontPatchPhiInLoopHeaderThatIsNotAVar) {
// From https://github.com/KhronosGroup/SPIRV-Tools/issues/826
// Don't try patching the (%16 %7) value/predecessor pair in the OpPhi.
// That OpPhi is unrelated to this optimization: we did not set that up
// in the SSA initialization for the loop header block.
// The pass should be a no-op on this module.
const std::string before = R"(OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%3 = OpTypeFunction %void
%float = OpTypeFloat 32
%float_1 = OpConstant %float 1
%1 = OpFunction %void None %3
%6 = OpLabel
OpBranch %7
%7 = OpLabel
%8 = OpPhi %float %float_1 %6 %9 %7
%9 = OpFAdd %float %8 %float_1
OpLoopMerge %10 %7 None
OpBranch %7
%10 = OpLabel
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(before, before, true, true);
elim-multi-store: only patch loop header phis that we created There can already be OpPhi instructions in a loop header that are unrelated to the optimization. We should not be patching those. Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/826
2017-09-20 15:07:55 +00:00
}
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
TEST_F(LocalSSAElimTest, OptInitializedVariableLikeStore) {
// Note: SPIR-V edited to change store to v into variable initialization
//
// #version 450
//
// layout (location=0) in vec4 iColor;
// layout (location=1) in float fi;
// layout (location=0) out vec4 oColor;
//
// void main()
// {
// vec4 v = vec4(0.0);
// if (fi < 0.0)
// v.x = iColor.x;
// oColor = v;
// }
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
const std::string predefs =
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
R"(OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %fi %iColor %oColor
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 450
OpName %main "main"
OpName %v "v"
OpName %fi "fi"
OpName %iColor "iColor"
OpName %oColor "oColor"
OpDecorate %fi Location 1
OpDecorate %iColor Location 0
OpDecorate %oColor Location 0
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%float_0 = OpConstant %float 0
%13 = OpConstantComposite %v4float %float_0 %float_0 %float_0 %float_0
%_ptr_Input_float = OpTypePointer Input %float
%fi = OpVariable %_ptr_Input_float Input
%bool = OpTypeBool
%_ptr_Input_v4float = OpTypePointer Input %v4float
%iColor = OpVariable %_ptr_Input_v4float Input
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%_ptr_Function_float = OpTypePointer Function %float
%_ptr_Output_v4float = OpTypePointer Output %v4float
%oColor = OpVariable %_ptr_Output_v4float Output
)";
const std::string func_before =
R"(%main = OpFunction %void None %8
%21 = OpLabel
%v = OpVariable %_ptr_Function_v4float Function %13
%22 = OpLoad %float %fi
%23 = OpFOrdLessThan %bool %22 %float_0
OpSelectionMerge %24 None
OpBranchConditional %23 %25 %24
%25 = OpLabel
%26 = OpAccessChain %_ptr_Input_float %iColor %uint_0
%27 = OpLoad %float %26
%28 = OpLoad %v4float %v
%29 = OpCompositeInsert %v4float %27 %28 0
OpStore %v %29
OpBranch %24
%24 = OpLabel
%30 = OpLoad %v4float %v
OpStore %oColor %30
OpReturn
OpFunctionEnd
)";
const std::string func_after =
R"(%main = OpFunction %void None %8
%21 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%v = OpVariable %_ptr_Function_v4float Function %13
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
%22 = OpLoad %float %fi
%23 = OpFOrdLessThan %bool %22 %float_0
OpSelectionMerge %24 None
OpBranchConditional %23 %25 %24
%25 = OpLabel
%26 = OpAccessChain %_ptr_Input_float %iColor %uint_0
%27 = OpLoad %float %26
%29 = OpCompositeInsert %v4float %27 %13 0
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
OpStore %v %29
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
OpBranch %24
%24 = OpLabel
%31 = OpPhi %v4float %13 %21 %29 %25
OpStore %oColor %31
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(predefs + func_before,
predefs + func_after, true, true);
MultiStore: Support OpVariable Initialization Treat an OpVariable with initialization as if it was an OpStore. With PR #1073, this completes work for issue #1017.
2017-12-08 17:44:15 +00:00
}
Allow pointers to pointers in logical addressing mode. A few optimizations are updates to handle code that is suppose to be using the logical addressing mode, but still has variables that contain pointers as long as the pointer are to opaque objects. This is called "relaxed logical addressing". |Instruction::GetBaseAddress| will check that pointers that are use meet the relaxed logical addressing rules. Optimization that now handle relaxed logical addressing instead of logical addressing are: - aggressive dead-code elimination - local access chain convert - local store elimination passes.
2017-12-11 18:10:24 +00:00
TEST_F(LocalSSAElimTest, PointerVariable) {
// Test that checks if a pointer variable is removed.
const std::string before =
R"(OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %1 "main" %2
OpExecutionMode %1 OriginUpperLeft
OpMemberDecorate %_struct_3 0 Offset 0
OpDecorate %_runtimearr__struct_3 ArrayStride 16
OpMemberDecorate %_struct_5 0 Offset 0
OpDecorate %_struct_5 BufferBlock
OpMemberDecorate %_struct_6 0 Offset 0
OpDecorate %_struct_6 BufferBlock
OpDecorate %2 Location 0
OpDecorate %7 DescriptorSet 0
OpDecorate %7 Binding 0
%void = OpTypeVoid
%10 = OpTypeFunction %void
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Output_v4float = OpTypePointer Output %v4float
%_ptr_Uniform_v4float = OpTypePointer Uniform %v4float
%_struct_3 = OpTypeStruct %v4float
%_runtimearr__struct_3 = OpTypeRuntimeArray %_struct_3
%_struct_5 = OpTypeStruct %_runtimearr__struct_3
%_ptr_Uniform__struct_5 = OpTypePointer Uniform %_struct_5
%_struct_6 = OpTypeStruct %int
%_ptr_Uniform__struct_6 = OpTypePointer Uniform %_struct_6
%_ptr_Function__ptr_Uniform__struct_5 = OpTypePointer Function %_ptr_Uniform__struct_5
%_ptr_Function__ptr_Uniform__struct_6 = OpTypePointer Function %_ptr_Uniform__struct_6
%int_0 = OpConstant %int 0
%uint_0 = OpConstant %uint 0
%2 = OpVariable %_ptr_Output_v4float Output
%7 = OpVariable %_ptr_Uniform__struct_5 Uniform
%1 = OpFunction %void None %10
%23 = OpLabel
%24 = OpVariable %_ptr_Function__ptr_Uniform__struct_5 Function
OpStore %24 %7
%26 = OpLoad %_ptr_Uniform__struct_5 %24
%27 = OpAccessChain %_ptr_Uniform_v4float %26 %int_0 %uint_0 %int_0
%28 = OpLoad %v4float %27
%29 = OpCopyObject %v4float %28
OpStore %2 %28
OpReturn
OpFunctionEnd
)";
const std::string after =
R"(OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %1 "main" %2
OpExecutionMode %1 OriginUpperLeft
OpMemberDecorate %_struct_3 0 Offset 0
OpDecorate %_runtimearr__struct_3 ArrayStride 16
OpMemberDecorate %_struct_5 0 Offset 0
OpDecorate %_struct_5 BufferBlock
OpMemberDecorate %_struct_6 0 Offset 0
OpDecorate %_struct_6 BufferBlock
OpDecorate %2 Location 0
OpDecorate %7 DescriptorSet 0
OpDecorate %7 Binding 0
%void = OpTypeVoid
%10 = OpTypeFunction %void
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Output_v4float = OpTypePointer Output %v4float
%_ptr_Uniform_v4float = OpTypePointer Uniform %v4float
%_struct_3 = OpTypeStruct %v4float
%_runtimearr__struct_3 = OpTypeRuntimeArray %_struct_3
%_struct_5 = OpTypeStruct %_runtimearr__struct_3
%_ptr_Uniform__struct_5 = OpTypePointer Uniform %_struct_5
%_struct_6 = OpTypeStruct %int
%_ptr_Uniform__struct_6 = OpTypePointer Uniform %_struct_6
%_ptr_Function__ptr_Uniform__struct_5 = OpTypePointer Function %_ptr_Uniform__struct_5
%_ptr_Function__ptr_Uniform__struct_6 = OpTypePointer Function %_ptr_Uniform__struct_6
%int_0 = OpConstant %int 0
%uint_0 = OpConstant %uint 0
%2 = OpVariable %_ptr_Output_v4float Output
%7 = OpVariable %_ptr_Uniform__struct_5 Uniform
%1 = OpFunction %void None %10
%23 = OpLabel
Remove uses DCEInst and call ADCE The algorithm used in DCEInst to remove dead code is very slow. It is fine if you only want to remove a small number of instructions, but, if you need to remove a large number of instructions, then the algorithm in ADCE is much faster. This PR removes the calls to DCEInst in the load-store removal passes and adds a pass of ADCE afterwards. A number of different iterations of the order of optimization, and I believe this is the best I could find. The results I have on 3 sets of shaders are: Legalization: Set 1: 5.39 -> 5.01 Set 2: 13.98 -> 8.38 Set 3: 98.00 -> 96.26 Performance passes: Set 1: 6.90 -> 5.23 Set 2: 10.11 -> 6.62 Set 3: 253.69 -> 253.74 Size reduction passes: Set 1: 7.16 -> 7.25 Set 2: 17.17 -> 16.81 Set 3: 112.06 -> 107.71 Note that the third set's compile time is large because of the large number of basic blocks, not so much because of the number of instructions. That is why we don't see much gain there.
2018-02-22 21:05:37 +00:00
%24 = OpVariable %_ptr_Function__ptr_Uniform__struct_5 Function
OpStore %24 %7
Allow pointers to pointers in logical addressing mode. A few optimizations are updates to handle code that is suppose to be using the logical addressing mode, but still has variables that contain pointers as long as the pointer are to opaque objects. This is called "relaxed logical addressing". |Instruction::GetBaseAddress| will check that pointers that are use meet the relaxed logical addressing rules. Optimization that now handle relaxed logical addressing instead of logical addressing are: - aggressive dead-code elimination - local access chain convert - local store elimination passes.
2017-12-11 18:10:24 +00:00
%27 = OpAccessChain %_ptr_Uniform_v4float %7 %int_0 %uint_0 %int_0
%28 = OpLoad %v4float %27
%29 = OpCopyObject %v4float %28
OpStore %2 %28
OpReturn
OpFunctionEnd
)";
Validate pointer variables (#2111) Fixes #2104 * Checks the rules for logical addressing and variable pointers * Has an out for relaxed logical pointers * Updated PassFixture to expose validator options * enabled relaxed logical pointers for some tests * New validator tests
2018-11-27 21:47:10 +00:00
// Relax logical pointers to allow pointer allocations.
Allow pointers to pointers in logical addressing mode. A few optimizations are updates to handle code that is suppose to be using the logical addressing mode, but still has variables that contain pointers as long as the pointer are to opaque objects. This is called "relaxed logical addressing". |Instruction::GetBaseAddress| will check that pointers that are use meet the relaxed logical addressing rules. Optimization that now handle relaxed logical addressing instead of logical addressing are: - aggressive dead-code elimination - local access chain convert - local store elimination passes.
2017-12-11 18:10:24 +00:00
SetAssembleOptions(SPV_TEXT_TO_BINARY_OPTION_PRESERVE_NUMERIC_IDS);
Validate pointer variables (#2111) Fixes #2104 * Checks the rules for logical addressing and variable pointers * Has an out for relaxed logical pointers * Updated PassFixture to expose validator options * enabled relaxed logical pointers for some tests * New validator tests
2018-11-27 21:47:10 +00:00
ValidatorOptions()->relax_logical_pointer = true;
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndCheck<SSARewritePass>(before, after, true, true);
Allow pointers to pointers in logical addressing mode. A few optimizations are updates to handle code that is suppose to be using the logical addressing mode, but still has variables that contain pointers as long as the pointer are to opaque objects. This is called "relaxed logical addressing". |Instruction::GetBaseAddress| will check that pointers that are use meet the relaxed logical addressing rules. Optimization that now handle relaxed logical addressing instead of logical addressing are: - aggressive dead-code elimination - local access chain convert - local store elimination passes.
2017-12-11 18:10:24 +00:00
}
Maintain instruction to block mapping in phi insertion * Changed MemPass::InsertPhiInstructions to set basic blocks for new phis * Local SSA elim now maintains instr to block mapping * Added a test and confirmed it fails without the updated phis * IRContext::set_instr_block no longer builds the map if the analysis is invalid * Added instruction to block mapping verification to IRContext::IsConsistent()
2018-01-11 20:05:39 +00:00
TEST_F(LocalSSAElimTest, VerifyInstToBlockMap) {
// #version 140
//
// in vec4 BC;
// out float fo;
//
// void main()
// {
// float f = 0.0;
// for (int i=0; i<4; i++) {
// f = f + BC[i];
// }
// fo = f;
// }
const std::string text = R"(
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
%main = OpFunction %void None %8
%22 = OpLabel
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
OpBranch %23
%23 = OpLabel
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%27 = OpLoad %int %i
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%30 = OpLoad %float %f
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %25
%25 = OpLabel
%35 = OpLoad %int %i
%36 = OpIAdd %int %35 %int_1
OpStore %i %36
OpBranch %23
%24 = OpLabel
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
Move tests into namespaces (#1689) This CL moves the test into namespaces based on their directories.
2018-07-11 13:24:49 +00:00
std::unique_ptr<IRContext> context =
Maintain instruction to block mapping in phi insertion * Changed MemPass::InsertPhiInstructions to set basic blocks for new phis * Local SSA elim now maintains instr to block mapping * Added a test and confirmed it fails without the updated phis * IRContext::set_instr_block no longer builds the map if the analysis is invalid * Added instruction to block mapping verification to IRContext::IsConsistent()
2018-01-11 20:05:39 +00:00
BuildModule(SPV_ENV_UNIVERSAL_1_1, nullptr, text,
SPV_TEXT_TO_BINARY_OPTION_PRESERVE_NUMERIC_IDS);
EXPECT_NE(nullptr, context);
// Force the instruction to block mapping to get built.
context->get_instr_block(27u);
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
auto pass = MakeUnique<SSARewritePass>();
Maintain instruction to block mapping in phi insertion * Changed MemPass::InsertPhiInstructions to set basic blocks for new phis * Local SSA elim now maintains instr to block mapping * Added a test and confirmed it fails without the updated phis * IRContext::set_instr_block no longer builds the map if the analysis is invalid * Added instruction to block mapping verification to IRContext::IsConsistent()
2018-01-11 20:05:39 +00:00
pass->SetMessageConsumer(nullptr);
const auto status = pass->Run(context.get());
Move tests into namespaces (#1689) This CL moves the test into namespaces based on their directories.
2018-07-11 13:24:49 +00:00
EXPECT_TRUE(status == Pass::Status::SuccessWithChange);
Maintain instruction to block mapping in phi insertion * Changed MemPass::InsertPhiInstructions to set basic blocks for new phis * Local SSA elim now maintains instr to block mapping * Added a test and confirmed it fails without the updated phis * IRContext::set_instr_block no longer builds the map if the analysis is invalid * Added instruction to block mapping verification to IRContext::IsConsistent()
2018-01-11 20:05:39 +00:00
}
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
TEST_F(LocalSSAElimTest, CompositeExtractProblem) {
const std::string spv_asm = R"(
OpCapability Tessellation
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
Validate static uses of interfaces Fixes #1120 Checks that all static uses of the Input and Output variables are listed as interfaces in each corresponding entry point declaration. * Changed validation state to track interface lists * updated many tests * Modified validation state to store entry point names * Combined with interface list and called EntryPointDescription * Updated uses * Changed interface validation error messages to output entry point name in addtion to ID
2018-05-31 13:07:09 +00:00
OpEntryPoint TessellationControl %2 "main" %16 %17 %18 %20 %22 %26 %27 %30 %31
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%void = OpTypeVoid
%4 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%uint = OpTypeInt 32 0
%uint_3 = OpConstant %uint 3
%v3float = OpTypeVector %float 3
%v2float = OpTypeVector %float 2
%_struct_11 = OpTypeStruct %v4float %v4float %v4float %v3float %v3float %v2float %v2float
%_arr__struct_11_uint_3 = OpTypeArray %_struct_11 %uint_3
%_ptr_Function__arr__struct_11_uint_3 = OpTypePointer Function %_arr__struct_11_uint_3
%_arr_v4float_uint_3 = OpTypeArray %v4float %uint_3
%_ptr_Input__arr_v4float_uint_3 = OpTypePointer Input %_arr_v4float_uint_3
%16 = OpVariable %_ptr_Input__arr_v4float_uint_3 Input
%17 = OpVariable %_ptr_Input__arr_v4float_uint_3 Input
%18 = OpVariable %_ptr_Input__arr_v4float_uint_3 Input
%_ptr_Input_uint = OpTypePointer Input %uint
%20 = OpVariable %_ptr_Input_uint Input
%_ptr_Output__arr_v4float_uint_3 = OpTypePointer Output %_arr_v4float_uint_3
%22 = OpVariable %_ptr_Output__arr_v4float_uint_3 Output
%_ptr_Output_v4float = OpTypePointer Output %v4float
%_arr_v3float_uint_3 = OpTypeArray %v3float %uint_3
%_ptr_Input__arr_v3float_uint_3 = OpTypePointer Input %_arr_v3float_uint_3
%26 = OpVariable %_ptr_Input__arr_v3float_uint_3 Input
%27 = OpVariable %_ptr_Input__arr_v3float_uint_3 Input
%_arr_v2float_uint_3 = OpTypeArray %v2float %uint_3
%_ptr_Input__arr_v2float_uint_3 = OpTypePointer Input %_arr_v2float_uint_3
%30 = OpVariable %_ptr_Input__arr_v2float_uint_3 Input
%31 = OpVariable %_ptr_Input__arr_v2float_uint_3 Input
%_ptr_Function__struct_11 = OpTypePointer Function %_struct_11
%2 = OpFunction %void None %4
%33 = OpLabel
Validator: function scope variable at start of entry block #1923 All OpVariable instructions in a function must be the first instructions in the first block.
2018-09-28 17:45:43 +00:00
%66 = OpVariable %_ptr_Function__arr__struct_11_uint_3 Function
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
%34 = OpLoad %_arr_v4float_uint_3 %16
%35 = OpLoad %_arr_v4float_uint_3 %17
%36 = OpLoad %_arr_v4float_uint_3 %18
%37 = OpLoad %_arr_v3float_uint_3 %26
%38 = OpLoad %_arr_v3float_uint_3 %27
%39 = OpLoad %_arr_v2float_uint_3 %30
%40 = OpLoad %_arr_v2float_uint_3 %31
%41 = OpCompositeExtract %v4float %34 0
%42 = OpCompositeExtract %v4float %35 0
%43 = OpCompositeExtract %v4float %36 0
%44 = OpCompositeExtract %v3float %37 0
%45 = OpCompositeExtract %v3float %38 0
%46 = OpCompositeExtract %v2float %39 0
%47 = OpCompositeExtract %v2float %40 0
%48 = OpCompositeConstruct %_struct_11 %41 %42 %43 %44 %45 %46 %47
%49 = OpCompositeExtract %v4float %34 1
%50 = OpCompositeExtract %v4float %35 1
%51 = OpCompositeExtract %v4float %36 1
%52 = OpCompositeExtract %v3float %37 1
%53 = OpCompositeExtract %v3float %38 1
%54 = OpCompositeExtract %v2float %39 1
%55 = OpCompositeExtract %v2float %40 1
%56 = OpCompositeConstruct %_struct_11 %49 %50 %51 %52 %53 %54 %55
%57 = OpCompositeExtract %v4float %34 2
%58 = OpCompositeExtract %v4float %35 2
%59 = OpCompositeExtract %v4float %36 2
%60 = OpCompositeExtract %v3float %37 2
%61 = OpCompositeExtract %v3float %38 2
%62 = OpCompositeExtract %v2float %39 2
%63 = OpCompositeExtract %v2float %40 2
%64 = OpCompositeConstruct %_struct_11 %57 %58 %59 %60 %61 %62 %63
%65 = OpCompositeConstruct %_arr__struct_11_uint_3 %48 %56 %64
%67 = OpLoad %uint %20
; CHECK OpStore {{%\d+}} [[store_source:%\d+]]
OpStore %66 %65
%68 = OpAccessChain %_ptr_Function__struct_11 %66 %67
; This load was being removed, because %_ptr_Function__struct_11 was being
; wrongfully considered an SSA target.
; CHECK OpLoad %_struct_11 %68
%69 = OpLoad %_struct_11 %68
; Similarly, %69 cannot be replaced with %65.
; CHECK-NOT: OpCompositeExtract %v4float [[store_source]] 0
%70 = OpCompositeExtract %v4float %69 0
%71 = OpAccessChain %_ptr_Output_v4float %22 %67
OpStore %71 %70
OpReturn
OpFunctionEnd)";
Move tests into namespaces (#1689) This CL moves the test into namespaces based on their directories.
2018-07-11 13:24:49 +00:00
SinglePassRunAndMatch<SSARewritePass>(spv_asm, true);
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
}
Copy decorations when creating new ids. (#1843) * Copy decorations when creating new ids. When creating a new value based on an old value, we need to copy the decorations to the new id. This change does this in 3 places: 1) The variable holding the return value of the function generated by merge return should get decorations from the function. 2) The results of the OpPhi instructions should get decorations from the variable they are replacing in the ssa writer. 3) In local access chain convert the intermediate struct (result of OpCompositeInsert) generated for the store replacement should get its decorations from the variable being stored to. Fixes #1787.
2018-08-24 15:55:39 +00:00
// Test that the RelaxedPrecision decoration on the variable to added to the
// result of the OpPhi instruction.
TEST_F(LocalSSAElimTest, DecoratedVariable) {
const std::string spv_asm = R"(
; CHECK: OpDecorate [[var:%\w+]] RelaxedPrecision
; CHECK: OpDecorate [[phi_id:%\w+]] RelaxedPrecision
; CHECK: [[phi_id]] = OpPhi
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %2 "main"
OpDecorate %v RelaxedPrecision
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%bool = OpTypeBool
%true = OpConstantTrue %bool
%int = OpTypeInt 32 0
%int_p = OpTypePointer Function %int
%int_1 = OpConstant %int 1
%int_0 = OpConstant %int 0
%2 = OpFunction %void None %func_t
%33 = OpLabel
%v = OpVariable %int_p Function
OpSelectionMerge %merge None
OpBranchConditional %true %l1 %l2
%l1 = OpLabel
OpStore %v %int_1
OpBranch %merge
%l2 = OpLabel
OpStore %v %int_0
OpBranch %merge
%merge = OpLabel
%ld = OpLoad %int %v
OpReturn
OpFunctionEnd)";
SinglePassRunAndMatch<SSARewritePass>(spv_asm, true);
}
SSA rewrite pass. This pass replaces the load/store elimination passes. It implements the SSA re-writing algorithm proposed in Simple and Efficient Construction of Static Single Assignment Form. Braun M., Buchwald S., Hack S., Leißa R., Mallon C., Zwinkau A. (2013) In: Jhala R., De Bosschere K. (eds) Compiler Construction. CC 2013. Lecture Notes in Computer Science, vol 7791. Springer, Berlin, Heidelberg https://link.springer.com/chapter/10.1007/978-3-642-37051-9_6 In contrast to common eager algorithms based on dominance and dominance frontier information, this algorithm works backwards from load operations. When a target variable is loaded, it queries the variable's reaching definition. If the reaching definition is unknown at the current location, it searches backwards in the CFG, inserting Phi instructions at join points in the CFG along the way until it finds the desired store instruction. The algorithm avoids repeated lookups using memoization. For reducible CFGs, which are a superset of the structured CFGs in SPIRV, this algorithm is proven to produce minimal SSA. That is, it inserts the minimal number of Phi instructions required to ensure the SSA property, but some Phi instructions may be dead (https://en.wikipedia.org/wiki/Static_single_assignment_form).
2018-02-22 21:18:29 +00:00
SSA-rewriter: make sure phi entries are unique. (#2206) If there are multiple edges to a basic block, then the ssa rewriter will create OpPhi instructions with duplicate entries. This is invalid, and it is fixed in this commit. Fixes #2202.
2018-12-18 18:14:27 +00:00
// Test that the RelaxedPrecision decoration on the variable to added to the
// result of the OpPhi instruction.
TEST_F(LocalSSAElimTest, MultipleEdges) {
const std::string spv_asm = R"(
; CHECK: OpSelectionMerge
; CHECK: [[header_bb:%\w+]] = OpLabel
; CHECK-NOT: OpLabel
; CHECK: OpSwitch {{%\w+}} {{%\w+}} 76 [[bb1:%\w+]] 17 [[bb2:%\w+]]
; CHECK-SAME: 4 [[bb2]]
; CHECK: [[bb2]] = OpLabel
; CHECK-NEXT: OpPhi [[type:%\w+]] [[val:%\w+]] [[header_bb]] %int_0 [[bb1]]
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %4 "main"
OpExecutionMode %4 OriginUpperLeft
OpSource ESSL 310
%void = OpTypeVoid
%3 = OpTypeFunction %void
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%bool = OpTypeBool
%true = OpConstantTrue %bool
%false = OpConstantFalse %bool
%int_1 = OpConstant %int 1
%4 = OpFunction %void None %3
%5 = OpLabel
%8 = OpVariable %_ptr_Function_int Function
OpBranch %10
%10 = OpLabel
OpLoopMerge %12 %13 None
OpBranch %14
%14 = OpLabel
OpBranchConditional %true %11 %12
%11 = OpLabel
OpSelectionMerge %19 None
OpBranchConditional %false %18 %19
%18 = OpLabel
OpSelectionMerge %22 None
OpSwitch %int_0 %22 76 %20 17 %21 4 %21
%20 = OpLabel
%23 = OpLoad %int %8
OpStore %8 %int_0
OpBranch %21
%21 = OpLabel
OpBranch %22
%22 = OpLabel
OpBranch %19
%19 = OpLabel
OpBranch %13
%13 = OpLabel
OpBranch %10
%12 = OpLabel
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(spv_asm, true);
}
Handle variable pointer in some optimizations (#2490) * Check var pointer capability in ADCE. * Check var ptr capability for common uniform. * Check var ptr capability in access chain convert. Since we want this pass to run even if there are variable pointer on storage buffers, we had to remove asserts that assumed there were no variable pointers. The functions with the asserts will now work, it becomes the responsibility of the callers to deal with the output as appropriate. * Single block elimination and variable pointers. It seems like the code in local single block elimination is able to handle cases with variable pointers already. This is because the function `HasOnlySupportedRefs` ensures that variables that feed a variable pointer are not candidates. * Single store elimination and variable pointers. It seems like the code in local single stroe elimination is able to handle cases with variable pointers already. This is because the function `FindSingleStoreAndCheckUses` ensures that variables that feed a variable pointer are not candidates. * SSA rewriter and variable pointers. It seems like the code in the two passes that call the SSA rewriter are able to handle cases with variable pointers already. This is because the function `HasOnlySupportedRefs` ensures that variables that feed a variable pointer are not candidates. Fixes #2458.
2019-04-03 16:47:51 +00:00
TEST_F(LocalSSAElimTest, VariablePointerTest1) {
// Check that the load of the first variable is still used and that the load
// of the third variable is propagated. The first load has to remain because
// of the store to the variable pointer.
const std::string text = R"(
; CHECK: [[v1:%\w+]] = OpVariable
; CHECK: [[v2:%\w+]] = OpVariable
; CHECK: [[v3:%\w+]] = OpVariable
; CHECK: [[ld1:%\w+]] = OpLoad %int [[v1]]
; CHECK: OpIAdd %int [[ld1]] %int_0
OpCapability Shader
OpCapability VariablePointers
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %2 "main"
OpExecutionMode %2 LocalSize 1 1 1
OpSource GLSL 450
OpMemberDecorate %_struct_3 0 Offset 0
OpMemberDecorate %_struct_3 1 Offset 4
%void = OpTypeVoid
%5 = OpTypeFunction %void
%int = OpTypeInt 32 1
%bool = OpTypeBool
%_struct_3 = OpTypeStruct %int %int
%_ptr_Function__struct_3 = OpTypePointer Function %_struct_3
%_ptr_Function_int = OpTypePointer Function %int
%true = OpConstantTrue %bool
%int_0 = OpConstant %int 0
%int_1 = OpConstant %int 1
%13 = OpConstantNull %_struct_3
%2 = OpFunction %void None %5
%14 = OpLabel
%15 = OpVariable %_ptr_Function_int Function
%16 = OpVariable %_ptr_Function_int Function
%17 = OpVariable %_ptr_Function_int Function
OpStore %15 %int_1
OpStore %17 %int_0
OpSelectionMerge %18 None
OpBranchConditional %true %19 %20
%19 = OpLabel
OpBranch %18
%20 = OpLabel
OpBranch %18
%18 = OpLabel
%21 = OpPhi %_ptr_Function_int %15 %19 %16 %20
OpStore %21 %int_0
%22 = OpLoad %int %15
%23 = OpLoad %int %17
%24 = OpIAdd %int %22 %23
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, false);
}
TEST_F(LocalSSAElimTest, VariablePointerTest2) {
// Check that the load of the first variable is still used and that the load
// of the third variable is propagated. The first load has to remain because
// of the store to the variable pointer.
const std::string text = R"(
; CHECK: [[v1:%\w+]] = OpVariable
; CHECK: [[v2:%\w+]] = OpVariable
; CHECK: [[v3:%\w+]] = OpVariable
; CHECK: [[ld1:%\w+]] = OpLoad %int [[v1]]
; CHECK: OpIAdd %int [[ld1]] %int_0
OpCapability Shader
OpCapability VariablePointers
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %2 "main"
OpExecutionMode %2 LocalSize 1 1 1
OpSource GLSL 450
OpMemberDecorate %_struct_3 0 Offset 0
OpMemberDecorate %_struct_3 1 Offset 4
%void = OpTypeVoid
%5 = OpTypeFunction %void
%int = OpTypeInt 32 1
%bool = OpTypeBool
%_struct_3 = OpTypeStruct %int %int
%_ptr_Function__struct_3 = OpTypePointer Function %_struct_3
%_ptr_Function_int = OpTypePointer Function %int
%true = OpConstantTrue %bool
%int_0 = OpConstant %int 0
%int_1 = OpConstant %int 1
%13 = OpConstantNull %_struct_3
%2 = OpFunction %void None %5
%14 = OpLabel
%15 = OpVariable %_ptr_Function_int Function
%16 = OpVariable %_ptr_Function_int Function
%17 = OpVariable %_ptr_Function_int Function
OpStore %15 %int_1
OpStore %17 %int_0
OpSelectionMerge %18 None
OpBranchConditional %true %19 %20
%19 = OpLabel
OpBranch %18
%20 = OpLabel
OpBranch %18
%18 = OpLabel
%21 = OpPhi %_ptr_Function_int %15 %19 %16 %20
OpStore %21 %int_0
%22 = OpLoad %int %15
%23 = OpLoad %int %17
%24 = OpIAdd %int %22 %23
OpReturn
OpFunctionEnd
)";
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
SinglePassRunAndMatch<SSARewritePass>(text, false);
Handle variable pointer in some optimizations (#2490) * Check var pointer capability in ADCE. * Check var ptr capability for common uniform. * Check var ptr capability in access chain convert. Since we want this pass to run even if there are variable pointer on storage buffers, we had to remove asserts that assumed there were no variable pointers. The functions with the asserts will now work, it becomes the responsibility of the callers to deal with the output as appropriate. * Single block elimination and variable pointers. It seems like the code in local single block elimination is able to handle cases with variable pointers already. This is because the function `HasOnlySupportedRefs` ensures that variables that feed a variable pointer are not candidates. * Single store elimination and variable pointers. It seems like the code in local single stroe elimination is able to handle cases with variable pointers already. This is because the function `FindSingleStoreAndCheckUses` ensures that variables that feed a variable pointer are not candidates. * SSA rewriter and variable pointers. It seems like the code in the two passes that call the SSA rewriter are able to handle cases with variable pointers already. This is because the function `HasOnlySupportedRefs` ensures that variables that feed a variable pointer are not candidates. Fixes #2458.
2019-04-03 16:47:51 +00:00
}
SSA rewriter: Don't use trivial phis (#2757) When a phi candidate is marked as trivial, we are suppose to update all of its uses to the reference the value that it is being folded to. However, the code updates the uses misses `defs_at_block_`. So at a later time, the id for the trivial phi can reemerge. Fixes #2744
2019-07-23 21:59:30 +00:00
TEST_F(LocalSSAElimTest, ChainedTrivialPhis) {
// Check that the copy object get the undef value implicitly assigned in the
// entry block.
const std::string text = R"(
; CHECK: [[undef:%\w+]] = OpUndef %v4float
; CHECK: OpCopyObject %v4float [[undef]]
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %2 "main"
OpExecutionMode %2 LocalSize 1 18 6
OpSource ESSL 310
%void = OpTypeVoid
%4 = OpTypeFunction %void
%bool = OpTypeBool
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%2 = OpFunction %void None %4
%9 = OpLabel
%10 = OpVariable %_ptr_Function_v4float Function
OpBranch %11
%11 = OpLabel
OpLoopMerge %12 %13 None
OpBranch %14
%14 = OpLabel
%15 = OpUndef %bool
OpBranchConditional %15 %16 %12
%16 = OpLabel
%17 = OpUndef %bool
OpSelectionMerge %18 None
OpBranchConditional %17 %19 %18
%19 = OpLabel
%20 = OpUndef %bool
OpLoopMerge %21 %22 None
OpBranchConditional %20 %23 %21
%23 = OpLabel
%24 = OpLoad %v4float %10
%25 = OpCopyObject %v4float %24
%26 = OpUndef %bool
OpBranch %22
%22 = OpLabel
OpBranch %19
%21 = OpLabel
OpBranch %12
%18 = OpLabel
OpBranch %13
%13 = OpLabel
OpBranch %11
%12 = OpLabel
%27 = OpLoad %v4float %10
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, false);
}
Handle id overflow in the ssa rewriter. (#2845) * Handle id overflow in the ssa rewriter. Remove LocalSSAElim pass at the same time. It does the same thing as the SSARewrite pass. Then even share almost all of the same code. Fixes crbug.com/997246
2019-09-10 13:38:23 +00:00
TEST_F(LocalSSAElimTest, Overflowtest1) {
// Check that the copy object get the undef value implicitly assigned in the
// entry block.
const std::string text = R"(
OpCapability Geometry
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %4 "P2Mai" %12 %17
OpExecutionMode %4 OriginUpperLeft
%2 = OpTypeVoid
%3 = OpTypeFunction %2
%6 = OpTypeFloat 32
%7 = OpTypeVector %6 4
%11 = OpTypePointer Input %7
%16 = OpTypePointer Output %7
%23 = OpTypePointer Function %7
%12 = OpVariable %11 Input
%17 = OpVariable %16 Output
%4 = OpFunction %2 None %3
%2177 = OpLabel
%4194302 = OpVariable %23 Function
%4194301 = OpLoad %7 %4194302
OpStore %17 %4194301
OpReturn
OpFunctionEnd
)";
SetAssembleOptions(SPV_TEXT_TO_BINARY_OPTION_PRESERVE_NUMERIC_IDS);
std::vector<Message> messages = {
{SPV_MSG_ERROR, "", 0, 0, "ID overflow. Try running compact-ids."}};
SetMessageConsumer(GetTestMessageConsumer(messages));
auto result = SinglePassRunToBinary<SSARewritePass>(text, true);
EXPECT_EQ(Pass::Status::Failure, std::get<1>(result));
}
Handle OpConstantNull in ssa-rewrite (#3362) ssa-rewrite fails in `MemPass::GetPtr()` when the SPIR-V code contains `OpLoad` for the result id of `OpConstantNull` because of the out of index access for an operand to get the base address. This commit fixes it. Fixes #3344
2020-05-20 16:00:51 +00:00
TEST_F(LocalSSAElimTest, OpConstantNull) {
const std::string text = R"(
OpCapability Addresses
OpCapability Kernel
OpCapability Int64
OpMemoryModel Physical64 OpenCL
OpEntryPoint Kernel %4 "A"
OpSource OpenCL_C 200000
%2 = OpTypeVoid
%3 = OpTypeFunction %2
%6 = OpTypeInt 32 0
%11 = OpTypePointer CrossWorkgroup %6
%16 = OpConstantNull %11
%20 = OpConstant %6 269484031
%4 = OpFunction %2 None %3
%17 = OpLabel
%18 = OpLoad %6 %16 Aligned 536870912
%19 = OpBitwiseXor %6 %18 %20
OpStore %16 %19 Aligned 536870912
OpReturn
OpFunctionEnd
)";
SinglePassRunToBinary<SSARewritePass>(text, false);
}
[spirv-opt] debug info preservation in ssa-rewrite (#3356) Add OpenCL.DebugInfo.100 `DebugValue` instructions for store and phi instructions of local variables to provide the debugger with the updated values of local variables correctly.
2020-06-19 18:57:43 +00:00
TEST_F(LocalSSAElimTest, DebugForLoop) {
// #version 140
//
// in vec4 BC;
// out float fo;
//
// void main()
// {
// float f = 0.0;
// for (int i=0; i<4; i++) {
// f = f + BC[i];
// }
// fo = f;
// }
const std::string text = R"(
; CHECK: [[f_name:%\w+]] = OpString "f"
; CHECK: [[i_name:%\w+]] = OpString "i"
; CHECK: [[dbg_f:%\w+]] = OpExtInst %void [[ext:%\d+]] DebugLocalVariable [[f_name]]
; CHECK: [[dbg_i:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[i_name]]
; CHECK: OpStore %f %float_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] %float_0
; CHECK-NEXT: OpStore %i %int_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] %int_0
; CHECK-NOT: DebugDeclare
; CHECK: [[loop_head:%\w+]] = OpLabel
; CHECK: [[phi0:%\w+]] = OpPhi %float %float_0
; CHECK: [[phi1:%\w+]] = OpPhi %int %int_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[phi0]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] [[phi1]]
; CHECK: OpLoopMerge [[loop_merge:%\w+]] [[loop_cont:%\w+]] None
; CHECK-NEXT: OpBranch [[loop_body:%\w+]]
; CHECK-NEXT: [[loop_body]] = OpLabel
; CHECK: OpBranchConditional {{%\w+}} [[bb:%\w+]] [[loop_merge]]
; CHECK: [[bb]] = OpLabel
; CHECK: OpStore %f [[f_val:%\w+]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[f_val]]
; CHECK-NEXT: OpBranch [[loop_cont]]
; CHECK: [[loop_cont]] = OpLabel
; CHECK: OpStore %i [[i_val:%\w+]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] [[i_val]]
; CHECK-NEXT: OpBranch [[loop_head]]
; CHECK: [[loop_merge]] = OpLabel
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
OpSource GLSL 140
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
%i_name = OpString "i"
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%dbg_v4f = OpExtInst %void %ext DebugTypeVector %dbg_tf 4
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %dbg_v4f %dbg_v4f
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_f = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_v4f %src 0 0 %dbg_main FlagIsLocal
%dbg_i = OpExtInst %void %ext DebugLocalVariable %i_name %dbg_v4f %src 0 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %8
%22 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
%decl0 = OpExtInst %void %ext DebugDeclare %dbg_f %f %null_expr
%decl1 = OpExtInst %void %ext DebugDeclare %dbg_i %i %null_expr
OpBranch %23
%23 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
spirv-opt: Switch from Vulkan.DebugInfo to Shader.DebugInfo (#4493) Includes: - Shift to use of spirv-header extinst.nonsemantic.shader grammar.json - Remove extinst.nonsemantic.vulkan.debuginfo.100.grammar.json - Enable all optimizations for Shader.DebugInfo Also fixes scalar replacement to only insert DebugValue after all OpVariables. This is not necessary for OpenCL.DebugInfo, but it is for Shader.DebugInfo. Likewise, fixes Private-to-Local to insert DebugDeclare after all OpVariables. Also fixes inlining to handle FunctionDefinition which can show up after first block if early return processing happens. Co-authored-by: baldurk <baldurk@baldurk.org>
2021-09-15 18:38:53 +00:00
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%27 = OpLoad %int %i
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%30 = OpLoad %float %f
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %25
%25 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %int %i
%36 = OpIAdd %int %35 %int_1
OpStore %i %36
OpBranch %23
%24 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
TEST_F(LocalSSAElimTest, ShaderDebugForLoop) {
const std::string text = R"(
; CHECK: [[f_name:%\w+]] = OpString "f"
; CHECK: [[i_name:%\w+]] = OpString "i"
; CHECK: [[dbg_f:%\w+]] = OpExtInst %void [[ext:%\d+]] DebugLocalVariable [[f_name]]
; CHECK: [[dbg_i:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[i_name]]
; CHECK: OpStore %f %float_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] %float_0
; CHECK-NEXT: OpStore %i %int_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] %int_0
; CHECK-NOT: DebugDeclare
; CHECK: [[loop_head:%\w+]] = OpLabel
; CHECK: [[phi0:%\w+]] = OpPhi %float %float_0
; CHECK: [[phi1:%\w+]] = OpPhi %int %int_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[phi0]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] [[phi1]]
; CHECK: OpLoopMerge [[loop_merge:%\w+]] [[loop_cont:%\w+]] None
; CHECK-NEXT: OpBranch [[loop_body:%\w+]]
; CHECK-NEXT: [[loop_body]] = OpLabel
; CHECK: OpBranchConditional {{%\w+}} [[bb:%\w+]] [[loop_merge]]
; CHECK: [[bb]] = OpLabel
; CHECK: OpStore %f [[f_val:%\w+]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[f_val]]
; CHECK-NEXT: OpBranch [[loop_cont]]
; CHECK: [[loop_cont]] = OpLabel
; CHECK: OpStore %i [[i_val:%\w+]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] [[i_val]]
; CHECK-NEXT: OpBranch [[loop_head]]
; CHECK: [[loop_merge]] = OpLabel
OpCapability Shader
OpExtension "SPV_KHR_non_semantic_info"
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "NonSemantic.Shader.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
OpSource GLSL 140
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
%i_name = OpString "i"
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_1 = OpConstant %uint 1
%uint_3 = OpConstant %uint 3
%uint_4 = OpConstant %uint 4
%uint_5 = OpConstant %uint 5
%uint_10 = OpConstant %uint 10
%uint_32 = OpConstant %uint 32
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit %uint_1 %uint_4 %src %uint_5
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 %uint_3 %uint_0
%dbg_v4f = OpExtInst %void %ext DebugTypeVector %dbg_tf %uint_4
%main_ty = OpExtInst %void %ext DebugTypeFunction %uint_3 %dbg_v4f %dbg_v4f
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src %uint_0 %uint_0 %cu %main_name %uint_3 %uint_10
%dbg_f = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_v4f %src %uint_0 %uint_0 %dbg_main %uint_4
%dbg_i = OpExtInst %void %ext DebugLocalVariable %i_name %dbg_v4f %src %uint_0 %uint_0 %dbg_main %uint_4
%main = OpFunction %void None %8
%22 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
%decl0 = OpExtInst %void %ext DebugDeclare %dbg_f %f %null_expr
%decl1 = OpExtInst %void %ext DebugDeclare %dbg_i %i %null_expr
OpBranch %23
%23 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
[spirv-opt] debug info preservation in ssa-rewrite (#3356) Add OpenCL.DebugInfo.100 `DebugValue` instructions for store and phi instructions of local variables to provide the debugger with the updated values of local variables correctly.
2020-06-19 18:57:43 +00:00
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%27 = OpLoad %int %i
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%30 = OpLoad %float %f
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %25
%25 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %int %i
%36 = OpIAdd %int %35 %int_1
OpStore %i %36
OpBranch %23
%24 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
TEST_F(LocalSSAElimTest, AddDebugValueForFunctionParameterWithPhi) {
// Test the distribution of DebugValue for a parameter of an inlined function
// and the visibility of Phi instruction. The ssa-rewrite pass must add
// DebugValue for the value assignment of function argument even when it is an
// inlined function. It has to check the visibility Phi through all its value
// operands. See the DebugValue for "int i" of "foo()" in the following code.
//
// struct VS_OUTPUT {
// float4 pos : SV_POSITION;
// float4 color : COLOR;
// };
//
// float4 foo(int i, float4 pos) {
// while (i < pos.x) {
// pos = pos.x + i;
// ++i;
// }
// return pos;
// }
//
// VS_OUTPUT main(float4 pos : POSITION,
// float4 color : COLOR) {
// VS_OUTPUT vout;
// vout.pos = foo(4, pos);
// vout.color = color;
// return vout;
// }
const std::string text = R"(
OpCapability Shader
%1 = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Vertex %main "main" %in_var_POSITION %in_var_COLOR %gl_Position %out_var_COLOR
%7 = OpString "vertex.hlsl"
%8 = OpString "float"
%9 = OpString "VS_OUTPUT"
%10 = OpString "color"
%11 = OpString "pos"
%12 = OpString "int"
%13 = OpString "foo"
%14 = OpString ""
%15 = OpString "i"
%16 = OpString "main"
%17 = OpString "vout"
OpName %in_var_POSITION "in.var.POSITION"
OpName %in_var_COLOR "in.var.COLOR"
OpName %out_var_COLOR "out.var.COLOR"
OpName %main "main"
OpName %param_var_pos "param.var.pos"
OpName %param_var_color "param.var.color"
OpName %VS_OUTPUT "VS_OUTPUT"
OpMemberName %VS_OUTPUT 0 "pos"
OpMemberName %VS_OUTPUT 1 "color"
OpDecorate %gl_Position BuiltIn Position
OpDecorate %in_var_POSITION Location 0
OpDecorate %in_var_COLOR Location 1
OpDecorate %out_var_COLOR Location 0
%int = OpTypeInt 32 1
%int_4 = OpConstant %int 4
%int_0 = OpConstant %int 0
%int_1 = OpConstant %int 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%_ptr_Output_v4float = OpTypePointer Output %v4float
%void = OpTypeVoid
%uint_256 = OpConstant %uint 256
%uint_128 = OpConstant %uint 128
%uint_0 = OpConstant %uint 0
%50 = OpTypeFunction %void
%_ptr_Function_v4float = OpTypePointer Function %v4float
%VS_OUTPUT = OpTypeStruct %v4float %v4float
%_ptr_Function_int = OpTypePointer Function %int
%_ptr_Function_float = OpTypePointer Function %float
%bool = OpTypeBool
%in_var_POSITION = OpVariable %_ptr_Input_v4float Input
%in_var_COLOR = OpVariable %_ptr_Input_v4float Input
%gl_Position = OpVariable %_ptr_Output_v4float Output
%out_var_COLOR = OpVariable %_ptr_Output_v4float Output
%156 = OpExtInst %void %1 DebugInfoNone
%77 = OpExtInst %void %1 DebugExpression
%58 = OpExtInst %void %1 DebugTypeBasic %8 %uint_32 Float
%59 = OpExtInst %void %1 DebugTypeVector %58 4
%60 = OpExtInst %void %1 DebugSource %7
%61 = OpExtInst %void %1 DebugCompilationUnit 1 4 %60 HLSL
%62 = OpExtInst %void %1 DebugTypeComposite %9 Structure %60 1 8 %61 %9 %uint_256 FlagIsProtected|FlagIsPrivate %63 %64
%64 = OpExtInst %void %1 DebugTypeMember %10 %59 %60 3 10 %62 %uint_128 %uint_128 FlagIsProtected|FlagIsPrivate
%63 = OpExtInst %void %1 DebugTypeMember %11 %59 %60 2 10 %62 %uint_0 %uint_128 FlagIsProtected|FlagIsPrivate
%65 = OpExtInst %void %1 DebugTypeBasic %12 %uint_32 Signed
%66 = OpExtInst %void %1 DebugTypeFunction FlagIsProtected|FlagIsPrivate %59 %65 %59
%67 = OpExtInst %void %1 DebugFunction %13 %66 %60 6 1 %61 %14 FlagIsProtected|FlagIsPrivate 6 %156
%68 = OpExtInst %void %1 DebugLexicalBlock %60 6 31 %67
%69 = OpExtInst %void %1 DebugLexicalBlock %60 7 21 %68
%70 = OpExtInst %void %1 DebugLocalVariable %11 %59 %60 6 26 %67 FlagIsLocal 2
Set correct scope and line info for DebugValue (#4125) The existing spirv-opt `DebugInfoManager::AddDebugValueForDecl()` sets the scope and line info of the new added DebugValue using the scope and line of DebugDeclare. This is wrong because only a single DebugDeclare must exist under a scope while we have to add DebugValue for all the places where the variable's value is updated. Therefore, we have to set the scope and line of DebugValue based on the places of the variable updates. This bug makes https://github.com/google/amber/blob/main/tests/cases/debugger_hlsl_shadowed_vars.amber fail. This commit fixes the bug.
2021-01-28 17:57:35 +00:00
; CHECK: [[color_name:%\w+]] = OpString "color"
; CHECK: [[pos_name:%\w+]] = OpString "pos"
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
; CHECK: [[i_name:%\w+]] = OpString "i"
; CHECK: [[null_expr:%\w+]] = OpExtInst %void [[ext:%\w+]] DebugExpression
; CHECK: [[dbg_i:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[i_name]] {{%\w+}} {{%\w+}} 6 16 {{%\w+}} FlagIsLocal 1
Set correct scope and line info for DebugValue (#4125) The existing spirv-opt `DebugInfoManager::AddDebugValueForDecl()` sets the scope and line info of the new added DebugValue using the scope and line of DebugDeclare. This is wrong because only a single DebugDeclare must exist under a scope while we have to add DebugValue for all the places where the variable's value is updated. Therefore, we have to set the scope and line of DebugValue based on the places of the variable updates. This bug makes https://github.com/google/amber/blob/main/tests/cases/debugger_hlsl_shadowed_vars.amber fail. This commit fixes the bug.
2021-01-28 17:57:35 +00:00
; CHECK: [[dbg_color:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[color_name]] {{%\w+}} {{%\w+}} 15 23
; CHECK: [[dbg_pos:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[pos_name]] {{%\w+}} {{%\w+}} 14 23
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
%71 = OpExtInst %void %1 DebugLocalVariable %15 %65 %60 6 16 %67 FlagIsLocal 1
%72 = OpExtInst %void %1 DebugTypeFunction FlagIsProtected|FlagIsPrivate %62 %59 %59
%73 = OpExtInst %void %1 DebugFunction %16 %72 %60 14 1 %61 %14 FlagIsProtected|FlagIsPrivate 15 %156
%74 = OpExtInst %void %1 DebugLexicalBlock %60 15 38 %73
%75 = OpExtInst %void %1 DebugLocalVariable %17 %62 %60 16 13 %74 FlagIsLocal
%76 = OpExtInst %void %1 DebugLocalVariable %10 %59 %60 15 23 %73 FlagIsLocal 2
%78 = OpExtInst %void %1 DebugLocalVariable %11 %59 %60 14 23 %73 FlagIsLocal 1
%155 = OpExtInst %void %1 DebugInlinedAt 17 %74
%main = OpFunction %void None %50
%79 = OpLabel
%168 = OpExtInst %void %1 DebugScope %74
; CHECK: [[i:%\w+]] = OpVariable %_ptr_Function_int Function
%120 = OpVariable %_ptr_Function_int Function
%121 = OpVariable %_ptr_Function_v4float Function
%169 = OpExtInst %void %1 DebugNoScope
%param_var_pos = OpVariable %_ptr_Function_v4float Function
%param_var_color = OpVariable %_ptr_Function_v4float Function
Set correct scope and line info for DebugValue (#4125) The existing spirv-opt `DebugInfoManager::AddDebugValueForDecl()` sets the scope and line info of the new added DebugValue using the scope and line of DebugDeclare. This is wrong because only a single DebugDeclare must exist under a scope while we have to add DebugValue for all the places where the variable's value is updated. Therefore, we have to set the scope and line of DebugValue based on the places of the variable updates. This bug makes https://github.com/google/amber/blob/main/tests/cases/debugger_hlsl_shadowed_vars.amber fail. This commit fixes the bug.
2021-01-28 17:57:35 +00:00
OpLine %7 100 105
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
%80 = OpLoad %v4float %in_var_POSITION
OpStore %param_var_pos %80
Set correct scope and line info for DebugValue (#4125) The existing spirv-opt `DebugInfoManager::AddDebugValueForDecl()` sets the scope and line info of the new added DebugValue using the scope and line of DebugDeclare. This is wrong because only a single DebugDeclare must exist under a scope while we have to add DebugValue for all the places where the variable's value is updated. Therefore, we have to set the scope and line of DebugValue based on the places of the variable updates. This bug makes https://github.com/google/amber/blob/main/tests/cases/debugger_hlsl_shadowed_vars.amber fail. This commit fixes the bug.
2021-01-28 17:57:35 +00:00
OpNoLine
OpLine %7 200 205
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
%81 = OpLoad %v4float %in_var_COLOR
OpStore %param_var_color %81
Set correct scope and line info for DebugValue (#4125) The existing spirv-opt `DebugInfoManager::AddDebugValueForDecl()` sets the scope and line info of the new added DebugValue using the scope and line of DebugDeclare. This is wrong because only a single DebugDeclare must exist under a scope while we have to add DebugValue for all the places where the variable's value is updated. Therefore, we have to set the scope and line of DebugValue based on the places of the variable updates. This bug makes https://github.com/google/amber/blob/main/tests/cases/debugger_hlsl_shadowed_vars.amber fail. This commit fixes the bug.
2021-01-28 17:57:35 +00:00
OpNoLine
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
%170 = OpExtInst %void %1 DebugScope %73
Set correct scope and line info for DebugValue (#4125) The existing spirv-opt `DebugInfoManager::AddDebugValueForDecl()` sets the scope and line info of the new added DebugValue using the scope and line of DebugDeclare. This is wrong because only a single DebugDeclare must exist under a scope while we have to add DebugValue for all the places where the variable's value is updated. Therefore, we have to set the scope and line of DebugValue based on the places of the variable updates. This bug makes https://github.com/google/amber/blob/main/tests/cases/debugger_hlsl_shadowed_vars.amber fail. This commit fixes the bug.
2021-01-28 17:57:35 +00:00
; CHECK: OpLine {{%\w+}} 100 105
; CHECK: DebugValue [[dbg_pos]]
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
%124 = OpExtInst %void %1 DebugDeclare %78 %param_var_pos %77
Set correct scope and line info for DebugValue (#4125) The existing spirv-opt `DebugInfoManager::AddDebugValueForDecl()` sets the scope and line info of the new added DebugValue using the scope and line of DebugDeclare. This is wrong because only a single DebugDeclare must exist under a scope while we have to add DebugValue for all the places where the variable's value is updated. Therefore, we have to set the scope and line of DebugValue based on the places of the variable updates. This bug makes https://github.com/google/amber/blob/main/tests/cases/debugger_hlsl_shadowed_vars.amber fail. This commit fixes the bug.
2021-01-28 17:57:35 +00:00
; CHECK: OpLine {{%\w+}} 200 205
; CHECK: DebugValue [[dbg_color]]
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
%125 = OpExtInst %void %1 DebugDeclare %76 %param_var_color %77
Set correct scope and line info for DebugValue (#4125) The existing spirv-opt `DebugInfoManager::AddDebugValueForDecl()` sets the scope and line info of the new added DebugValue using the scope and line of DebugDeclare. This is wrong because only a single DebugDeclare must exist under a scope while we have to add DebugValue for all the places where the variable's value is updated. Therefore, we have to set the scope and line of DebugValue based on the places of the variable updates. This bug makes https://github.com/google/amber/blob/main/tests/cases/debugger_hlsl_shadowed_vars.amber fail. This commit fixes the bug.
2021-01-28 17:57:35 +00:00
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
%171 = OpExtInst %void %1 DebugScope %74
OpLine %7 17 18
; CHECK: OpStore {{%\w+}} %int_4
; CHECK: DebugValue [[dbg_i]] %int_4 [[null_expr]]
OpStore %120 %int_4
OpStore %121 %80
%172 = OpExtInst %void %1 DebugScope %67 %155
%135 = OpExtInst %void %1 DebugDeclare %71 %120 %77
%136 = OpExtInst %void %1 DebugDeclare %70 %121 %77
%173 = OpExtInst %void %1 DebugScope %68 %155
OpLine %7 7 3
OpBranch %137
%174 = OpExtInst %void %1 DebugNoScope
%137 = OpLabel
; CHECK: [[phi:%\w+]] = OpPhi %int %int_4
; CHECK: DebugValue [[dbg_i]] [[phi]] [[null_expr]]
%175 = OpExtInst %void %1 DebugScope %68 %155
OpLine %7 7 10
%138 = OpLoad %int %120
%139 = OpConvertSToF %float %138
OpLine %7 7 14
%140 = OpAccessChain %_ptr_Function_float %121 %int_0
%141 = OpLoad %float %140
OpLine %7 7 12
%142 = OpFOrdLessThan %bool %139 %141
OpLine %7 7 3
%176 = OpExtInst %void %1 DebugNoScope
OpLoopMerge %153 %152 None
OpBranchConditional %142 %143 %153
%177 = OpExtInst %void %1 DebugNoScope
%143 = OpLabel
%178 = OpExtInst %void %1 DebugScope %69 %155
OpLine %7 8 11
%144 = OpAccessChain %_ptr_Function_float %121 %int_0
%145 = OpLoad %float %144
OpLine %7 8 19
%146 = OpLoad %int %120
%147 = OpConvertSToF %float %146
OpLine %7 8 17
%148 = OpFAdd %float %145 %147
OpLine %7 8 11
%149 = OpCompositeConstruct %v4float %148 %148 %148 %148
OpLine %7 8 5
OpStore %121 %149
OpLine %7 9 5
%151 = OpIAdd %int %146 %int_1
OpLine %7 9 7
; CHECK: OpStore [[i]] [[value:%\w+]]
; CHECK: DebugValue [[dbg_i]] [[value]] [[null_expr]]
OpStore %120 %151
%179 = OpExtInst %void %1 DebugScope %68 %155
OpLine %7 10 3
OpBranch %152
%180 = OpExtInst %void %1 DebugNoScope
%152 = OpLabel
%181 = OpExtInst %void %1 DebugScope %68 %155
OpBranch %137
%182 = OpExtInst %void %1 DebugNoScope
%153 = OpLabel
%183 = OpExtInst %void %1 DebugScope %68 %155
OpLine %7 11 10
%154 = OpLoad %v4float %121
%184 = OpExtInst %void %1 DebugScope %74
%167 = OpExtInst %void %1 DebugValue %75 %154 %77 %int_0
%166 = OpExtInst %void %1 DebugValue %75 %81 %77 %int_1
OpLine %7 19 10
%165 = OpCompositeConstruct %VS_OUTPUT %154 %81
%185 = OpExtInst %void %1 DebugNoScope
%83 = OpCompositeExtract %v4float %165 0
OpStore %gl_Position %83
%84 = OpCompositeExtract %v4float %165 1
OpStore %out_var_COLOR %84
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
spirv-opt: properly preserve DebugValue indexes operand (#4022) spirv-opt has a bug that `DebugInfoManager::AddDebugValueWithIndex()` does not preserve `Indexes` operands of [DebugValue](https://www.khronos.org/registry/spir-v/specs/unified1/OpenCL.DebugInfo.100.html#DebugValue). It has to preserve all of those `Indexes` operands, but it preserves only the first index operand. This PR removes `DebugInfoManager::AddDebugValueWithIndex()` and lets the spirv-opt use `DebugInfoManager::AddDebugValueForDecl()`. `DebugInfoManager::AddDebugValueForDecl()` preserves the Indexes operand correctly.
2020-11-13 17:06:38 +00:00
TEST_F(LocalSSAElimTest, DebugValueWithIndexesInForLoop) {
// #version 140
//
// in vec4 BC;
// out float fo;
//
// struct T {
// float a;
// float f;
// };
//
// struct value {
// int x;
// int y;
// T z;
// };
//
// void main()
// {
// value v;
// v.z.f = 0.0;
// for (int i=0; i<4; i++) {
// v.z.f = v.z.f + BC[i];
// }
// fo = v.z.f;
// }
const std::string text = R"(
; CHECK: [[f_name:%\w+]] = OpString "f"
; CHECK: [[dbg_f:%\w+]] = OpExtInst %void [[ext:%\d+]] DebugLocalVariable [[f_name]]
; CHECK: OpStore %f %float_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] %float_0 [[null_expr:%\d+]] %int_2 %int_1
; CHECK-NOT: DebugDeclare
; CHECK: [[loop_head:%\w+]] = OpLabel
; CHECK: [[phi0:%\w+]] = OpPhi %float %float_0
; CHECK: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[phi0]] [[null_expr]] %int_2 %int_1
; CHECK: OpLoopMerge [[loop_merge:%\w+]] [[loop_cont:%\w+]] None
; CHECK-NEXT: OpBranch [[loop_body:%\w+]]
; CHECK-NEXT: [[loop_body]] = OpLabel
; CHECK: OpBranchConditional {{%\w+}} [[bb:%\w+]] [[loop_merge]]
; CHECK: [[bb]] = OpLabel
; CHECK: OpStore %f [[f_val:%\w+]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[f_val]] [[null_expr]] %int_2 %int_1
; CHECK-NEXT: OpBranch [[loop_cont]]
; CHECK: [[loop_cont]] = OpLabel
; CHECK: OpBranch [[loop_head]]
; CHECK: [[loop_merge]] = OpLabel
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
OpSource GLSL 140
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
%i_name = OpString "i"
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%int_2 = OpConstant %int 2
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
%deref = OpExtInst %void %ext DebugOperation Deref
%null_expr = OpExtInst %void %ext DebugExpression
%deref_expr = OpExtInst %void %ext DebugExpression %deref
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%dbg_v4f = OpExtInst %void %ext DebugTypeVector %dbg_tf 4
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %dbg_v4f %dbg_v4f
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_f = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_v4f %src 0 0 %dbg_main FlagIsLocal
%dbg_i = OpExtInst %void %ext DebugLocalVariable %i_name %dbg_v4f %src 0 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %8
%22 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
%decl0 = OpExtInst %void %ext DebugValue %dbg_f %f %deref_expr %int_2 %int_1
%decl1 = OpExtInst %void %ext DebugDeclare %dbg_i %i %null_expr
OpBranch %23
%23 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%27 = OpLoad %int %i
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%30 = OpLoad %float %f
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %25
%25 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %int %i
%36 = OpIAdd %int %35 %int_1
OpStore %i %36
OpBranch %23
%24 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
Remove DebugDeclare only for target variables in ssa-rewrite (#3511) For each local variable, ssa-rewrite should remove its DebugDeclare if and only if it is replaced by any number of DebugValues for store and phi instructions. For example, when we have two variables `a` whose DebugDeclare will be replaced to DebugValues by ssa-rewrite pass and `b` whose DebugDeclare will not be replaced, we have to remove only DebugDeclare for `a`, not `b`.
2020-07-31 14:00:30 +00:00
TEST_F(LocalSSAElimTest, PartiallyKillDebugDeclare) {
// For a reference variable e.g., int i in the following example,
// we do not propagate DebugValue for a store or phi instruction
// out of the variable's scope. In that case, we should not remove
// DebugDeclare for the variable that we did not add its DebugValue.
//
// #version 140
//
// in vec4 BC;
// out float fo;
//
// int j;
// void main()
// {
// float f = 0.0;
// for (j=0; j<4; j++) {
// int& i = j;
// f = f + BC[i];
// }
// fo = f;
// }
const std::string text = R"(
; CHECK: [[f_name:%\w+]] = OpString "f"
; CHECK: [[i_name:%\w+]] = OpString "i"
; CHECK: [[fn:%\w+]] = OpExtInst %void [[ext:%\d+]] DebugFunction
; CHECK: [[bb:%\w+]] = OpExtInst %void [[ext]] DebugLexicalBlock
; CHECK: [[dbg_f:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[f_name]] {{%\w+}} {{%\w+}} 0 0 [[fn]]
; CHECK: [[dbg_i:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[i_name]] {{%\w+}} {{%\w+}} 0 0 [[bb]]
; CHECK: OpStore %f %float_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] %float_0
; CHECK-NOT: DebugDeclare [[dbg_f]]
; CHECK: OpExtInst %void [[ext]] DebugDeclare [[dbg_i]] %j
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
OpSource GLSL 140
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
%i_name = OpString "i"
%j_name = OpString "j"
OpName %main "main"
OpName %f "f"
OpName %j "j"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%_ptr_Function_int = OpTypePointer Function %int
%_ptr_Private_int = OpTypePointer Private %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
%j = OpVariable %_ptr_Private_int Private
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%dbg_v4f = OpExtInst %void %ext DebugTypeVector %dbg_tf 4
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %dbg_v4f %dbg_v4f
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%bb = OpExtInst %void %ext DebugLexicalBlock %src 0 0 %dbg_main
%dbg_f = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_v4f %src 0 0 %dbg_main FlagIsLocal
%dbg_i = OpExtInst %void %ext DebugLocalVariable %i_name %dbg_v4f %src 0 0 %bb FlagIsLocal
%dbg_j = OpExtInst %void %ext DebugGlobalVariable %j_name %dbg_v4f %src 0 0 %dbg_main %j_name %j FlagIsPrivate
%main = OpFunction %void None %8
%22 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
OpStore %f %float_0
OpStore %j %int_0
%decl0 = OpExtInst %void %ext DebugDeclare %dbg_f %f %null_expr
OpBranch %23
%23 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%27 = OpLoad %int %j
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %bb
%decl1 = OpExtInst %void %ext DebugDeclare %dbg_i %j %null_expr
%30 = OpLoad %float %f
%31 = OpLoad %int %j
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %25
%25 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %int %j
%36 = OpIAdd %int %35 %int_1
OpStore %j %36
OpBranch %23
%24 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
[spirv-opt] debug info preservation in ssa-rewrite (#3356) Add OpenCL.DebugInfo.100 `DebugValue` instructions for store and phi instructions of local variables to provide the debugger with the updated values of local variables correctly.
2020-06-19 18:57:43 +00:00
TEST_F(LocalSSAElimTest, DebugValueForReferenceVariable) {
// #version 140
//
// in vec4 BC;
// out float fo;
//
// void main()
// {
// float f = 0.0;
// float& x = f;
// for (int i=0; i<4; i++) {
// x = x + BC[i];
// }
// fo = f;
// }
const std::string text = R"(
; CHECK: [[f_name:%\w+]] = OpString "f"
; CHECK: [[i_name:%\w+]] = OpString "i"
; CHECK: [[x_name:%\w+]] = OpString "x"
; CHECK: [[dbg_f:%\w+]] = OpExtInst %void [[ext:%\d+]] DebugLocalVariable [[f_name]]
; CHECK: [[dbg_i:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[i_name]]
; CHECK: [[dbg_x:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[x_name]]
; CHECK: OpStore %f %float_0
Clear debug information for kill and replacement (#3459) For many spirv-opt passes such as simplify-instructions pass, we have to correctly clear the OpenCL.DebugInfo.100 debug information for KillInst() and ReplaceAllUses(). If we keep some debug information that disappeared because of KillInst() and ReplaceAllUses(), adding new DebugValue instructions based on the existing DebugDeclare information will generate incorrect information. This CL update DebugInfoManager and IRContext to correctly clear debug information.
2020-06-25 19:48:26 +00:00
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_f]] %float_0
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_x]] %float_0
; CHECK: OpStore %i %int_0
[spirv-opt] debug info preservation in ssa-rewrite (#3356) Add OpenCL.DebugInfo.100 `DebugValue` instructions for store and phi instructions of local variables to provide the debugger with the updated values of local variables correctly.
2020-06-19 18:57:43 +00:00
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] %int_0
; CHECK-NOT: DebugDeclare
; CHECK: [[loop_head:%\w+]] = OpLabel
; CHECK: [[phi0:%\w+]] = OpPhi %float %float_0
; CHECK: [[phi1:%\w+]] = OpPhi %int %int_0
Clear debug information for kill and replacement (#3459) For many spirv-opt passes such as simplify-instructions pass, we have to correctly clear the OpenCL.DebugInfo.100 debug information for KillInst() and ReplaceAllUses(). If we keep some debug information that disappeared because of KillInst() and ReplaceAllUses(), adding new DebugValue instructions based on the existing DebugDeclare information will generate incorrect information. This CL update DebugInfoManager and IRContext to correctly clear debug information.
2020-06-25 19:48:26 +00:00
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[phi0]]
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_x]] [[phi0]]
; CHECK: OpExtInst %void [[ext]] DebugValue [[dbg_i]] [[phi1]]
[spirv-opt] debug info preservation in ssa-rewrite (#3356) Add OpenCL.DebugInfo.100 `DebugValue` instructions for store and phi instructions of local variables to provide the debugger with the updated values of local variables correctly.
2020-06-19 18:57:43 +00:00
; CHECK: OpLoopMerge [[loop_merge:%\w+]] [[loop_cont:%\w+]] None
; CHECK-NEXT: OpBranch [[loop_body:%\w+]]
; CHECK: [[loop_body]] = OpLabel
; CHECK: OpBranchConditional {{%\w+}} [[bb:%\w+]] [[loop_merge]]
; CHECK: [[bb]] = OpLabel
; CHECK: OpStore %f [[f_val:%\w+]]
Clear debug information for kill and replacement (#3459) For many spirv-opt passes such as simplify-instructions pass, we have to correctly clear the OpenCL.DebugInfo.100 debug information for KillInst() and ReplaceAllUses(). If we keep some debug information that disappeared because of KillInst() and ReplaceAllUses(), adding new DebugValue instructions based on the existing DebugDeclare information will generate incorrect information. This CL update DebugInfoManager and IRContext to correctly clear debug information.
2020-06-25 19:48:26 +00:00
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[f_val]]
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_x]] [[f_val]]
; CHECK: OpBranch [[loop_cont]]
[spirv-opt] debug info preservation in ssa-rewrite (#3356) Add OpenCL.DebugInfo.100 `DebugValue` instructions for store and phi instructions of local variables to provide the debugger with the updated values of local variables correctly.
2020-06-19 18:57:43 +00:00
; CHECK: [[loop_cont]] = OpLabel
; CHECK: OpStore %i [[i_val:%\w+]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] [[i_val]]
; CHECK-NEXT: OpBranch [[loop_head]]
; CHECK: [[loop_merge]] = OpLabel
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
OpSource GLSL 140
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
%i_name = OpString "i"
%x_name = OpString "x"
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%dbg_v4f = OpExtInst %void %ext DebugTypeVector %dbg_tf 4
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %dbg_v4f %dbg_v4f
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_f = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_v4f %src 0 0 %dbg_main FlagIsLocal
%dbg_i = OpExtInst %void %ext DebugLocalVariable %i_name %dbg_v4f %src 1 0 %dbg_main FlagIsLocal
%dbg_x = OpExtInst %void %ext DebugLocalVariable %x_name %dbg_v4f %src 2 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %8
%22 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
%decl0 = OpExtInst %void %ext DebugDeclare %dbg_f %f %null_expr
%decl1 = OpExtInst %void %ext DebugDeclare %dbg_i %i %null_expr
%decl2 = OpExtInst %void %ext DebugDeclare %dbg_x %f %null_expr
OpBranch %23
%23 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%27 = OpLoad %int %i
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%30 = OpLoad %float %f
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %25
%25 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %int %i
%36 = OpIAdd %int %35 %int_1
OpStore %i %36
OpBranch %23
%24 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
TEST_F(LocalSSAElimTest, DebugValueForReferenceVariableInBB) {
// #version 140
//
// in vec4 BC;
// out float fo;
//
// void main()
// {
// float f = 0.0;
// for (int i=0; i<4; i++) {
// float& x = f;
// x = x + BC[i];
// {
// x = x + BC[i];
// }
// }
// fo = f;
// }
const std::string text = R"(
; CHECK: [[f_name:%\w+]] = OpString "f"
; CHECK: [[i_name:%\w+]] = OpString "i"
; CHECK: [[x_name:%\w+]] = OpString "x"
; CHECK: [[dbg_main:%\w+]] = OpExtInst %void [[ext:%\d+]] DebugFunction
; CHECK: [[dbg_bb:%\w+]] = OpExtInst %void [[ext]] DebugLexicalBlock
; CHECK: [[dbg_bb_child:%\w+]] = OpExtInst %void [[ext]] DebugLexicalBlock
; CHECK: [[dbg_f:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[f_name]]
; CHECK: [[dbg_i:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[i_name]]
; CHECK: [[dbg_x:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[x_name]]
; CHECK: OpExtInst %void [[ext]] DebugScope [[dbg_main]]
; CHECK: OpStore %f %float_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] %float_0
; CHECK-NEXT: OpStore %i %int_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] %int_0
; CHECK-NOT: DebugDeclare
; CHECK: [[loop_head:%\w+]] = OpLabel
; CHECK: OpExtInst %void [[ext]] DebugScope [[dbg_main]]
; CHECK: [[phi0:%\w+]] = OpPhi %float %float_0
; CHECK: [[phi1:%\w+]] = OpPhi %int %int_0
Add DebugValue for DebugDecl invisible to value assignment (#3973) For some cases, we have DebugDecl invisible to a value assignment, but the value assignment information is important i.e., debugger cannot inspect the variable without the information. For example, a parameter of an inlined function must have its value assignment i.e., argument passing out of its function scope. If we simply remove DebugDecl because it is invisible to the argument passing, we cannot inspec the variable. This PR - Adds DebugValue for DebugDecl invisible to a value assignment. We use the value of the variable in the basic block that contains DebugDecl, which is found by ssa-rewrite. If the value instruction does not dominate DebugDecl, we use the value of the variable in the immediate dominator of the basic block. - Checks the visibility of DebugDecl for Phi value assignment based on the all value operands of the Phi. Since Phi just references multiple values from multiple basic blocks, scopes of value operands must be regarded as the scope of the Phi.
2020-10-27 19:10:08 +00:00
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[phi0]]
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_x]] [[phi0]]
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_i]] [[phi1]]
[spirv-opt] debug info preservation in ssa-rewrite (#3356) Add OpenCL.DebugInfo.100 `DebugValue` instructions for store and phi instructions of local variables to provide the debugger with the updated values of local variables correctly.
2020-06-19 18:57:43 +00:00
; CHECK: OpLoopMerge [[loop_merge:%\w+]] [[loop_cont:%\w+]] None
; CHECK-NEXT: OpBranch [[loop_body:%\w+]]
; CHECK-NEXT: [[loop_body]] = OpLabel
; CHECK: OpBranchConditional {{%\w+}} [[bb:%\w+]] [[loop_merge]]
; CHECK: [[bb]] = OpLabel
; CHECK: OpExtInst %void [[ext]] DebugScope [[dbg_bb]]
; CHECK: OpStore %f [[f_val:%\w+]]
Clear debug information for kill and replacement (#3459) For many spirv-opt passes such as simplify-instructions pass, we have to correctly clear the OpenCL.DebugInfo.100 debug information for KillInst() and ReplaceAllUses(). If we keep some debug information that disappeared because of KillInst() and ReplaceAllUses(), adding new DebugValue instructions based on the existing DebugDeclare information will generate incorrect information. This CL update DebugInfoManager and IRContext to correctly clear debug information.
2020-06-25 19:48:26 +00:00
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[f_val]]
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_x]] [[f_val]]
; CHECK: OpBranch [[bb_child:%\w+]]
[spirv-opt] debug info preservation in ssa-rewrite (#3356) Add OpenCL.DebugInfo.100 `DebugValue` instructions for store and phi instructions of local variables to provide the debugger with the updated values of local variables correctly.
2020-06-19 18:57:43 +00:00
; CHECK: [[bb_child]] = OpLabel
; CHECK: OpExtInst %void [[ext]] DebugScope [[dbg_bb_child]]
; CHECK: OpStore %f [[new_f_val:%\w+]]
Clear debug information for kill and replacement (#3459) For many spirv-opt passes such as simplify-instructions pass, we have to correctly clear the OpenCL.DebugInfo.100 debug information for KillInst() and ReplaceAllUses(). If we keep some debug information that disappeared because of KillInst() and ReplaceAllUses(), adding new DebugValue instructions based on the existing DebugDeclare information will generate incorrect information. This CL update DebugInfoManager and IRContext to correctly clear debug information.
2020-06-25 19:48:26 +00:00
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[new_f_val]]
; CHECK-DAG: OpExtInst %void [[ext]] DebugValue [[dbg_x]] [[new_f_val]]
; CHECK: OpBranch [[loop_cont]]
[spirv-opt] debug info preservation in ssa-rewrite (#3356) Add OpenCL.DebugInfo.100 `DebugValue` instructions for store and phi instructions of local variables to provide the debugger with the updated values of local variables correctly.
2020-06-19 18:57:43 +00:00
; CHECK: [[loop_cont]] = OpLabel
; CHECK: OpStore %i [[i_val:%\w+]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] [[i_val]]
; CHECK-NEXT: OpBranch [[loop_head]]
; CHECK: [[loop_merge]] = OpLabel
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
OpSource GLSL 140
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
%i_name = OpString "i"
%x_name = OpString "x"
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%dbg_v4f = OpExtInst %void %ext DebugTypeVector %dbg_tf 4
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %dbg_v4f %dbg_v4f
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%bb = OpExtInst %void %ext DebugLexicalBlock %src 0 0 %dbg_main
%bb_child = OpExtInst %void %ext DebugLexicalBlock %src 1 0 %bb
%dbg_f = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_v4f %src 0 0 %dbg_main FlagIsLocal
%dbg_i = OpExtInst %void %ext DebugLocalVariable %i_name %dbg_v4f %src 1 0 %dbg_main FlagIsLocal
%dbg_x = OpExtInst %void %ext DebugLocalVariable %x_name %dbg_v4f %src 2 0 %bb FlagIsLocal
%main = OpFunction %void None %8
%22 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
%decl0 = OpExtInst %void %ext DebugDeclare %dbg_f %f %null_expr
%decl1 = OpExtInst %void %ext DebugDeclare %dbg_i %i %null_expr
OpBranch %23
%23 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%27 = OpLoad %int %i
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%scope = OpExtInst %void %ext DebugScope %bb
%decl2 = OpExtInst %void %ext DebugDeclare %dbg_x %f %null_expr
%30 = OpLoad %float %f
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %38
%38 = OpLabel
%child_scope = OpExtInst %void %ext DebugScope %bb_child
%39 = OpLoad %float %f
%40 = OpFAdd %float %39 %33
OpStore %f %40
OpBranch %25
%25 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %int %i
%36 = OpIAdd %int %35 %int_1
OpStore %i %36
OpBranch %23
%24 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
TEST_F(LocalSSAElimTest, DebugForLoopUseDebugValueInsteadOfDebugDeclare) {
// #version 140
//
// in vec4 BC;
// out float fo;
//
// struct S {
// float f;
// int i;
// };
//
// void main()
// {
// S foo = {0.0, 0};
// for (; foo.i<4; foo.i++) {
// foo.f = foo.f + BC[foo.i];
// }
// fo = foo.f;
// }
const std::string text = R"(
; CHECK: [[f_name:%\w+]] = OpString "f"
; CHECK: [[empty_expr:%\w+]] = OpExtInst %void [[ext:%\d+]] DebugExpression
; CHECK: [[deref_op:%\w+]] = OpExtInst %void [[ext]] DebugOperation Deref
; CHECK: [[deref:%\w+]] = OpExtInst %void [[ext]] DebugExpression [[deref_op]]
; CHECK: [[dbg_foo:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[f_name]]
; CHECK: OpStore %f %float_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_foo]] %float_0 [[empty_expr]] %uint_0
; CHECK-NEXT: OpStore %i %int_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_foo]] %int_0 [[empty_expr]] %uint_1
; CHECK: [[loop_head:%\w+]] = OpLabel
; CHECK: [[phi0:%\w+]] = OpPhi %float %float_0
; CHECK: [[phi1:%\w+]] = OpPhi %int %int_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_foo]] [[phi0]] [[empty_expr]] %uint_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_foo]] [[phi1]] [[empty_expr]] %uint_1
; CHECK: OpLoopMerge [[loop_merge:%\w+]] [[loop_cont:%\w+]] None
; CHECK-NEXT: OpBranch [[loop_body:%\w+]]
; CHECK: [[loop_body]] = OpLabel
; CHECK: OpBranchConditional {{%\w+}} [[bb:%\w+]] [[loop_merge]]
; CHECK: [[bb]] = OpLabel
; CHECK: OpStore %f [[f_val:%\w+]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_foo]] [[f_val]] [[empty_expr]] %uint_0
; CHECK-NEXT: OpBranch [[loop_cont]]
; CHECK: [[loop_cont]] = OpLabel
; CHECK: OpStore %i [[i_val:%\w+]]
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_foo]] [[i_val]] [[empty_expr]] %uint_1
; CHECK-NEXT: OpBranch [[loop_head]]
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
OpSource GLSL 140
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%uint_0 = OpConstant %uint 0
%uint_1 = OpConstant %uint 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
%deref_op = OpExtInst %void %ext DebugOperation Deref
%deref = OpExtInst %void %ext DebugExpression %deref_op
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%dbg_v4f = OpExtInst %void %ext DebugTypeVector %dbg_tf 4
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %dbg_v4f %dbg_v4f
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_foo = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_v4f %src 0 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %8
%22 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
%decl0 = OpExtInst %void %ext DebugValue %dbg_foo %f %deref %uint_0
%decl1 = OpExtInst %void %ext DebugValue %dbg_foo %i %deref %uint_1
OpBranch %23
%23 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%27 = OpLoad %int %i
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%30 = OpLoad %float %f
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %25
%25 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %int %i
%36 = OpIAdd %int %35 %int_1
OpStore %i %36
OpBranch %23
%24 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
TEST_F(LocalSSAElimTest, DebugValueNotUsedForDebugDeclare) {
// #version 140
//
// in vec4 BC;
// out float fo;
//
// void main()
// {
// float f = 0.0;
// for (int i=0; i<4; i++) {
// f = f + BC[i];
// }
// fo = f;
// }
const std::string text = R"(
; CHECK: [[f_name:%\w+]] = OpString "f"
; CHECK: [[i_name:%\w+]] = OpString "i"
; CHECK: [[dbg_f:%\w+]] = OpExtInst %void [[ext:%\d+]] DebugLocalVariable [[f_name]]
; CHECK: [[dbg_i:%\w+]] = OpExtInst %void [[ext]] DebugLocalVariable [[i_name]]
; CHECK: OpStore %f %float_0
; CHECK-NEXT: OpStore %i %int_0
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_f]] %f
; CHECK-NEXT: OpExtInst %void [[ext]] DebugValue [[dbg_i]] %i
; CHECK-NOT: DebugValue
; CHECK-NOT: DebugDeclare
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
OpSource GLSL 140
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
%i_name = OpString "i"
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%dbg_v4f = OpExtInst %void %ext DebugTypeVector %dbg_tf 4
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %dbg_v4f %dbg_v4f
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_f = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_v4f %src 0 0 %dbg_main FlagIsLocal
%dbg_i = OpExtInst %void %ext DebugLocalVariable %i_name %dbg_v4f %src 1 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %8
%22 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
OpStore %f %float_0
OpStore %i %int_0
%decl0 = OpExtInst %void %ext DebugValue %dbg_f %f %null_expr
%decl1 = OpExtInst %void %ext DebugValue %dbg_i %i %null_expr
OpBranch %23
%23 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
OpLoopMerge %24 %25 None
OpBranch %26
%26 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%27 = OpLoad %int %i
%28 = OpSLessThan %bool %27 %int_4
OpBranchConditional %28 %29 %24
%29 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%30 = OpLoad %float %f
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
%34 = OpFAdd %float %30 %33
OpStore %f %34
OpBranch %25
%25 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %int %i
%36 = OpIAdd %int %35 %int_1
OpStore %i %36
OpBranch %23
%24 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %float %f
OpStore %fo %37
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
TEST_F(LocalSSAElimTest, DebugNestedForLoop) {
const std::string text = R"(
; CHECK: = OpFunction
; CHECK-NEXT: [[entry:%\w+]] = OpLabel
; CHECK: OpStore %f %float_0
; CHECK-NEXT: = OpExtInst %void [[ext:%\w+]] DebugValue [[dbg_f:%\w+]] %float_0
; CHECK: [[outer_header:%\w+]] = OpLabel
; CHECK: [[outer_f:%\w+]] = OpPhi %float %float_0 [[entry]] [[inner_f:%\w+]] [[outer_be:%\w+]]
; CHECK: = OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[outer_f]]
; CHECK: [[inner_pre_header:%\w+]] = OpLabel
; CHECK: [[inner_header:%\w+]] = OpLabel
; CHECK: [[inner_f]] = OpPhi %float [[outer_f]] [[inner_pre_header]] [[f_next:%\w+]] [[inner_be:%\w+]]
; CHECK: = OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[inner_f]]
; CHECK: [[inner_be]] = OpLabel
; CHECK: [[f_next]] = OpFAdd %float [[inner_f]]
; CHECK-NEXT: OpStore %f [[f_next]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[f_next]]
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
OpSource GLSL 450
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %j "j"
OpName %BC "BC"
OpName %fo "fo"
OpDecorate %BC Location 0
OpDecorate %fo Location 0
%void = OpTypeVoid
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%mat4v4float = OpTypeMatrix %v4float 4
%_ptr_Input_mat4v4float = OpTypePointer Input %mat4v4float
%BC = OpVariable %_ptr_Input_mat4v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
; Debug information
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %void
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_f = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_tf %src 0 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %9
%24 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
%j = OpVariable %_ptr_Function_int Function
; DebugDeclare
OpStore %f %float_0
%decl = OpExtInst %void %ext DebugDeclare %dbg_f %f %null_expr
OpStore %i %int_0
OpBranch %25
%25 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%26 = OpLoad %int %i
%27 = OpSLessThan %bool %26 %int_4
OpLoopMerge %28 %29 None
OpBranchConditional %27 %30 %28
%30 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
OpStore %j %int_0
OpBranch %31
%31 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%32 = OpLoad %int %j
%33 = OpSLessThan %bool %32 %int_4
OpLoopMerge %50 %34 None
OpBranchConditional %33 %34 %50
%34 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %float %f
%36 = OpLoad %int %i
%37 = OpLoad %int %j
%38 = OpAccessChain %_ptr_Input_float %BC %36 %37
%39 = OpLoad %float %38
%40 = OpFAdd %float %35 %39
OpStore %f %40
%41 = OpLoad %int %j
%42 = OpIAdd %int %41 %int_1
OpStore %j %42
OpBranch %31
%50 = OpLabel
%s6 = OpExtInst %void %ext DebugScope %dbg_main
OpBranch %29
%29 = OpLabel
%s7 = OpExtInst %void %ext DebugScope %dbg_main
%43 = OpLoad %int %i
%44 = OpIAdd %int %43 %int_1
OpStore %i %44
OpBranch %25
%28 = OpLabel
%s8 = OpExtInst %void %ext DebugScope %dbg_main
%45 = OpLoad %float %f
OpStore %fo %45
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
TEST_F(LocalSSAElimTest, DebugForLoopWithContinue) {
const std::string text = R"(
; CHECK: = OpFunction
; CHECK-NEXT: [[entry:%\w+]] = OpLabel
; CHECK: OpStore %f %float_0
; CHECK-NEXT: = OpExtInst %void [[ext:%\w+]] DebugValue [[dbg_f:%\w+]] %float_0
; CHECK: [[outer_header:%\w+]] = OpLabel
; CHECK: [[outer_f:%\w+]] = OpPhi %float %float_0 [[entry]] [[inner_f:%\w+]] [[cont:%\w+]]
; CHECK: = OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[outer_f]]
; CHECK: [[f_next:%\w+]] = OpFAdd %float [[outer_f]]
; CHECK-NEXT: OpStore %f [[f_next]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[f_next]]
; CHECK: [[cont]] = OpLabel
; CHECK: [[inner_f]] = OpPhi %float [[outer_f]] {{%\d+}} [[f_next]] {{%\d+}}
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_f]] [[inner_f]]
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BC %fo
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
%float_name = OpString "float"
%main_name = OpString "main"
%f_name = OpString "f"
OpSource GLSL 140
OpName %main "main"
OpName %f "f"
OpName %i "i"
OpName %t "t"
OpName %BC "BC"
OpName %fo "fo"
%void = OpTypeVoid
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%int_0 = OpConstant %int 0
%int_4 = OpConstant %int 4
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BC = OpVariable %_ptr_Input_v4float Input
%_ptr_Input_float = OpTypePointer Input %float
%int_1 = OpConstant %int 1
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
; Debug information
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %void
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_f = OpExtInst %void %ext DebugLocalVariable %f_name %dbg_tf %src 0 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %9
%23 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f = OpVariable %_ptr_Function_float Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
; DebugDeclare
OpStore %f %float_0
%decl = OpExtInst %void %ext DebugDeclare %dbg_f %f %null_expr
OpStore %i %int_0
OpBranch %24
%24 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
OpLoopMerge %25 %26 None
OpBranch %27
%27 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%28 = OpLoad %int %i
%29 = OpSLessThan %bool %28 %int_4
OpBranchConditional %29 %30 %25
%30 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%31 = OpLoad %int %i
%32 = OpAccessChain %_ptr_Input_float %BC %31
%33 = OpLoad %float %32
OpStore %t %33
%34 = OpLoad %float %t
%35 = OpFOrdLessThan %bool %34 %float_0
OpSelectionMerge %36 None
OpBranchConditional %35 %37 %36
%37 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
OpBranch %26
%36 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%38 = OpLoad %float %f
%39 = OpLoad %float %t
%40 = OpFAdd %float %38 %39
OpStore %f %40
OpBranch %26
%26 = OpLabel
%s6 = OpExtInst %void %ext DebugScope %dbg_main
%41 = OpLoad %int %i
%42 = OpIAdd %int %41 %int_1
OpStore %i %42
OpBranch %24
%25 = OpLabel
%s7 = OpExtInst %void %ext DebugScope %dbg_main
%43 = OpLoad %float %f
OpStore %fo %43
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
TEST_F(LocalSSAElimTest, DebugIfElse) {
const std::string text = R"(
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %f %BaseColor %gl_FragColor
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
%float_name = OpString "float"
%main_name = OpString "main"
%v_name = OpString "v"
OpSource GLSL 140
OpName %main "main"
OpName %f "f"
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %gl_FragColor "gl_FragColor"
%void = OpTypeVoid
%8 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Input_float = OpTypePointer Input %float
%f = OpVariable %_ptr_Input_float Input
%float_0 = OpConstant %float 0
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%bool = OpTypeBool
%v4float = OpTypeVector %float 4
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%float_0_5 = OpConstant %float 0.5
%float_1 = OpConstant %float 1
%18 = OpConstantComposite %v4float %float_1 %float_1 %float_1 %float_1
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
; Debug information
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %void
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_v = OpExtInst %void %ext DebugLocalVariable %v_name %dbg_tf %src 0 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %8
%20 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
; DebugDeclare
%v = OpVariable %_ptr_Function_v4float Function
%decl = OpExtInst %void %ext DebugDeclare %dbg_v %v %null_expr
%21 = OpLoad %float %f
%22 = OpFOrdGreaterThanEqual %bool %21 %float_0
OpSelectionMerge %23 None
OpBranchConditional %22 %24 %25
; CHECK: OpBranchConditional
; CHECK-NEXT: [[br0:%\w+]] = OpLabel
; CHECK: OpStore %v [[v0:%\w+]]
; CHECK-NEXT: = OpExtInst %void [[ext:%\w+]] DebugValue [[dbg_v:%\w+]] [[v0]]
%24 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
%26 = OpLoad %v4float %BaseColor
%27 = OpVectorTimesScalar %v4float %26 %float_0_5
OpStore %v %27
OpBranch %23
; CHECK: [[br1:%\w+]] = OpLabel
; CHECK: OpStore %v [[v1:%\w+]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_v]] [[v1]]
%25 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%28 = OpLoad %v4float %BaseColor
%29 = OpFAdd %v4float %28 %18
OpStore %v %29
OpBranch %23
; CHECK: [[phi:%\w+]] = OpPhi %v4float [[v0]] [[br0]] [[v1]] [[br1]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_v]] [[phi]]
%23 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%30 = OpLoad %v4float %v
OpStore %gl_FragColor %30
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
TEST_F(LocalSSAElimTest, DebugSwitch) {
const std::string text = R"(
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %BaseColor %f %gl_FragColor
OpExecutionMode %main OriginUpperLeft
%file_name = OpString "test"
%float_name = OpString "float"
%main_name = OpString "main"
%v_name = OpString "v"
OpSource GLSL 140
OpName %main "main"
OpName %v "v"
OpName %BaseColor "BaseColor"
OpName %i "i"
OpName %f "f"
OpName %gl_FragColor "gl_FragColor"
%void = OpTypeVoid
%9 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%BaseColor = OpVariable %_ptr_Input_v4float Input
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%_ptr_Input_float = OpTypePointer Input %float
%f = OpVariable %_ptr_Input_float Input
%float_0_25 = OpConstant %float 0.25
%float_0_75 = OpConstant %float 0.75
%_ptr_Output_v4float = OpTypePointer Output %v4float
%gl_FragColor = OpVariable %_ptr_Output_v4float Output
; Debug information
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %void
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_v = OpExtInst %void %ext DebugLocalVariable %v_name %dbg_tf %src 0 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %9
%20 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%v = OpVariable %_ptr_Function_v4float Function
%i = OpVariable %_ptr_Function_int Function
%21 = OpLoad %v4float %BaseColor
; DebugDeclare
OpStore %v %21
%decl = OpExtInst %void %ext DebugDeclare %dbg_v %v %null_expr
; CHECK: %main = OpFunction %void None
; CHECK-NEXT: [[entry:%\w+]] = OpLabel
; CHECK: OpStore %v [[v0:%\w+]]
; CHECK-NEXT: = OpExtInst %void [[ext:%\w+]] DebugValue [[dbg_v:%\w+]] [[v0]]
; CHECK: OpSwitch {{%\w+}} [[case0:%\w+]] 0 [[case1:%\w+]] 1 [[case2:%\w+]] 2 [[case3:%\w+]]
; CHECK: OpStore %v [[v1:%\w+]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_v]] [[v1]]
; CHECK: OpStore %v [[v2:%\w+]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_v]] [[v2]]
; CHECK: [[phi0:%\w+]] = OpPhi %v4float [[v0]] [[entry]] [[v2]] [[case2]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_v]] [[phi0]]
; CHECK: OpStore %v [[v3:%\w+]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_v]] [[v3]]
; CHECK: [[phi1:%\w+]] = OpPhi %v4float [[v0]] [[case0]] [[v1]] [[case1]] [[v3]] [[case3]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_v]] [[phi1]]
%22 = OpLoad %float %f
%23 = OpConvertFToS %int %22
OpStore %i %23
%24 = OpLoad %int %i
OpSelectionMerge %25 None
OpSwitch %24 %26 0 %27 1 %28 2 %29
%26 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
OpBranch %25
%27 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%30 = OpLoad %v4float %v
%31 = OpVectorTimesScalar %v4float %30 %float_0_25
OpStore %v %31
OpBranch %25
%28 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%32 = OpLoad %v4float %v
%33 = OpCompositeConstruct %v4float %float_0_25 %float_0_25 %float_0_25 %float_0_25
%34 = OpFAdd %v4float %32 %33
OpStore %v %34
OpBranch %29
%29 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%35 = OpLoad %v4float %v
%36 = OpVectorTimesScalar %v4float %35 %float_0_75
OpStore %v %36
OpBranch %25
%25 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %v4float %v
OpStore %gl_FragColor %37
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
TEST_F(LocalSSAElimTest, DebugSwapProblem) {
// #version 140
//
// in float fe;
// out float fo;
//
// void main()
// {
// float f1 = 0.0;
// float f2 = 1.0;
// int ie = int(fe);
// for (int i=0; i<ie; i++) {
// float t = f1;
// f1 = f2;
// f2 = t;
// }
// fo = f1;
// }
//
// Because of the swap in the for loop, it generates the following phi
// instructions:
//
// [[phi_f2]] = OpPhi %float %float_1 [[entry]] [[phi_f1]] ..
// [[phi_f1]] = OpPhi %float %float_0 [[entry]] [[phi_f2]] ..
//
// Since they are used as operands by each other, we want to clearly check
// what DebugValue we have to add for them.
const std::string text = R"(
OpCapability Shader
%1 = OpExtInstImport "GLSL.std.450"
%ext = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %main "main" %fe %fo
OpExecutionMode %main OriginUpperLeft
OpSource GLSL 140
%file_name = OpString "test"
%float_name = OpString "float"
%main_name = OpString "main"
%t_name = OpString "t"
OpName %main "main"
OpName %f1 "f1"
OpName %f2 "f2"
OpName %ie "ie"
OpName %fe "fe"
OpName %i "i"
OpName %t "t"
OpName %fo "fo"
%void = OpTypeVoid
%11 = OpTypeFunction %void
%float = OpTypeFloat 32
%_ptr_Function_float = OpTypePointer Function %float
%float_0 = OpConstant %float 0
%float_1 = OpConstant %float 1
%int = OpTypeInt 32 1
%_ptr_Function_int = OpTypePointer Function %int
%_ptr_Input_float = OpTypePointer Input %float
%fe = OpVariable %_ptr_Input_float Input
%int_0 = OpConstant %int 0
%bool = OpTypeBool
%int_1 = OpConstant %int 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%_ptr_Output_float = OpTypePointer Output %float
%fo = OpVariable %_ptr_Output_float Output
; Debug information
%null_expr = OpExtInst %void %ext DebugExpression
%src = OpExtInst %void %ext DebugSource %file_name
%cu = OpExtInst %void %ext DebugCompilationUnit 1 4 %src HLSL
%dbg_tf = OpExtInst %void %ext DebugTypeBasic %float_name %uint_32 Float
%main_ty = OpExtInst %void %ext DebugTypeFunction FlagIsProtected|FlagIsPrivate %void
%dbg_main = OpExtInst %void %ext DebugFunction %main_name %main_ty %src 0 0 %cu %main_name FlagIsProtected|FlagIsPrivate 10 %main
%dbg_f1 = OpExtInst %void %ext DebugLocalVariable %t_name %dbg_tf %src 0 0 %dbg_main FlagIsLocal
%dbg_f2 = OpExtInst %void %ext DebugLocalVariable %t_name %dbg_tf %src 0 0 %dbg_main FlagIsLocal
%dbg_i = OpExtInst %void %ext DebugLocalVariable %t_name %dbg_tf %src 0 0 %dbg_main FlagIsLocal
%main = OpFunction %void None %11
%23 = OpLabel
%s0 = OpExtInst %void %ext DebugScope %dbg_main
%f1 = OpVariable %_ptr_Function_float Function
%f2 = OpVariable %_ptr_Function_float Function
%ie = OpVariable %_ptr_Function_int Function
%i = OpVariable %_ptr_Function_int Function
%t = OpVariable %_ptr_Function_float Function
OpStore %f1 %float_0
OpStore %f2 %float_1
%24 = OpLoad %float %fe
%25 = OpConvertFToS %int %24
OpStore %ie %25
OpStore %i %int_0
; DebugDeclare
%decl0 = OpExtInst %void %ext DebugDeclare %dbg_f1 %f1 %null_expr
%decl1 = OpExtInst %void %ext DebugDeclare %dbg_f2 %f2 %null_expr
%decl2 = OpExtInst %void %ext DebugDeclare %dbg_i %i %null_expr
; CHECK: %main = OpFunction %void None
; CHECK-NEXT: [[entry:%\w+]] = OpLabel
; CHECK: OpStore %f1 %float_0
; CHECK-NEXT: = OpExtInst %void [[ext:%\w+]] DebugValue [[dbg_f1:%\w+]] %float_0
; CHECK: OpStore %f2 %float_1
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_f2:%\w+]] %float_1
; CHECK: OpStore %i %int_0
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_i:%\w+]] %int_0
; CHECK: [[phi_f2:%\w+]] = OpPhi %float %float_1 [[entry]] [[phi_f1:%\w+]]
; CHECK: [[phi_f1]] = OpPhi %float %float_0 [[entry]] [[phi_f2]]
; CHECK: [[phi_i:%\w+]] = OpPhi %int %int_0 [[entry]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_f2]] [[phi_f2]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_f1]] [[phi_f1]]
; CHECK-NEXT: = OpExtInst %void [[ext]] DebugValue [[dbg_i]] [[phi_i]]
OpBranch %26
%26 = OpLabel
%s1 = OpExtInst %void %ext DebugScope %dbg_main
OpLoopMerge %27 %28 None
OpBranch %29
%29 = OpLabel
%s2 = OpExtInst %void %ext DebugScope %dbg_main
%30 = OpLoad %int %i
%31 = OpLoad %int %ie
%32 = OpSLessThan %bool %30 %31
OpBranchConditional %32 %33 %27
%33 = OpLabel
%s3 = OpExtInst %void %ext DebugScope %dbg_main
%34 = OpLoad %float %f1
OpStore %t %34
%35 = OpLoad %float %f2
OpStore %f1 %35
%36 = OpLoad %float %t
OpStore %f2 %36
OpBranch %28
%28 = OpLabel
%s4 = OpExtInst %void %ext DebugScope %dbg_main
%37 = OpLoad %int %i
%38 = OpIAdd %int %37 %int_1
OpStore %i %38
OpBranch %26
%27 = OpLabel
%s5 = OpExtInst %void %ext DebugScope %dbg_main
%39 = OpLoad %float %f1
OpStore %fo %39
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
Fix SSA-rewrite to remove DebugDeclare for variables without loads (#3719)
2020-08-24 19:33:01 +00:00
TEST_F(LocalSSAElimTest, RemoveDebugDeclareWithoutLoads) {
// Check that the DebugDeclare for c is removed even though its loads
// had been removed previously by single block store/load optimization.
// In the presence of DebugDeclare, single-block can and does remove loads,
// but cannot change the stores into DebugValues and remove the DebugDeclare
// because it is only a per block optimization, not a function optimization.
// So SSA-rewrite must perform this role.
//
// Texture2D g_tColor;
// SamplerState g_sAniso;
//
// struct PS_INPUT
// {
// float2 vTextureCoords2 : TEXCOORD2;
// float2 vTextureCoords3 : TEXCOORD3;
// };
//
// struct PS_OUTPUT
// {
// float4 vColor : SV_Target0;
// };
//
// PS_OUTPUT MainPs(PS_INPUT i)
// {
// PS_OUTPUT ps_output;
// float4 c;
// c = g_tColor.Sample(g_sAniso, i.vTextureCoords2.xy);
// c += g_tColor.Sample(g_sAniso, i.vTextureCoords3.xy);
// ps_output.vColor = c;
// return ps_output;
// }
const std::string text = R"(
OpCapability Shader
%1 = OpExtInstImport "OpenCL.DebugInfo.100"
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %MainPs "MainPs" %g_tColor %g_sAniso %in_var_TEXCOORD2 %in_var_TEXCOORD3 %out_var_SV_Target0
OpExecutionMode %MainPs OriginUpperLeft
%22 = OpString "foo.frag"
%26 = OpString "PS_OUTPUT"
%30 = OpString "float"
%33 = OpString "vColor"
%35 = OpString "PS_INPUT"
%40 = OpString "vTextureCoords3"
%42 = OpString "vTextureCoords2"
%44 = OpString "@type.2d.image"
%45 = OpString "type.2d.image"
%47 = OpString "Texture2D.TemplateParam"
%51 = OpString "src.MainPs"
%55 = OpString "c"
%57 = OpString "ps_output"
%60 = OpString "i"
%62 = OpString "@type.sampler"
%63 = OpString "type.sampler"
%65 = OpString "g_sAniso"
%67 = OpString "g_tColor"
OpName %type_2d_image "type.2d.image"
OpName %g_tColor "g_tColor"
OpName %type_sampler "type.sampler"
OpName %g_sAniso "g_sAniso"
OpName %in_var_TEXCOORD2 "in.var.TEXCOORD2"
OpName %in_var_TEXCOORD3 "in.var.TEXCOORD3"
OpName %out_var_SV_Target0 "out.var.SV_Target0"
OpName %MainPs "MainPs"
OpName %PS_INPUT "PS_INPUT"
OpMemberName %PS_INPUT 0 "vTextureCoords2"
OpMemberName %PS_INPUT 1 "vTextureCoords3"
OpName %param_var_i "param.var.i"
OpName %PS_OUTPUT "PS_OUTPUT"
OpMemberName %PS_OUTPUT 0 "vColor"
OpName %type_sampled_image "type.sampled.image"
OpDecorate %in_var_TEXCOORD2 Location 0
OpDecorate %in_var_TEXCOORD3 Location 1
OpDecorate %out_var_SV_Target0 Location 0
OpDecorate %g_tColor DescriptorSet 0
OpDecorate %g_tColor Binding 0
OpDecorate %g_sAniso DescriptorSet 0
OpDecorate %g_sAniso Binding 1
%int = OpTypeInt 32 1
%int_0 = OpConstant %int 0
%int_1 = OpConstant %int 1
%uint = OpTypeInt 32 0
%uint_32 = OpConstant %uint 32
%float = OpTypeFloat 32
%type_2d_image = OpTypeImage %float 2D 2 0 0 1 Unknown
%_ptr_UniformConstant_type_2d_image = OpTypePointer UniformConstant %type_2d_image
%type_sampler = OpTypeSampler
%_ptr_UniformConstant_type_sampler = OpTypePointer UniformConstant %type_sampler
%v2float = OpTypeVector %float 2
%_ptr_Input_v2float = OpTypePointer Input %v2float
%v4float = OpTypeVector %float 4
%_ptr_Output_v4float = OpTypePointer Output %v4float
%void = OpTypeVoid
%uint_128 = OpConstant %uint 128
%uint_0 = OpConstant %uint 0
%uint_64 = OpConstant %uint 64
%69 = OpTypeFunction %void
%PS_INPUT = OpTypeStruct %v2float %v2float
%_ptr_Function_PS_INPUT = OpTypePointer Function %PS_INPUT
%PS_OUTPUT = OpTypeStruct %v4float
%_ptr_Function_PS_OUTPUT = OpTypePointer Function %PS_OUTPUT
%_ptr_Function_v4float = OpTypePointer Function %v4float
%_ptr_Function_v2float = OpTypePointer Function %v2float
%type_sampled_image = OpTypeSampledImage %type_2d_image
%g_tColor = OpVariable %_ptr_UniformConstant_type_2d_image UniformConstant
%g_sAniso = OpVariable %_ptr_UniformConstant_type_sampler UniformConstant
%in_var_TEXCOORD2 = OpVariable %_ptr_Input_v2float Input
%in_var_TEXCOORD3 = OpVariable %_ptr_Input_v2float Input
%out_var_SV_Target0 = OpVariable %_ptr_Output_v4float Output
%43 = OpExtInst %void %1 DebugInfoNone
%59 = OpExtInst %void %1 DebugExpression
%24 = OpExtInst %void %1 DebugSource %22
%25 = OpExtInst %void %1 DebugCompilationUnit 1 4 %24 HLSL
%28 = OpExtInst %void %1 DebugTypeComposite %26 Structure %24 11 1 %25 %26 %uint_128 FlagIsProtected|FlagIsPrivate %29
%31 = OpExtInst %void %1 DebugTypeBasic %30 %uint_32 Float
%32 = OpExtInst %void %1 DebugTypeVector %31 4
%29 = OpExtInst %void %1 DebugTypeMember %33 %32 %24 13 5 %28 %uint_0 %uint_128 FlagIsProtected|FlagIsPrivate
%36 = OpExtInst %void %1 DebugTypeComposite %35 Structure %24 5 1 %25 %35 %uint_128 FlagIsProtected|FlagIsPrivate %37 %38
%39 = OpExtInst %void %1 DebugTypeVector %31 2
%38 = OpExtInst %void %1 DebugTypeMember %40 %39 %24 8 5 %36 %uint_64 %uint_64 FlagIsProtected|FlagIsPrivate
%37 = OpExtInst %void %1 DebugTypeMember %42 %39 %24 7 5 %36 %uint_0 %uint_64 FlagIsProtected|FlagIsPrivate
%46 = OpExtInst %void %1 DebugTypeComposite %44 Class %24 0 0 %25 %45 %43 FlagIsProtected|FlagIsPrivate
%50 = OpExtInst %void %1 DebugTypeFunction FlagIsProtected|FlagIsPrivate %28 %36
%52 = OpExtInst %void %1 DebugFunction %51 %50 %24 16 1 %25 %51 FlagIsProtected|FlagIsPrivate 17 %43
%54 = OpExtInst %void %1 DebugLexicalBlock %24 17 1 %52
%56 = OpExtInst %void %1 DebugLocalVariable %55 %32 %24 20 12 %54 FlagIsLocal
%58 = OpExtInst %void %1 DebugLocalVariable %57 %28 %24 18 15 %54 FlagIsLocal
%61 = OpExtInst %void %1 DebugLocalVariable %60 %36 %24 16 29 %52 FlagIsLocal 1
%64 = OpExtInst %void %1 DebugTypeComposite %62 Structure %24 0 0 %25 %63 %43 FlagIsProtected|FlagIsPrivate
%66 = OpExtInst %void %1 DebugGlobalVariable %65 %64 %24 3 14 %25 %65 %g_sAniso FlagIsDefinition
%68 = OpExtInst %void %1 DebugGlobalVariable %67 %46 %24 1 11 %25 %67 %g_tColor FlagIsDefinition
%MainPs = OpFunction %void None %69
%70 = OpLabel
%135 = OpExtInst %void %1 DebugScope %54
%111 = OpVariable %_ptr_Function_PS_OUTPUT Function
%112 = OpVariable %_ptr_Function_v4float Function
%136 = OpExtInst %void %1 DebugNoScope
%param_var_i = OpVariable %_ptr_Function_PS_INPUT Function
%74 = OpLoad %v2float %in_var_TEXCOORD2
%75 = OpLoad %v2float %in_var_TEXCOORD3
%76 = OpCompositeConstruct %PS_INPUT %74 %75
OpStore %param_var_i %76
%137 = OpExtInst %void %1 DebugScope %52
%115 = OpExtInst %void %1 DebugDeclare %61 %param_var_i %59
%138 = OpExtInst %void %1 DebugScope %54
%116 = OpExtInst %void %1 DebugDeclare %58 %111 %59
%117 = OpExtInst %void %1 DebugDeclare %56 %112 %59
;CHECK-NOT: %117 = OpExtInst %void %1 DebugDeclare %56 %112 %59
OpLine %22 21 9
%118 = OpLoad %type_2d_image %g_tColor
OpLine %22 21 29
%119 = OpLoad %type_sampler %g_sAniso
OpLine %22 21 40
%120 = OpAccessChain %_ptr_Function_v2float %param_var_i %int_0
%121 = OpLoad %v2float %120
OpLine %22 21 9
%122 = OpSampledImage %type_sampled_image %118 %119
%123 = OpImageSampleImplicitLod %v4float %122 %121 None
OpLine %22 21 5
OpStore %112 %123
;CHECK: %140 = OpExtInst %void %1 DebugValue %56 %123 %59
OpLine %22 22 10
%124 = OpLoad %type_2d_image %g_tColor
OpLine %22 22 30
%125 = OpLoad %type_sampler %g_sAniso
OpLine %22 22 41
%126 = OpAccessChain %_ptr_Function_v2float %param_var_i %int_1
%127 = OpLoad %v2float %126
OpLine %22 22 10
%128 = OpSampledImage %type_sampled_image %124 %125
%129 = OpImageSampleImplicitLod %v4float %128 %127 None
OpLine %22 22 7
%131 = OpFAdd %v4float %123 %129
OpLine %22 22 5
OpStore %112 %131
;CHECK: %141 = OpExtInst %void %1 DebugValue %56 %131 %59
OpLine %22 23 5
%133 = OpAccessChain %_ptr_Function_v4float %111 %int_0
OpStore %133 %131
OpLine %22 24 12
%134 = OpLoad %PS_OUTPUT %111
%139 = OpExtInst %void %1 DebugNoScope
%79 = OpCompositeExtract %v4float %134 0
OpStore %out_var_SV_Target0 %79
OpReturn
OpFunctionEnd
)";
SetTargetEnv(SPV_ENV_VULKAN_1_2);
SetAssembleOptions(SPV_TEXT_TO_BINARY_OPTION_PRESERVE_NUMERIC_IDS);
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
Fix SSA re-writing in the presence of variable pointers. (#4010) This fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/3873. In the presence of variable pointers, the reaching definition may be another pointer. For example, the following fragment: %2 = OpVariable %_ptr_Input_float Input %11 = OpVariable %_ptr_Function__ptr_Input_float Function OpStore %11 %2 %12 = OpLoad %_ptr_Input_float %11 %13 = OpLoad %float %12 corresponds to the pseudo-code: layout(location = 0) in flat float *%2 float %13; float *%12; float **%11; *%11 = %2; %12 = *%11; %13 = *%12; which ultimately, should correspond to: %13 = *%2; During rewriting, the pointer %12 is found to be replaceable by %2. However, when processing the load %13 = *%12, the type of %12's reaching definition is another float pointer (%2), instead of a float value. When this happens, we need to continue looking up the reaching definition chain until we get to a float value or a non-target var (i.e. a variable that cannot be SSA replaced, like %2 in this case since it is a function argument).
2020-11-04 15:23:53 +00:00
// Check support for pointer variables. When pointer variables are used, the
// computation of reaching definitions may need to follow pointer chains.
// See https://github.com/KhronosGroup/SPIRV-Tools/issues/3873 for details.
TEST_F(LocalSSAElimTest, PointerVariables) {
const std::string text = R"(
OpCapability Shader
OpCapability VariablePointers
OpExtension "SPV_KHR_variable_pointers"
OpMemoryModel Logical Simple
OpEntryPoint Fragment %1 "main" %2 %3
OpExecutionMode %1 OriginUpperLeft
%float = OpTypeFloat 32
%void = OpTypeVoid
%6 = OpTypeFunction %void
%_ptr_Input_float = OpTypePointer Input %float
%_ptr_Output_float = OpTypePointer Output %float
%_ptr_Function__ptr_Input_float = OpTypePointer Function %_ptr_Input_float
%2 = OpVariable %_ptr_Input_float Input
%3 = OpVariable %_ptr_Output_float Output
%1 = OpFunction %void None %6
%10 = OpLabel
%11 = OpVariable %_ptr_Function__ptr_Input_float Function
OpStore %11 %2
; CHECK-NOT: %12 = OpLoad %_ptr_Input_float %11
%12 = OpLoad %_ptr_Input_float %11
; CHECK: %13 = OpLoad %float %2
%13 = OpLoad %float %12
OpStore %3 %13
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<SSARewritePass>(text, true);
}
Update passes to handle function declarations (#4599) Spirv-opt has not had to handle module with function declarations. This lead many passes to assume that every function has a body. This is not always true. This commit will modify a number of passes to handle function declarations. Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/4443
2021-10-28 15:54:37 +00:00
TEST_F(LocalSSAElimTest, FunctionDeclaration) {
// Make sure the pass works with a function declaration that is called.
const std::string text = R"(OpCapability Addresses
OpCapability Linkage
OpCapability Kernel
OpCapability Int8
%1 = OpExtInstImport "OpenCL.std"
OpMemoryModel Physical64 OpenCL
OpEntryPoint Kernel %2 "_Z23julia__1166_kernel_77094Bool"
OpExecutionMode %2 ContractionOff
OpSource Unknown 0
OpDecorate %3 LinkageAttributes "julia_error_7712" Import
%void = OpTypeVoid
%5 = OpTypeFunction %void
%3 = OpFunction %void None %5
OpFunctionEnd
%2 = OpFunction %void None %5
%6 = OpLabel
%7 = OpFunctionCall %void %3
OpReturn
OpFunctionEnd
)";
SinglePassRunAndCheck<SSARewritePass>(text, text, false);
}
Add LocalMultiStoreElim pass A SSA local variable load/store elimination pass. For every entry point function, eliminate all loads and stores of function scope variables only referenced with non-access-chain loads and stores. Eliminate the variables as well. The presence of access chain references and function calls can inhibit the above optimization. Only shader modules with logical addressing are currently processed. Currently modules with any extensions enabled are not processed. This is left for future work. This pass is most effective if preceeded by Inlining and LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim will reduce the work that this pass has to do.
2017-06-16 21:37:31 +00:00
// TODO(greg-lunarg): Add tests to verify handling of these cases:
//
// No optimization in the presence of
// access chains
// function calls
// OpCopyMemory?
// unsupported extensions
// Others?
Move tests into namespaces (#1689) This CL moves the test into namespaces based on their directories.
2018-07-11 13:24:49 +00:00
} // namespace
} // namespace opt
} // namespace spvtools
Reference in New Issue Copy Permalink