We currently place the load instructions at the start of the basic block
that dominates all of the loads. If that basic block contains OpPhi
instructions, then this will generate invalid code. We just need to
search for a location that comes after all of the OpPhi instructions.
Fixes#2204.
* Don't fold specialized branchs in loop unswitch
Folding branches can have a lot of special cases, and can be a little
error prone. So I only want it in one place. That will be in dead
branch elimination. I will change loop unswitching to set the branches
that were being folded to have a constant condition. Then subsequent
pass of dead branch elimination will be able to remove the code.
At the same time, I added a check that loop unswitching will not
unswitch a branch with a constant condition. It is not useful to do it
because dead branch elimination will simple fold the branch anyway.
Also it avoid an infinite loop that would other wise be introduced by my
first change.
Fixes#2203.
Loop unswitching is unswitching the conditional branch that creates the
back-edge. In the version of the loop, where the bachedge is not taken,
there is no back-edge. This is what causes the validator to complain.
The solution I will go with will be to now unswitch a condition with a
back-edge. At this time we do not now if loop unswitching is used. We do
not include it in the optimization sets provided, nor is it used in
glslang's set. When there are opportunities and no breaks from the loop,
the loop with either be a single iteration loop, or an infinite loop.
There is no performance advantage to performing loop unswitching in
either of those cases. If there is a break, maintaining structured
control flow will be tricky. Unless we see a clear advantage to handling
these case, I would go with the safer simpler solution.
Fixes#2201.
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.
* Invalidate the decoration manager at the start of ADCE.
If the decoration manager is kept live the the contex will try to keep
it up to date. ADCE deals with group decorations by changing the
operands in |OpGroupDecorate| instructions directly without informing
the decoration manager. This puts it in an invalid state, which will
cause an error when the context tries to update it. To Avoid this
problem, we will invalidate the decoration manager upfront.
At the same time, the decoration manager is now considered when checking
the consistency of the decoration manager.
* Fix invalid OpPhi generated by merge-return.
When we create a new phi node for a value say %10, we have to replace
all of the uses of %10 that are no longer dominated by the def of %10
by the result id of the new phi. However, if the use is in a phi node,
it is possible that the bb contains the use is not dominated by either.
In this case, needs to be handled differently.
* Split loop headers before add a new branch to them.
In merge return, Phi node in loop header that are also merges for loop
do not get updated correctly. Those cases do not fit in with our
current analysis. Doing this will simplify the code by reducing the
number of cases that have to be handled.
Added documentation to the ir context to indicates that TakeNextId()
returns 0 when the max id is reached. TODOs were added to each call
sight so that we know where we have to start to handle this case.
Handle id overflow in |SplitLoopHeader|.
Handle id overflow in |GetOrCreatePreHeaderBlock|.
Handle failure to create preheader in LICM.
Part of https://github.com/KhronosGroup/SPIRV-Tools/issues/1841.
We currently simulate all shift operations when the two operand are
constants. The problem is that if the shift amount is larger than
32, the result is undefined.
I'm changing the folder to return 0 if the shift value is too high.
That way, we will have defined behaviour.
https://crbug.com/910937.
Upgrade to VulkanKHR memory model
* Converts Logical GLSL450 memory model to Logical VulkanKHR
* Adds extension and capability
* Removes deprecated decorations and replaces them with appropriate
flags on downstream instructions
* Support for Workgroup upgrades
* Support for copy memory
* Adding support for image functions
* Adding barrier upgrades and tests
* Use QueueFamilyKHR scope instead of device
* Move ProcessFunction* function from pass to the context.
There are a few functions that are used to traverse the call tree.
They currently live in the Pass class, but they have nothing to do with
a pass, and may be needed outside of a pass. They would be better in
the ir context, or in a specific call tree class if we ever have a need
for it.
* Don't inline recursive functions.
Inlining does not check if a function is recursive or not. This has
been fine as long as the shader was a Vulkan shader, which forbid
recursive functions. However, not all shaders are vulkan, so either
we limit inlining to Vulkan shaders or we teach it to look for recursive
functions.
I prefer to keep the passes as general as is reasonable. The change
does not require much new code in inlining and gives a reason to refactor
some other code.
The changes are to add a member function to the Function class that
checks if that function is recursive or not.
Then this is used in inlining to not inlining a function call if it calls
a recursive function.
* Add id to function analysis
There are a few places that build a map from ids to Function whose
result is that id. I decided to add an analysis to the context for this
to reduce that code, and simplify some of the functions.
* Add missing file.
* Added a reduction pass to replace ids with ids of the same type that dominate them.
* Introduce helper method for querying whether an operand type is an input id.
These are bookend passes designed to help preserve line information
across passes which delete, move and clone instructions. The propagation
pass attaches a debug line instruction to every instruction based on
SPIR-V line propagation rules. It should be performed before optimization.
The redundant line elimination pass eliminates all line instructions
which match the previous line instruction. This pass should be performed
at the end of optimization to reduce physical SPIR-V file size.
Fixes#2027.
The type manager in spirv-opt currently asserts if a function parameter
has type void. It is not exactly clear from the spec that this is
disallowed, even if it probably will be disallowed. In either case,
asserts should be used to verify assumptions that will actually make a
difference to the code. As far as the optimizer is concerned, a void
parameter does not matter. I don't see the point of the assert. I'll
just remove it and let the validator decide whether to accept it or not.
No test was added because it is not clear that it is legal, and should
not force us to accept it in the future unless the spec make it clear
that it is legal.
Fixes crbug.com/903088.
That function currently only handled OpPtrAccessChain if it was in the
middle of the chain, but not at the start. Fixing that up.
Fixes crbug.com/905271.
* Add base and core bindless validation instrumentation classes
* Fix formatting.
* Few more formatting fixes
* Fix build failure
* More build fixes
* Need to call non-const functions in order.
Specifically, these are functions which call TakeNextId(). These need to
be called in a specific order to guarantee that tests which do exact
compares will work across all platforms. c++ pretty much does not
guarantee order of evaluation of operands, so any such functions need to
be called separately in individual statements to guarantee order.
* More ordering.
* And more ordering.
* And more formatting.
* Attempt to fix NDK build
* Another attempt to address NDK build problem.
* One more attempt at NDK build failure
* Add instrument.hpp to BUILD.gn
* Some name improvement in instrument.hpp
* Change all types in instrument.hpp to int.
* Improve documentation in instrument.hpp
* Format fixes
* Comment clean up in instrument.hpp
* imageInst -> image_inst
* Fix GetLabel() issue.
If there is only 1 return and it is in a loop, then the function cannot be inlined.
Fix condition when inlined code needs one-trip loop wrapper. The dummy loop is needed when there is a return inside a selection construct. Even if there is only 1 return.
When looking for a break from a selection construct, we do not realize
that a jump to the continue target of a loop containing the selection
is a break. This causes and infinit loop, or possibly other failures.
Fixes#2004.
When looking for a break from a selection construct, we do not need to
look inside nested constructs. However, if a loop header has an
unconditional branch, then we enter the loop. Entering the loop causes
an infinite loop because we keep going through the loop.
The solution is to look for a merge block, if one exsits, even for block
terminated by an OpBranch.
Fixes#1979.
ADCE liveness algorithm should treat OpUnreachable at least like other
branch instructions. It was being treated as always live which was
preventing useless structured constructs from being eliminated.
OpUnreachable is generated by dead branch elimination which is now
being required by merge return, so this fix should accompany that
change.
We currently run merge-return on all functions, but
dead-branch-elimination only runs on function reachable from an entry
point or exported function. Since dead-branch-elimination is needed for
merge-return, they have to match.
Fixes#1976.
Was removing control structures which didn't have data dependency
with enclosed live loop and otherwise did not contain live code.
An example is a counting loop around a live loop.
Fixes#1967.
Consider atomics that load when analyzing live stores in ADCE.
Previously it asserted that the base of an OpImageTexelPointer should
be an image. It is actually a pointer to an image, so IsValidBasePointer
should suffice.
Merge return assumes that the only unreachable blocks are those needed
to keep the structured cfg valid. Even those must be essentially empty
blocks.
If this is not the case, we get unpredictable behaviour. This commit
add a check in merge return, and emits an error if it is not the case.
Added a pass of dead branch elimination before merge return in both the
performance and size passes. It is a precondition of merge return.
Fixes#1962.
The current implementation in the folder when seeing a division by zero
is to assert. In the release build, the compiler will attempt to
compute the value, which causes its own problems.
The solution I will go with is to fold the division, and just give it
the value of 0. The same goes for remainder and mod operations.
Fixes#1961.
The HlslCounterBufferGOOGLE that was introduced changed the OpDecorateId
so that is can now reference an id other than the target. If that other
id is used only in the decoration, then the definition of the id will be
removed because decoration do not count as real uses.
However, if the target of the decoration is still live the decoration
will not be removed. This leaves a reference to an id that is not
defined.
There are two solutions to consider. The first is that is the decoration
is kept, then the definition of the id should be kept live. Implementing
this change would be involved because the way ADCE handles decorations
will have to be reimplemented.
The other solution is to remove the decoration the id is otherwise dead.
This works for this specific case. Also this is the more desirable
behaviour in this case. The id will always be the id of a variable that
belongs to a descriptor set. If that variable is not bound and we do
not remove it, the driver will complain.
I chose to implement the second solution. The first will be left to when
a case for it comes up.
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1885.
There are a few spots where copy propagate arrays is trying
to go from a Type to an id, but the type is not unique. When generating
code this pass needs specific ids, otherwise we get type mismatches.
However, the ambigous types means we can sometimes get the wrong type
and generate invalid code.
That code has been rewritten to not rely on the type manager, and just
look at the instructions instead.
I have opened https://github.com/KhronosGroup/SPIRV-Tools/issues/1939 to
try to get a way to make this more robust.
* Analyze uses for all instructions.
The def-use manager needs to fill in the `inst_to_used_ids_` field for
every instruction. This means we have to analyze the uses for every
instruction, even if they do not have any uses.
This mistake was not found earlier because there was a typo in the
equality check for def-use managers. No new tests are needed.
While looking into this I found redundant work in block merge. Cleaning
that up at the same time.
* Fix other transformations
Aggressive dead code elimination did not update the OpGroupDecorate
and the OpGroupMemberDecorate instructions properly when they are
updated. That is fixed.
Dead branch elimination did not analyze the OpUnreachable instructions
that is would add. That is taken care of.
In DecorationManager::RemoveDecorationsFrom, we do not remove the id
from a decoration group if the group has no decorations. This causes
problems because KillNamesAndDecorates is suppose to remove all
references to the id, but in this case, there is still a reference.
This is fixed by adding a special case.
Also, there is the possibility of a double free because
RemoveDecorationsFrom will delete the instructions defining |id| when
|id| is a decoration group. Later, KillInst would later write to memory
that has been deleted when trying to turn it into a Nop. To fix this,
we will only remove the decorations that use |id| and not its definition
in RemoveDecorationsFrom.
Add code to keep the def-use manger and the inst-to-block mapping up-to-date. This means we do not have to rebuild them later.
To make this work, we will have to have to find places to update the
def-use manager. Updating the def-use manager is not straight forward
because we are unrolling loops, and we have circular references.
This forces one pass to register all of the definitions. A second one
to analyze the uses. Also because there will be references to the new
instructions in the old code, we want to register the definitions of the
new instructions early, so we can update the uses of the older code as
we go along.
The inst-to-block mapping is not too difficult. It can be done as instructions are created.
Fixes#1928.
A limit of 0 for the scalar replacement options it used to indicate that
there is no limit. The current implementation does not allow 0. This
should be fixed.
It seems like the current implementation of KillNameAndDecorates does
not handle group decorations correctly. The id being removed is not
removed from the OpGroupDecorate instructions. Even worst, any
decorations that apply to that group are removed.
The solution is to use the function in the decoration manager that will
remove the decorations and update the instructions instead of doing the
work itself.
Adds unrolling to the legalization passes.
After enabling unrolling I found a bug when there is a self-referencing
phi node. That has been fixed.
The test that checks for that the order of optimizations is correct also
needed to be updated.
The current implementation of merge return can create bad, but correct,
code. When it is not in a loop construct, it will insert a lot of
extra branch around code. The potentially large number of branches are
bad. At the same time, it can separate code store to variables from
its uses hiding the fact that the store dominates the load.
This hurts the later analysis because the compiler thinks that multiple
values can reach a load, when there is really only 1. This poorer
analysis leads to missed optimizations.
The solution is to create a dummy loop around the entire body of the
function, then we can break from that loop with a single branch. Also
only new merge nodes would be those at the end of loops meaning that
most analysies will not be hurt.
Remove dead code for cases that are no longer possible.
It seems like some drivers expect there the be an OpSelectionMerge
before conditional branches, even if they are not strictly needed.
So we add them.
* Create structed cfg analysis.
There are lots of optimization that have to traverse the CFG in a
structured order just because it wants to know which constructs a
basic block in contained in. This adds extra complexity to these
optimizations, for causes too much refactoring of older optimizations.
To help with this problem, I have written an analysis that can give this
information.
* Identify branches breaking from loops.
Dead branch elimination does a search for a conditional branch to the
end of the current selection construct. This search assumes that the
only way to leave the construct is through the merge node. But that is
not true. The code can jump to the merge node of a loop that contains
the construct.
The search needs to take this into consideration.
When using lldb and/or gdb I frequently get odd std::string failures
when using the IR printing instructions we have now. This adds the
methods Instruction::Dump(), BasicBlock::Dump() and Function::Dump() to
emit the output of the pretty print to stderr.
With this I can now reliably print IR from gdb and lldb sessions.
In merge blocks, we do not allow the merging of two blocks with merge
instructions. This is because if the two block are merged only 1 of
those instructions can exists. However, if the successor block is the
merge block of the predecessor, then we can delete the merge instruction
in the predecessor. In this case, we are able to merge the blocks.
* Create a new entry point for the optimizer
Creates a new struct to hold the options for the optimizer, and creates
an entry point that take the optimizer options as a parameter.
The old entry point that takes validator options are now deprecated.
The validator options will be one of the optimizer options.
Part of the optimizer options will also be the upper bound on the id bound.
* Add a command line option to set the max value for the id bound. The default is 0x3FFFFF.
* Modify `TakeNextIdBound` to return 0 when the limit is reached.
* Have the constant manager take ownership of constants.
Right now the owner of an object of type contant that is in the
|const_pool_| of the constant manager is unclear. The constant
manager does not delete them, there is no other reasonable owner. This
causes memory leaks.
This change fixes the memory leaks by having the constant manager
take ownership of the constant that is stores in |const_pool_|. Other
changes include interface changes to make it explicit that the constant
manager takes ownership of the object when a constant is registered
with the constant manager.
Fixes#1865.
Right now the owner of an object of type contant that is in the
|const_pool_| of the constant manager is unclear. The constant
manager does not delete them, there is no other reasonable owner. This
causes memory leaks.
This change fixes the memory leaks by having the constant manager
take ownership of the constant that is stores in |const_pool_|. Other
changes include interface changes to make it explicit that the constant
manager takes ownership of the object when a constant is registered
with the constant manager.
* 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.
If seems like at least 1 driver does not like a condition jump to the end
of a selection construct. We are generating these in the merge return
pass. This change stops merge return from generating this sequence.
Part of #1861.
When doing predicate blocks, we need to traverse every block in
structured order in order to keep track of which construct a block is
contained in. The standard way of traversing code in structured order
is to create a list with all of the nodes in order. However, when
predicating blocks, new blocks are created, and those blocks are missed.
This causes branches that go too far.
The solution is to update the order as new blocks are created. Since
we are using an std::list, we do not have to worry about invalidation of
iterators when changing the list.
* Refactor PredicateBlocks
Refactor PredicateBlocks so that we know which constructs a return
is contained in. Will be used later.
* Have PredicateBlocks jump the existing merge blocks.
In PredicateBlocks, we currently skip instructions with side effects,
but it still follows the same control flow (sort-of). This causes a
problem, when we are trying to predicate code in a loop. We skip all
of the code with side effects (IV increment), but still follow the
same control flow (jump back the start of the loop). This creates an
infinite loop because the code will keep jumping back to the start of
the loop without changing the values that effect the exit condition.
This is a large change to merge-return. When predicating a block that
is in a loop or merge construct, it will jump to the merge block of the
construct. Once out of all constructs we will generate code as we did
before.
* Handle breaks from structured-ifs in DCE.
dead code elimination assumes that are conditional branches except for
breaks and continues in loops will have an OpSelectionMerge before them.
That is not true when breaking out of a selection construct.
The fix is to look for breaks in selection constructs in the same place
we look for breaks and continues for loops.
When dead-branch-elim folds a conditional branch, it also deletes the
OpSelectionMerge instruction. If that construct contains a
conditional branch to the merge node, it will not have its own
OpSelectionMerge. When the headers merge instruction is deleted, the
the inner conditional branch will no longer be legal. It will be a
selection to a node that is not a merge node.
We fix this up by moving the OpSelectionMerge to a new location if it is
still needed.
This forks the testing harness from https://github.com/google/shaderc
to allow testing CLI tools.
New features needed for SPIRV-Tools include:
1- A new PlaceHolder subclass for spirv shaders. This place holder
calls spirv-as to convert assembly input into SPIRV bytecode. This is
required for most tools in SPIRV-Tools.
2- A minimal testing file for testing basic functionality of spirv-opt.
Add tests for all flags in spirv-opt.
1. Adds tests to check that known flags match the names that each pass
advertises.
2. Adds tests to check that -O, -Os and --legalize-hlsl schedule the
expected passes.
3. Adds more functionality to Expect classes to support regular
expression matching on stderr.
4. Add checks for integer arguments to optimization flags.
5. Fixes#1817 by modifying the parsing of integer arguments in
flags that take them.
6. Fixes -Oconfig file parsing (#1778). It reads every line of the file
into a string and then parses that string by tokenizing every group of
characters between whitespaces (using the standard cin reading
operator). This mimics shell command-line parsing, but it does not
support quoting (and I'm not planning to).
In local-access-chain-convert, we replace loads by load the entire
variable, then doing the extract. The extract will have the same value
as the load. However, if the load has a decoration on it, the
decoration is lost because we do not copy any them to the new id.
This is fixed by rewritting the load into the extract and keeping the
same result id.
This change has the effect that we do not call DCEInst on the loads
because the load is not being deleted, but replaced. This could leave
OpAccessChain instructions around that are not used. This is not a
problem for -O and -Os. They run local_single_*_elim passes and then
dead code elimination. The dce will remove the unused access chains,
and the load elimination passes work even if there are unused access
chains. I have added test to them to ensure they will not loss
opportunities.
Fixes#1787.
In `TypeManager::RebuildType`, the base cases call `Clone`, which will
copy the decorations for the type. After that it breaks out of the
switch statement and copies the decorations again.
This has not causes any real problems yet because none of those types
are allowed to have decorations. However to make the code more robust
it is best to not copy twice because it should be empty.
This way if a new base type or decoration is added that changes this
rule the code will be correct.
* Run the validator in the optimization fuzzers.
The optimizers assumes that the input to the optimizer is valid. Since
the fuzzers do not check that the input is valid before passing the
spir-v to the optimizer, we are getting a few errors.
The solution is to run the validator in the optimizer to validate the
input.
For the legalization passes, we need to add an extra option to the
validator to accept certain types of variable pointers, even if the
capability is not given. At the same time, we changed the option
"--legalize-hlsl" to relax the validator in the same way instead of
turning it off.
In the merge return pass, we will split a block, but not update the phi
instructions that reference the block. Since the branch in the original
block is now part of the block with the new id, the phi nodes must be
updated.
This commit will change this.
I have also considered other places where an id of a basic block could
be referenced, and I don't think any of them need to change.
1) Branch and merge instructions: These jump to the start of the
original block, and so we want them to jump to the block that uses the
original id. Nothing needs to change.
2) Names and decorations: I don't think it matters with block keeps the
name, and there are no decorations that apply to basic blocks.
Fixes#1736.
Many of the files have using std::<foo> statements in them, but then the
use of <foo> will be inconsistently std::<foo> or <foo> scattered
through the file. This CL removes all of the using statements and
updates the code to have the required std:: prefix.
When creating a new phi for a value in the function, merge return will
rewrite all uses of an id that are no longer dominated by its
definition. Uses that are not in a basic block, like OpName or
decorations, are not dominated, but they should not be replaced.
Fixes#1736.
* Combines OpAccessChain, OpInBoundsAccessChain, OpPtrAccessChain and
OpInBoundsPtrAccessChain
* New folding rule to fold add with 0 for integers
* Converts to a bitcast if the result type does not match the operand
type
V
This re-implements the -Oconfig=<file> flag to use a new API that takes
a list of command-line flags representing optimization passes.
This moves the processing of flags that create new optimization passes
out of spirv-opt and into the library API. Useful for other tools that
want to incorporate a facility similar to -Oconfig.
The main changes are:
1- Add a new public function Optimizer::RegisterPassesFromFlags. This
takes a vector of strings. Each string is assumed to have the form
'--pass_name[=pass_args]'. It creates and registers into the pass
manager all the passes specified in the vector. Each pass is
validated internally. Failure to create a pass instance causes the
function to return false and a diagnostic is emitted to the
registered message consumer.
2- Re-implements -Oconfig in spirv-opt to use the new API.
Fixes#1731
* Updated folding rules related to vector shuffle to account for the
undef literal value:
* FoldVectorShuffleFeedingShuffle
* FoldVectorShuffleFeedingExtract
* FoldVectorShuffleWithConstants
* These rules would commit memory violations due to treating the undef
literal value as an accessible composite component
Currentlty opt::Instruction class holds a cache of the result_id and
type_id for the instruction. That cache needs to be updated if the
underlying operand values are changes.
This CL changes the cache to being a flag if there is a type or result
id for the instruction. We then retrieve the value if needed from the
operands.
Fixes#1727
* If the pass finds any dead branches it can optimize then at the end of
the pass it reorders basic blocks to ensure they satisfy block ordering
requirements
* Added some new tests
* While investigating this issue, found and fixed a non-deterministic
ordering of dominators
* Now the edges used to construct the dominator tree are sorted
according to posorder traversal indices
With current implementation, the constant manager does not keep around
two constant with the same value but different types when the types
hash to the same value. So when you start looking for that constant you
will get a constant with the wrong type back.
I've made a few changes to the constant manager to fix this. First off,
I have changed the map from constant to ids to be an std::multimap.
This way a single constant can be mapped to mutiple ids each
representing a different type.
Then when asking for an id of a constant, we can search all of the ids
associated with that constant in order to find the one with the correct
type.
When folding an OpVectorShuffle where the first operand is defined by
an OpVectorShuffle, is unused, and is equal to the second, we end up
with an infinite loop. This is because we think we change the
instruction, but it does not actually change. So we keep trying to
folding the same instruction.
This commit fixes up that specific issue. When the operand is unused,
we replace it with Null.
When folding a vector shuffle that feeds another vector shuffle causes
the size of the first operand to change, when other indices have to be
adjusted reletive to the new size.
The function class provides a {Set|Get}Parent call in order to provide
the context to the LoopDescriptor methods. This CL removes the module
from Function and provides the needed context directly to LoopDescriptor
on creation.
This CL removes the context() method from opt::Function. In the places
where the context() was used we can retrieve, or provide, the context in
another fashion.
Currently the IRContext is passed into the Pass::Process method. It is
then up to the individual pass to store the context into the context_
variable. This CL changes the Run method to store the context before
calling Process which no-longer receives the context as a parameter.
This CL moves the files in opt/ to consistenly be under the opt::
namespace. This frees up the ir:: namespace so it can be used to make a
shared ir represenation.
Currently the utils/ folder uses both spvutils:: and spvtools::utils.
This CL changes the namespace to consistenly be spvtools::utils to match
the rest of the codebase.
The folding routines are currently global functions. They also rely on
data in an std::map that holds the folding rules for each opcode. This
causes that map to not have a clear owner, and therefore never gets
deleted.
There has been a request to delete this map. To implement this, we will
create a InstructionFolder class that owns the maps. The IRContext will
own the InstructionFolder instance. Then the global functions will
become public memeber functions of the InstructionFolder.
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1659.
There are a few locations where we need to handle duplicate types. We
cannot merge them because they may be needed for reflection. When this
happens we need do some extra lookups in the type manager.
The specific fixes are:
1) When generating a constant through `GetDefiningInstruction` accept
and use an id for the desired type of the constant. This will make sure
you get the type that is needed.
2) In Private-to-local, make sure we to update the def-use chains when a
new pointer type is created.
3) In the type manager, make sure that `FindPointerToType` returns a
pointer that points to the given type and not a duplicate type.
4) In scalar replacment, make sure the null constants that are created
are the correct type.
Many optimization will run on function scope symbols only. When symbols
are moved from private scope to function scople, then these optimizations
can do more.
I believe it is a good idea to run this pass with both -O and -Os. To
get the most out of it it should be run ASAP after inlining and something
that remove all of the dead functions.
Revert "Don't merge types of resources"
This reverts commit f393b0e480, but leaves
the tests that were added. Added new test. These test are the so that,
if someone tries the same change I made, they will see the test that
they need to handle.
Don't run remove duplicates in -O and -Os
Romve duplicates was run to help reduce compile time when looking for
types in the type manager. I've run compile time test on three sets
of shaders, and the compile time does not seem to change.
It should be safe to remove it.
During the compact IDs optimization pass, the result IDs of some
basic blocks can change. In spite of this, GetPreservedAnalyses
indicated that the CFG was preserved. But the CFG relies on
the basic blocks having the same IDs. Simply removing this flag
resolves the issue by preventing the CFG check.
Also Removes combinators and namemap preserved analyses from
compact IDs pass.
When doing reflection users care about the names of the variable, the
name of the type, and the name of the members. Remove duplicates breaks
this because it removes the names one of the types when merging.
To fix this we have to keep the different types around for each
resource. This commit adds code to remove duplicates to look for the
types uses to describe resources, and make sure they do not get merged.
However, allow merging of a type used in a resource with something not
used in a resource. Was done when the non resource type came second.
This could have a negative effect on compile time, but it was not
expected to be much.
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1372.
We replace the std::vector in the Operand class by a new class that does
a small size optimization. This helps improve compile time on Windows.
Tested on three sets of shaders. Trying various values for the small
vector. The optimal value for the operand class was 2. However, for
the Instruction class, using an std::vector was optimal. Size of "0"
means that an std::vector was used.
Instruction size
0 4 8
Operand Size
0 489 544 684
1 593 487
2 469 570
4 473
8 505
This is a single thread run of ~120 shaders. For the multithreaded run
the results were the similar. The basline time was ~62sec. The
optimal configuration was an 2 for the OperandData and an
std::vector for the OperandList with a compile time of ~38sec. Similar
expiriments were done with other sets of shaders. The compile time still
improved, but not as much.
Contributes to https://github.com/KhronosGroup/SPIRV-Tools/issues/1609.
The following passes are updated to preserve the inst-to-block and
def-use analysies:
private-to-local
aggressive dead-code elimination
dead branch elimination
local-single-block elimination
local-single-store elimination
reduce load size
compact ids (inst-to-block only)
merge block
dead-insert elimination
ccp
The one execption is that compact ids still kills the def-use manager.
This is because it changes so many ids it is faster to kill and rebuild.
Does everything in
https://github.com/KhronosGroup/SPIRV-Tools/issues/1593 except for the
changes to merge return.
By using forward pointers, we are able to define a struct that has a
pointer to itself. This could be directly or indirectly. The current
implementation of the type manager did not handle this case. There are
three changes that are made in this commit inorder to handle this case:
1) Change the handling of OpTypeForwardPointer
The current handling of OpTypeForwardsPointer is broken if there is a
reference to the pointer before the real definition. When build the
type that contain the forward delared pointer, the type manager will ask
for the type for that ID, and will get a nullptr because it does not
exists. This nullptr is not handleded very well.
The change is to keep track of the incomplete types the first time
through all of the types. An incomplete type is a ForwardPointer or any
type that references an incomplete type.
Then we implement a second pass through the incomplete types that will
complete them.
2) Hashing types.
When hashing a type, we want to uses all of the subtypes as part of the
hash. However, with types that reference them selves, this creates an
infinite recursion. To get around this, we keep track of which types
have been seen on the path from the root type. If we have see the
current type already then we can stop the recursion.
3) Comparing types.
In order to check if two types are the same, we must check that all of
their subtypes are the same as well. This also causes an infinit
recursion. The solution is to stop comparing the subtypes if we are
trying to compare two pointer types that we are already in the middle of
comparing. The ideas is that if the two pointer are different, then in
progress compare will return false itself.
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1578.
We add a new rule to the folding rules to fold an FMix feeding an
extract when the alpha value for the element being extracted is either
0 or 1. In those case, we can simple extract from one of the operands
to the FMix.
With that change the simplification pass completely subsumes the
insert-extract elimination pass. So we remove the insert-extract
elimination passes and replce them with calls to the simplification
pass.
In a follow up PR, we should delete the insert-extract elimination pass.
Contributes to https://github.com/KhronosGroup/SPIRV-Tools/issues/1570.
Currently it's impossible for external code to register a pass because
the only source file that can create pass tokens is optimizer.cpp. This
makes it hard to add passes that can't be upstreamed since you can't run
them from the usual pass sequence without reimplementing Optimizer.
This change adds a PassToken constructor that takes unique_ptr to
opt::Pass; if out-of-tree code implements opt::Pass it can register a
custom pass without having to add it to SPIRV-Tools source code.
Removes the limit on scalar replacement for the lagalization passes.
This is done by adding an option to the pass (and command line option)
to set the limit on maximum size of the composite that scalar
replacement is willing to divide.
Fixes#1494.
ADCE does not treat OpCopyMemory as an instruction that references
memory. Because of that stores are removed that should not be.
This change teaches ADCE that OpCopyMemory and OpCopyMemorySize both
loads from and stores to memory. This will keep other stores live when
needed, and will allows ADCE to remove OpCopyMemory instructions as
well.
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1556.
Currently in scalar replacement, we create a new variable for every
memeber of the composite being divided. It is often overkill, because
not all of those members will be used. This change will check which
elements are used and only create variable for the members that are
used.
This reduces the compile time for one set of shader from 248s to 165s.
Part of https://github.com/KhronosGroup/SPIRV-Tools/issues/1494.
The code patterns generated by DXC around function calls can cause many
store to be storing the same value that was just loaded from the same
location:
```
%10 = OpLoad %type %var
OpStore %var %10
```
We want to clean these up very early on because they can cause other
transformations to do a lot of work. For the cases I see, they can be
removed during local-single-block-elim.
For one set of shaders the compile time goes from 248s to 182s. A 26%
improvement.
Part of https://github.com/KhronosGroup/SPIRV-Tools/issues/1494.
We have already disabled common uniform elimination because it created
sequences of loads an entire uniform object, then we extract just a
single element. This caused problems in some drivers, and is just
generally slow because it loads more memory than needed.
However, there are other way to get into this situation, so I've added
a pass that looks specifically for this pattern and removes it when only
a portion of the load is used.
Fixes#1547.
An FClamp instruction forces a values to be within a certain interval.
When the upper or lower bound of the FClamp is a constant and the value
being compared with is a constant, then in some case we can fold the
compared because the entire range is say less than the value.
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1549.
If there is a shader with a variable in the workgroup storage class that
is stored to, but not loadeds, then we know nothing will read those
loads. It should be safe to remove them.
This is implemented in ADCE by treating workgroup variables the same
way that private variables are treated.
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1550.
At this time, DCE will only remove an instruction if it is a combinator.
However, there are certain non-combinator instructions that can be
safely removed if their results are not used. The derivative
instructions are on example.
We are also missing some instructions from the list of combinators
those are added as the same time.
When doing if-conversion, we do not currently move code out of the side
nodes. The reason for this is that it can increase the number of
instructions that get executed because both side nods will have to be
executed now.
In this commit, we add code to move an instruction, and all of the
instructions it depends on, out of a side node and into the header of
the selection construct. However to keep the cost down, we only do it
when the two values in the OpPhi node compute the same value. This way
we have to move only one of the instructions and the other becomes
unused most of the time. So no real extra cost.
Makes the value number table an alalysis in the ir context.
Added more opcodes to list of code motion safe opcodes.
Fixes#1526.
Previously, the loop class used the terms latch and continue block
interchangeably. This patch splits the two and corrects and tests some
uses of the old uses of GetLatchBlock.
This pass will look for adjacent loops that are compatible and legal to
be fused.
Loops are compatible if:
- they both have one induction variable
- they have the same upper and lower bounds
- same initial value
- same condition
- they have the same update step
- they are adjacent
- there are no break/continue in either of them
Fusion is legal if:
- fused loops do not have any dependencies with dependence distance
greater than 0 that did not exist in the original loops.
- there are no function calls in the loops (could have side-effects)
- there are no barriers in the loops
It will fuse all such loops as long as the number of registers used for
the fused loop stays under the threshold defined by
max_registers_per_loop.
Adds support for spliting loops whose register pressure exceeds a user
provided level. This pass will split a loop into two or more loops given
that the loop is a top level loop and that spliting the loop is legal.
Control flow is left intact for dead code elimination to remove.
This pass is enabled with the --loop-fission flag to spirv-opt.
Track live scalars in VDCE as if they were single element vectors.
Handle the extended instructions for GLSL in VDCE.
Handle composite construct instructions in VDCE.
If one of the operands to an OpVectorTimesScalar instruction is zero,
then the result will be the 0 vector. Currently we do not fold the
insturction unless both operands are constants. This change fixes that.
We also allow folding of OpPhi instructions where the incoming values
are either an OpUndef or the OpPhi instruction itself. As with other
cases, this can be simplified to the OpUndef.
Track live scalars in VDCE as if they were single element vectors.
Handle the extended instructions for GLSL in VDCE.
Handle composite construct instructions in VDCE.
Fixes#1511.
Eliminate unused store to variable if followed by store to same
variable in same block.
Most significantly, this cleans up stores made unused by this pass.
These useless stores can inhibit subsequent optimizations, specifically
LocalSingleStoreElim. Eliminating them makes subsequent optimization more
effective.
The main effect of this pass is to simplify the work done by the SSA
rewriter. It catches many local loads/stores that help speeding up the
work done by the main rewriter.
Introduce a pass that does a DCE type analysis for vector elements
instead of the whole vector as a single element.
It will then rewrite instructions that are not used with something else.
For example, an instruction whose value are not used, even though it is
referenced, is replaced with an OpUndef.
For each function, the analysis determine which SSA registers are live
at the beginning of each basic block and which one are killed at
the end of the basic block.
It also includes utilities to simulate the register pressure for loop
fusion and fission.
The implementation is based on the paper "A non-iterative data-flow
algorithm for computing liveness sets in strict ssa programs" from
Boissinot et al.
The local-single-store-elim algorithm is not fundamentally bad.
However, when there are a large number of variables, some of the
maps that are used can become very large. These large data structures
then take a very long time to be destroyed. I've seen cases around 40%
if the time.
I've rewritten that algorithm to not use as much memory. This give a
significant improvement when running a large number of shader through
DXC.
I've also made a small change to local-single-block-elim to delete the
loads that is has replaced. That way local-single-store-elim will not
have to look at those. local-single-store-elim now does the same thing.
The time for one set goes from 309s down to 126s. For another set, the
time goes from 102s down to 88s.
GCD MIV test as described in Chapter 3 of "Optimizing Compilers for
Modern Architectures: A Dependence-Based Approach" by Randy Allen, and
Ken Kennedy.
Delta test as described in Figure 3 of "Practical Dependence Testing" by
Gina Goff, Ken Kennedy, and Chau-Wen Tseng from PLDI '91.
The unordered_set in ADCE that holds all of the live instructions takes
a very long time to be destroyed. In some shaders, it takes over 40% of
the time.
If we look at the unique ids of the live instructions, I believe they
are dense enough make a simple bit vector a good choice for to hold that
data. When I check the density of the bit vector for larger shaders, we
are usually using less than 4 bytes per element in the vector, and
almost always less than 16.
So, in this commit, I introduce a simple bit vector class, and
use it in ADCE.
This help improve the compile time for some shaders on windows by the
40% mentioned above.
Contributes to https://github.com/KhronosGroup/SPIRV-Tools/issues/1328.
For each loop in a function, the pass walks the loops from inner to outer most loop
and tries to peel loop for which a certain amount of iteration can be done before or after the loop.
To limit code growth, peeling will not happen if the growth in code size goes above a configurable threshold.
Provides functionality to perform ZIV and SIV dependency analysis tests
between a load and store within the same loop.
Dependency tests rely on scalar analysis to prove and disprove dependencies
with regard to the loop being analysed.
Based on the 1990 paper Practical Dependence Testing by Goff, Kennedy, Tseng
Adds support for marking loops in the loop nest as IRRELEVANT.
Loops are marked IRRELEVANT if the analysed instructions contain
no induction variables for the loops, i.e. the loops induction
variable is not relevent to the dependence of the store and load.
Adding three rules to fold OpDot (implemented as two).
- When an OpDot has two constants, then fold to the resulting const.
- When one of the inputs is the 0 vector, then fold to zero.
- When one of the inputs is a single 1 with 0s, then rewrite to an
OpCompositeExtract of the appropriate element. This will help find
even more folding opportunities.
Contributes to #709.
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.
Currently OpImageTexelPointer operations are treat like a use of the
pointer, but it does
not look for the memory being referenced to make sure stores are not
removed.
This change teaches it so identify the memory being accessed, and
treats it as if that memory is loaded.
Fixes to #1445.
OpImageTexelPointer acts like a special kind of load. It is not an
array load, but it also cannot be removed the same way a regular
load can. The type of propagation that needs to be done is similar
to what we do for arrays, so I want to merge that code into that
optmization.
Contributers to #1445.
OpImageTexelPointer acts like a special kind of load. It is still
safe to change the storage class of a variable used in a
OpImageTexalPointer instruction.
Contributes to #1445.
When the original code copies an entire array or struct one element at a
time, this turns into a series of OpCompositeInsert instruction followed
by a store of the whole array. We currently miss opportunities in copy
propagate arrays because we do not recognize this as a copy.
This commit adds code to copy propagate arrays to identify this code
pattern.
Also updates the performance passed to run array copy propagation.
The first implementation of MemroyObject, which is used in copy
propagate arrays, forced the access chain to be like the access chains
in OpCompositeExtract. This excluded the possibility of the memory
object from representing an array element that was extracted with a
variable index. Looking at the code, that restriction is not
neccessary. I also see some opportunities for doing this in some real
shaders.
Contributes to #1430.
This patch adds support for the analysis of scalars in loops. It works
by traversing the defuse chain to build a DAG of scalar operations and
then simplifies the DAG by folding constants and grouping like terms.
It represents induction variables as recurrent expressions with respect
to a given loop and can simplify DAGs containing recurrent expression by
rewritting the entire DAG to be a recurrent expression with respect to
the same loop.
When we change the type of an object that gets stored, we do not want to
change the type of the memory location being stored to. In order to
still be able to do the rewrite, we will decompose and rebuild the
object so it is the type that can be stored.
Fixes#1416.
The sprir-v generated from HLSL code contain many copyies of very large
arrays. Not only are these time consumming, but they also cause
problems for drivers because they require too much space.
To work around this, we will implement an array copy propagation. Note
that we will not implement a complete array data flow analysis in order
to implement this. We will be looking for very simple cases:
1) The source must never be stored to.
2) The target must be stored to exactly once.
3) The store to the target must be a store to the entire array, and be a
copy of the entire source.
4) All loads of the target must be dominated by the store.
The hard part is keeping all of the types correct. We do not want to
have to do too large a search to update everything, which may not be
possible, do we give up if we see any instruction that might be hard to
update.
Also in types.h, the element decorations are not stored in an std::map.
This change was done so the hashing algorithm for a Struct is
consistent. With the std::unordered_map, the traversal order was
non-deterministic leading to the same type getting hashed to different
values. See |Struct::GetExtraHashWords|.
Contributes to #1416.
This patch adds a new option --time-report to spirv-opt. For each pass
executed by spirv-opt, the flag prints resource utilization for the pass
(CPU time, wall time, RSS and page faults)
This fixes issue #1378
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).
The loop peeler util takes a loop as input and create a new one before.
The iterator of the duplicated loop then set to accommodate the number
of iteration required for the peeling.
The loop peeling pass that decided to do the peeling and profitability
analysis is left for a follow-up PR.
We are seeing shaders that have multiple returns in a functions. These
functions must get inlined for legalization purposes; however, the
inliner does not know how to inline functions that have multiple
returns.
The solution we will go with it to improve the merge return pass to
handle structured control flow.
Note that the merge return pass will assume the cfg has been cleanedup
by dead branch elimination.
Fixes#857.
Strips reflection info. This is limited to decorations and
decoration instructions related to the SPV_GOOGLE_hlsl_functionality1
extension.
It will remove the OpExtension for SPV_GOOGLE_hlsl_functionality1.
It will also remove the OpExtension for SPV_GOOGLE_decorate_string
if there are no further remaining uses of OpDecorateStringGOOGLE.
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1398
Optimizations should work in the presence of recent
SPV_GOOGLE_decorate_string and SPV_GOOGLE_hlsl_functionality1
SPV_GOOGLE_decorate_string:
- Adds operation OpDecorateStringGOOGLE to decorate an object with decorations
having string operands.
SPV_GOOGLE_hlsl_functionality1:
- Adds HlslSemanticGOOGLE, used to decorate an interface variable with
an HLSL semantic string. Optimizations already preserve those variables
as required because they are interface variables (with uses), independent
of whether they have HLSL decorations.
- Adds HlslCounterBufferGOOGLE, used to associate a buffer with a
counter variable.
Fixes#1391
This reimplementation fixes several issues when removing decorations associated
to an ID (partially addresses #1174 and gives tools for fixing #898), as well
as making it easier to remove groups; a few additional tests have been added.
DecorationManager::RemoveDecoration() will still not delete dead decorations it
created, but I do not think it is its job either; given the following input
```
OpCapability Shader
OpCapability Linkage
OpMemoryModel Logical GLSL450
OpDecorate %2 Restrict
%2 = OpDecorationGroup
OpGroupDecorate %2 %1 %3
OpDecorate %4 Invariant
%4 = OpDecorationGroup
OpGroupDecorate %4 %2
%uint = OpTypeInt 32 0
%1 = OpVariable %uint Uniform
%3 = OpVariable %uint Uniform
```
which of the following two outputs would you expect RemoveDecoration(2) to produce:
```
OpCapability Shader
OpCapability Linkage
OpMemoryModel Logical GLSL450
%uint = OpTypeInt 32 0
%1 = OpVariable %uint Uniform
%3 = OpVariable %uint Uniform
```
or
```
OpCapability Shader
OpCapability Linkage
OpMemoryModel Logical GLSL450
OpDecorate %4 Invariant
%4 = OpDecorationGroup
%uint = OpTypeInt 32 0
%1 = OpVariable %uint Uniform
%3 = OpVariable %uint Uniform
```
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/924
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1174
Remove extension whitelists from transforms that are essentially
combinatorial (and avoiding pointers) or which affect only control flow.
It's very very unlikely an extension will add a new control flow construct.
Remove from:
- dead branch elimination
- dead insertion elimination
- insert extract elimination
- block merge
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1392
The merging types we do not remove other information related to the
types. We simply leave it duplicated, and hope it is removed later.
This is what happens with decorations. They are removed in the next
phase of remove duplicates. However, for OpNames that is not the case.
We end up with two different names for the same id, which does not make
sense.
The solution is to remove the names and decorations for the type being
removed instead of rewriting them to refer to the other type.
Note that it is possible that if the first type does not have a name,
then the types will end up with no name. That is fine because the names
should not have any semantic significance anyway.
The was identified in issue #1372, but this does not fix that issue.
* Also mark function parameters as varying
* Conservatively mark assignment instructions as varying if any input is
varying after attempting to fold
* Added a test to catch this case
* getFloatConstantKind() now handles OpConstantNull
* PerformOperation() now handles OpConstantNull for vectors
* Fixed some instances where we would attempt to merge a division by 0
* added tests
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.
Use indirection through latest_version_spirv.h
Also, when generating enum tables, use the unified1 JSON grammar since
it now has FragmentFullyCoveredEXT but the other JSON grammars don't.
They are starting to fall behind.
Adding basis of arithmetic merging
* Refactored constant collection in ConstantManager
* New rules:
* consecutive negates
* negate of arithmetic op with a constant
* consecutive muls
* reciprocal of div
* Removed IRContext::CanFoldFloatingPoint
* replaced by Instruction::IsFloatingPointFoldingAllowed
* Fixed some bad tests
* added some header comments
Added PerformIntegerOperation
* minor fixes to constants and tests
* fixed IntMultiplyBy1 to work with 64 bit ints
* added tests for integer mul merging
Adding test for vector integer multiply merging
Adding support for merging integer add and sub through negate
* Added tests
Adding rules to merge mult with preceding divide
* Has a couple tests, but needs more
* Added more comments
Fixed bug in integer division folding
* Will no longer merge through integer division if there would be a
remainder in the division
* Added a bunch more tests
Adding rules to merge divide and multiply through divide
* Improved comments
* Added tests
Adding rules to handle mul or div of a negation
* Added tests
Changes for review
* Early exit if no constants are involved in more functions
* fixed some comments
* removed unused declaration
* clarified some logic
Adding new rules for add and subtract
* Fold adds of adds, subtracts or negates
* Fold subtracts of adds, subtracts or negates
* Added tests
It moves all conditional branching and switch whose conditions are loop
invariant and uniform. Before performing the loop unswitch we check that
the loop does not contain any instruction that would prevent it
(barriers, group instructions etc.).
In some shaders there are a lot of very large and deeply nested
structures. This creates a lot of work for scalar replacement. Also,
since commit ca4457b we have been very aggressive as rewriting
variables. This has causes a large increase in compile time in creating
and then deleting the instructions.
To help low the costs, I want to run a cleanup of some of the easy loads
and stores to remove. This reduces the number of symbols sroa has to
work on. It also reduces the amount of code the simplifier has to
simplify because it was not generated by sroa.
To confirm the improvement, I ran numbers on three different sets of
shaders:
Time to run --legalize-hlsl:
Set #1: 55.89s -> 12.0s
Set #2: 1m44s -> 1m40.5s
Set #3: 6.8s -> 5.7s
Time to run -O
Set #1: 18.8s -> 10.9s
Set #2: 5m44s -> 4m17s
Set #3: 7.8s -> 7.8s
Contributes to #1328.
Fixes a bug at the same time. In `UpdateDefUse`, if the definition
already exists, we are not suppose to analyse it again. When you do
the entries for the definition are deleted, and we don't want that.
The check for this was wrong.
This function now checks for side-effects before adding operand
instructions to the dead instruction work list.
Because this fix puts more pressure on IsCombinatorInstruction() to
be correct, this commit adds all OpConstant* and OpType* instructions
to combinator_ops_ set.
Fixes#1341.
When inlining a function call the instructions in the same basic block
as the call get cloned. The clone is added to the set of new blocks
containing the inlined code, and the original instructions are deleted.
This PR will change this so that we simply move the instructions to the
new blocks. This saves on the creation and deletion of the
instructions.
Contributes to #1328.
This change implements instruction folding for arithmetic operations
that are redundant, specifically:
x + 0 = 0 + x = x
x - 0 = x
0 - x = -x
x * 0 = 0 * x = 0
x * 1 = 1 * x = x
0 / x = 0
x / 1 = x
mix(a, b, 0) = a
mix(a, b, 1) = b
Cache ExtInst import id in feature manager
This allows us to avoid string lookups during optimization; for now we
just cache GLSL std450 import id but I can imagine caching more sets as
they become utilized by the optimizer.
Add tests for add/sub/mul/div/mix folding
The tests cover scalar float/double cases, and some vector cases.
Since most of the code for floating point folding is shared, the tests
for vector folding are not as exhaustive as scalar.
To test sub->negate folding I had to implement a custom fixture.
Building the def-use chains is very expensive, so we do not want to
invalidate them it if is not necessary. At the moment, it seems like
most optimizatoins are good at not invalidating the def-use chains, but
simplification does.
This PR get the simlification pass to keep the analysies valid.
Contributes to #1328.
On some shader code we have in our testsuite, Phi insertion is showing
massive compile time slowdowns, particularly during destruction. The
specific shader I was looking at has about 600 variables to keep track
of and around 3200 basic blocks. The algorithm is currently O(var x
blocks), which means maps with around 2M entries. This was taking about
8 minutes of compile time.
This patch changes the tracking of stored variables to be more sparse.
Instead of having every basic block contain all the tracked variables in
the map, they now have only the variables actually stored in that block.
This speeds up deallocation, which brings down compile time to about
1m20s.
Note that this is not the definite fix for this. I will re-write Phi
insertion to use a standard SSA rewriting algorithm
(https://github.com/KhronosGroup/SPIRV-Tools/issues/893).
This contributes to
https://github.com/KhronosGroup/SPIRV-Tools/issues/1328.
I mixed up two cases when folding an OpCompositeExtract that is feed by
and OpCompositeInsert. The specific cases are demonstracted in the new
test. I mixed up the conditions for the cases, and treated one like the
other.
Fixes#1323.
* Now track propagation status and assert on bad statuses
* Added helper methods to access instruction propagation status
* Modified the phi meet operator to properly reflect the paper it is
based on
* Modified SSA edge addition so that all edge are added, but only on
state changes
* Fixed a bug in instruction simulation where interesting conditional
branches would not mark the interesting edge as executed
* Added a test to catch this bug
* Added an ostream operator for SSAPropagator::PropStatus
The simplification pass works better after all of the dead branches are
removed. So swapping them around in the legalization passes. Also
adding the simplification pass to performance passes right after dead
branch elimination.
Added CCP to the legalization passes so we can propagate the constants
into the branchs, and remove as many branches a possible. CCP is
designed to still get opportunities even if the branches are dead, so it
is a good place for it.
Fixes#1118
This change handles all 6 regular comparison types in two variations,
ordered (true if values are ordered *and* comparison is true) and
unordered (true if values are unordered *or* comparison is true).
Ordered comparison matches the default floating-point behavior on host
but we use std::isnan to check ordering explicitly anyway.
This change also slightly reworks the floating-point folding support
code to make it possible to define a folding operation that returns
boolean instead of floating point.
These tests exhaustively test ordered/unordered comparisons for
float/double.
Since for NaN inputs the comparison result doesn't depend on the
comparison function, we just test == and !=; NaN inputs result in true
unordered comparisons and false ordered comparisons.
Registering a constant in constant manager establishes a relation
between instruction that defined it and constant object. On complex
shaders this could result in the constant definition getting removed as
part of one of the DCE pass, and a subsequent simplification pass trying
to use the defining instruction for the constant.
To fix this, we now remove associated constant entries from constant
manager when killing constant instructions; the constant object is still
registered and can be remapped to a new instruction later.
GetDefiningInstruction shouldn't ever return nullptr after this change
so add an assertion to check for that.
In dead branch elimination, we already recognize unreachable continue
blocks, and update OpPhi instruction accordingly. This change adds an
extra check: if the head block has exactly 1 other incoming edge, then
replace the OpPhi with the value from that edge.
Fixes#1314.
unordered_map is not POD. Using it as static may cause problems
when operator new() and operator delete() is customized.
Also changed some function signatures to use const char* instead
of std::string, which will give caller the flexibility to avoid
creating a std::string.
We can fold OpSelect into one of the operands in two cases:
- condition is constant
- both results are the same
Even if the original shader doesn't have either of these, if-conversion
pass sometimes ends up generating instructions like
%7127 = OpSelect %int %3220 %7058 %7058
And this optimization cleans them up.
Adding a map from an id to it set of OpName and OpMemberName
instructions. This will be used in KillNameAndDecorates to kill the
names for the ids that are being removed.
In my test, the compile time for 50 shaders went from 1m57s to 55s.
This was on linux using the release build.
Fixes#1290.
This patch adds initial support for loop unrolling in the form of a
series of utility classes which perform the unrolling. The pass can
be run with the command spirv-opt --loop-unroll. This will unroll
loops within the module which have the unroll hint set. The unroller
imposes a number of requirements on the loops it can unroll. These are
documented in the comments for the LoopUtils::CanPerformUnroll method in
loop_utils.h. Some of the restrictions will be lifted in future patches.
There seems to only be a single location where the def-use manager is
used. It is to get information about a type. We can do that with the
type manager instead.
Fixes#1285
Implementation of the simplification pass.
- Create pass that calls the instruction folder on each instruction and
propagate instructions that fold to a copy. This will do copy
propagation as well.
- Did not use the propagator engine because I want to modify the instruction
as we go along.
- Change folding to not allocate new instructions, but make changes in
place. This change had a big impact on compile time.
- Add simplification pass to the legalization passes in place of
insert-extract elimination.
- Added test cases for new folding rules.
- Added tests for the simplification pass
- Added a method to the CFG to apply a function to the basic blocks in
reverse post order.
Contributes to #1164.
* Added TypeManager::RebuildType
* rebuilds the type and its constituent types in terms of memory owned
by the manager.
* Used by TypeManager::RegisterType to properly allocate memory
* Adding an unit test to expose the issue
* Added some tests to provide coverage of RebuildType
* Added an accessor to the target pointer for a forward pointer
The combinator initialization was only looking at the capabilities
in the shader and not the inferred capabilities. Geometry and tessellation
shaders were not setting the Shader capability which is inferred. So the
combinator set was not initialized correctly causing problems for ADCE.
Create the folding engine that will
1) attempt to fold an instruction.
2) iterates on the folding so small folding rules can be easily combined.
3) insert new instructions when needed.
I've added the minimum number of rules needed to test the features above.
* Moved initial insert/extract passes later to cover more opportunities
* Added an extra set of passes to clean up opportunities exposed later
in the pipeline
This patch adds LoopUtils class to handle some loop related transformations. For now it has 2 transformations that simplifies other transformations such as loop unroll or unswitch:
- Dedicate exit blocks: this ensure that all exit basic block
(out-of-loop basic blocks that have a predecessor in the loop)
have all their predecessors in the loop;
- Loop Closed SSA (LCSSA): this ensure that all definitions in a loop are used inside the loop
or in a phi instruction in an exit basic block.
It also adds the following capabilities:
- Loop::IsLCSSA to test if the loop is in a LCSSA form
- Loop::GetOrCreatePreHeaderBlock that can build a loop preheader if required;
- New methods to allow on the fly updates of the loop descriptors.
- New methods to allow on the fly updates of the CFG analysis.
- Instruction::SetOperand to allow expression of the index relative to Instruction::NumOperands (to be compatible with the index returned by DefUseManager::ForEachUse)
Creates a pass that will remove instructions that are invalid for the
current shader stage. For the instruction to be considered for replacement
1) The opcode must be valid for a shader modules.
2) The opcode must be invalid for the current shader stage.
3) All entry points to the module must be for the same shader stage.
4) The function containing the instruction must be reachable from an entry point.
Fixes#1247.
* Had to remove templating from InstructionBuilder as a result
* now preserved analyses are specified as a constructor argument
* updated tests and uses
* changed static_assert to a runtime assert
* this should probably get further changes in the future
* When handling unreachable merges and continues, do not optimize to the
same IR
* pass did not check whether the unreachable blocks were in the
optimized form before transforming them
* added a test to catch this issue
* Should handle all possibilities
* Stricter checks for what is disallowed:
* header and header
* merge and merge
* Allow header and merge blocks to be merged
* Erases the structured control declaration if merging header and
merge blocks together.
* If the dead branch elim is performed on a module without structured
control flow, the OpSelectionMerge may not be present
* Add a check for pointer validity before dereferencing
* Added a test to catch the bug
* Forces traversal of phis if the def has changed to varying
* Mark a phi as varying if all incoming values are varying
* added a test to catch the bug
This adds Dead Insert Elimination to the end of the
--eliminate-insert-extract pass. See the new tests for examples of code
that will benefit.
Essentially, this removes OpCompositeInsert instructions which are not
used, either because there is no instruction which uses the value at the
index it is inserted, or because a subsequent insert intercepts any such
use.
This code has been seen to remove significant amounts of dead code from
real-life HLSL shaders being ported to Vulkan. In fact, it is needed to
remove dead texture samples which cause Vulkan validation layer errors
(unbound textures and samplers) if not removed . Such DCE is thus
required for fxc equivalence and legalization.
This analysis operates across "chains" of Inserts which can also contain
Phi instructions.
* Handles simple cases only
* Identifies phis in blocks with two predecessors and attempts to
convert the phi to an select
* does not perform code motion currently so the converted values must
dominate the join point (e.g. can't be defined in the branches)
* limited for now to two predecessors, but can be extended to handle
more cases
* Adding if conversion to -O and -Os
The current folding routines have a very cumbersome interface, make them
harder to use, and not a obvious how to extend.
This change is to create a new interface for the folding routines, and
show how it can be used by calling it from CCP.
This does not make a significant change to the behaviour of CCP. In
general it should produce the same code as before; however it is
possible that an instruction that takes 32-bit integers as inputs and
the result is not a 32-bit integer or bool will not be folded as before.
It seems like andriod has a problem with INT32_MAX and the like. I'll
explicitly define those if the are not already defined.
The class factorize the instruction building process.
Def-use manager analysis can be updated on the fly to maintain coherency.
To be updated to take into account more analysis.
* AddToWorklist can now be called unconditionally
* It will only add instructions that have not already been marked as
live
* Fixes a case where a merge was not added to the worklist because the
branch was already marked as live
* Added two similar tests that fail without the fix
We have come across a driver bug where and OpUnreachable inside a loop
is causing the shader to go into an infinite loop. This commit will try
to avoid this bug by turning OpUnreachable instructions that are
contained in a loop into branches to the loop merge block.
This is not added to "-O" and "-Os" because it should only be used if
the driver being targeted has this problem.
Fixes#1209.
This ensure that all basic blocks in a function have a valid entry the CFG object.
The entry block has no predecessors but remains a valid basic block
for which we might want to query the number of predecessors.
Some unreachable basic blocks may not have predecessors as well.
At the moment specialization constants look like constants to ccp. This
causes a problem because they are handled differently by the constant
manager.
I choose to simply skip over them, and not try to add them to the value
table. We can do specialization before ccp if we want to be able to
propagate these values.
Fixes#1199.
With work that Alan has done, some passes have become redundant. ADCE
now removed unused variables. Dead branch elimination removes
unreachable blocks. This means we can remove CFG Cleanup and dead
variable elimination.
The current code expects the users of the constant manager to initialize
it with all of the constants in the module. The problem is that you do
not want to redo the work multiple times. So I decided to move that
code to the constructor of the constant manager. This way it will
always be initialized on first use.
I also removed an assert that expects all constant instructions to be
successfully mapped. This is because not all OpConstant* instruction
can map to a constant, and neither do the OpSpecConstant* instructions.
The real problem is that an OpConstantComposite can contain a member
that is OpUndef. I tried to treat OpUndef like OpConstantNull, but this
failed because an OpSpecConstantComposite with an OpUndef cannot be
changed to an OpConstantComposite. Since I feel this case will not be
common, I decided to not complicate the code.
Fixes#1193.
* Added for Instruction, BasicBlock, Function and Module
* Uses new disassembly functionality that can disassemble individual
instructions
* For debug use only (no caching is done)
* Each output converts module to binary, parses and outputs an
individual instruction
* Added a test for whole module output
* Disabling Microsoft checked iterator warnings
* Updated check_copyright.py to accept 2018
* 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()
This improves Extract replacement to continue through VectorShuffle.
It will also handle Mix with 0.0 or 1.0 in the a-value of the desired
component.
To facilitate optimization of VectorShuffle, the algorithm was refactored
to pass around the indices of the extract in a vector rather than pass the
extract instruction itself. This allows the indices to be modified as the
algorithm progresses.
The current folding routines have a very cumbersome interface, make them
harder to use, and not a obvious how to extend.
This change is to create a new interface for the folding routines, and
show how it can be used by calling it from CCP.
This does not make a significant change to the behaviour of CCP. In
general it should produce the same code as before; however it is
possible that an instruction that takes 32-bit integers as inputs and
the result is not a 32-bit integer or bool will not be folded as before.
Modified ADCE to remove dead globals.
* Entry point and execution mode instructions are marked as alive
* Reachable functions and their parameters are marked as alive
* Instruction deletion now deferred until the end of the pass
* Eliminated dead insts set, added IsDead to calculate that value
instead
* Ported applicable dead variable elimination tests
* Ported dead constant elim tests
Added dead function elimination to ADCE
* ported dead function elim tests
Added handling of decoration groups in ADCE
* Uses a custom sorter to traverse decorations in a specific order
* Simplifies necessary checks
Updated -O and -Os pass lists.
Pass now paints live blocks and fixes constant branches and switches as
it goes. No longer requires structured control flow. It also removes
unreachable blocks as a side effect. It fixes the IR (phis) before doing
any code removal (other than terminator changes).
Added several unit tests for updated/new functionality.
Does not remove dead edge from a phi node:
* Checks that incoming edges are live in order to retain them
* Added BasicBlock::IsSuccessor
* added test
Fixing phi updates in the presence of extra backedge blocks
* Added tests to catch bug
Reworked how phis are updated
* Instead of creating a new Phi and RAUW'ing the old phi with it, I now
replace the phi operands, but maintain the def/use manager correctly.
For unreachable merge:
* When considering unreachable continue blocks the code now properly
checks whether the incoming edge will continue to be live.
Major refactoring for review
* Broke into 4 major functions
* marking live blocks
* marking structured targets
* fixing phis
* deleting blocks
This fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1143.
When an instruction transitions from constant to bottom (varying) in the
lattice, we were telling the propagator that the instruction was
varying, but never updating the actual value in the values table.
This led to incorrect value substitutions at the end of propagation.
The patch also re-enables CCP in -O and -Os.
Add post-order tree iterator.
Add DominatorTreeNode extensions:
- Add begin/end methods to do pre-order and post-order tree traversal from a given DominatorTreeNode
Add DominatorTree extensions:
- Add begin/end methods to do pre-order and post-order tree traversal
- Tree traversal ignore by default the pseudo entry block
- Retrieve a DominatorTreeNode from a basic block
Add loop descriptor:
- Add a LoopDescriptor class to register all loops in a given function.
- Add a Loop class to describe a loop:
- Loop parent
- Nested loops
- Loop depth
- Loop header, merge, continue and preheader
- Basic blocks that belong to the loop
Correct a bug that forced dominator tree to be constantly rebuilt.
Turn `Linker::Link()` into free functions
As very little information was kept in the Linker class, we can get rid
of the whole class and have the `Link()` as free functions instead; the
environment target as well as the consumer are passed along through an
`spv_context` object.
The resulting linked_binary is passed as a pointer rather than a
reference to follow the Google C++ Style guidelines.
Addresses remaining comments from
https://github.com/KhronosGroup/SPIRV-Tools/pull/693 about the SPIR-V
linker.
Fix variable naming in the linker
Some of the variables were using mixed case, which did not follow the
Google C++ Style guidelines.
Linker: Use EXPECT_EQ when possible and update some test
* Replace occurrences of ASSERT_EQ by EXPECT_EQ when possible;
* Reformulated some of the error messages;
* Added the symbol name in the error message when there is a type or
decoration mismatch between the imported and exported declarations.
Opt: List all duplicates removed by RemoveDuplicatePass in the header
Opt: Make the const version of GetLabelInst() return a pointer
For consistency with the non-const version, as well as other similar
functions.
Opt: Rename function_end to EndInst()
As pointed out by dneto0 the previous name was quite confusing and could
be mistaken with a function returning an end iterator.
Also change the return type of the const version to a pointer rather
than a reference, for consistency.
Opt: Add performance comment to RemoveDuplicateTypes and decorations
This comment was requested during the review of
https://github.com/KhronosGroup/SPIRV-Tools/pull/693.
Opt: Add comments and fix variable naming in RemoveDuplicatePass
* Add missing comments to private functions;
* Rename variables that were using mixed case;
* Add TODO for moving AreTypesEqual out.
Linker: Remove commented out code and add TODOs
Linker: Merged together strings that were too much splitted
Implement a C++ RAII wrapper around spv_context
In value numbering, we treat loads and stores of images, ie OpImageLoad,
as a memory operation where it is interested in the "base address" of
the instruction. In those cases, it is an image instruction.
The problem is that `Instruction::GetBaseAddress()` does not account for
the image instructions, so the assert at the end to make sure it found
a valid base address for its addressing mode fails.
The solution is to look at the load/store instruction to determine how
the assertion should be done.
Fixes#1160.
This fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1159. I
had missed a nuance in the original algorithm. When simulating Phi
instructions, the SSA edges out of a Phi instruction should never be
added to the list of edges to simulate.
Phi instructions can be in SSA def-use cycles with other Phi
instructions. This was causing the propagator to fall into an infinite
loop when the same def-use edge kept being added to the queue.
The original algorithm in the paper specifically separates the visit of
a Phi instruction vs the visit of a regular instruction. This fix makes
the implementation match the original algorithm.
When deleting branches and blocks, also remove them from
the backedges set, in case they were there.
This prevents us from keeping stale pointers to deleted Instruction
objects. That memory could be used later by another instruction,
incorrectly signaling that something has a backedge reference, and
the dead branch eliminator could end up deleting live blocks.
Adds accessor method ir::BasicBlock::terminator
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1168
Adds optimizer API to write disassembly to a given output stream
before each pass, and after the last pass.
Adds spirv-opt --print-all option to write disassembly to stderr
before each pass, and after the last pass.
I've a few passes the legalization passes. The first is to add the
more specialized load-store removal passes to help improve the compile
time, as was suggested in #1118.
I've also added dead branch elimination while we wait for the behaviour
of dead branch elimination to be folded into CFG cleanup.
I did not add CCP because it seems like most of the constant propagation
what is needed is already being done by the load-store removal passes,
which call `ReplaceAllUsesWith`. We can reconsider this if needed.
Calling `ToNop` leaves around instructions that are pointless. In
general it is better to remove the instruction completely. That way
other optimizations will not need to look at them.
Fixes https://github.com/KhronosGroup/SPIRV-Tools/issues/1003.
In CCP we should not need to insert Phi nodes because CCP never looks at
loads/stores. This required adjusting two tests that relied on Phi
instructions being inserted. I changed the tests to have the Phi
instructions pre-inserted.
I also added a new test to make sure that CCP does not try to look
through stores and loads.
Finally, given that CCP does not handle loads/stores, it's better to run
mem2reg before it. I've changed the -O/-Os schedules to run local
multi-store elimination before CCP.
Although this is just an efficiency fix for CCP, it is
also working around a bug in Phi insertion. When Phi instructions are
inserted, they are never associated a basic block. This causes a
segfault when the propagator tries to lookup CFG edges when analyzing
Phi instructions.
This addresses review feedback for the CCP implementation (which fixes
https://github.com/KhronosGroup/SPIRV-Tools/issues/889).
This adds more protection around the folding of instructions that would
not be supported by the folder.
This implements the conditional constant propagation pass proposed in
Constant propagation with conditional branches,
Wegman and Zadeck, ACM TOPLAS 13(2):181-210.
The main logic resides in CCPPass::VisitInstruction. Instruction that
may produce a constant value are evaluated with the constant folder. If
they produce a new constant, the instruction is considered interesting.
Otherwise, it's considered varying (for unfoldable instructions) or
just not interesting (when not enough operands have a constant value).
The other main piece of logic is in CCPPass::VisitBranch. This
evaluates the selector of the branch. When it's found to be a known
value, it computes the destination basic block and sets it. This tells
the propagator which branches to follow.
The patch required extensions to the constant manager as well. Instead
of hashing the Constant pointers, this patch changes the constant pool
to hash the contents of the Constant. This allows the lookups to be
done using the actual values of the Constant, preventing duplicate
definitions.
In order to keep track of all of the implicit capabilities as well as
the explicit ones, we will add them all to the feature manager. That is
the object that needs to be queried when checking if a capability is
enabled.
The name of the "HasCapability" function in the module was changed to
make it more obvious that it does not check for implied capabilities.
Keep an spv_context and AssemblyGrammar in IRContext
