This change adds a folding rule which transforms x * y - a and a - x * y
into FMA(x, y, -a) and FMA(-x, y, a), respectively.
While the SPIR-V instruction count remains the same, target instruction
sets typically feature FMA instruction variants that can negate an
operand. Also this transformation may unlock further optimizations which
eliminate the negation.
(Google bug: b/226145988)
* Add move folding for composite instructions
Fold chains of insert into construct
If a chain of OpCompositeInsert instruction write to every element of a
composite object, then we can replace it with an OpCompositeConstruct.
Fold a construct fed by extracts to a single extract
We already fold an OpCompositeConstruct when it is simlpy reconstructing
an object that was decomposed by a series of OpCompositeExtract
instructions. However, we do not do that if that object is an element
of a larger object.
I have updated the rule, so that if the original object is a an element
of a larger object, then the OpCompositeConstruct is replaced with a
single OpCompositeExtract from the larger object.
Fixes#4371.
This reverts commit 671f6e633f.
PR #4783 was reverted because it caused OpenCL CTS failures for clvk.
The was in clspv, which was not adding the no contract decoration when
it was required. This has been fixed in
https://github.com/google/clspv/pull/845. We can now reapply #4783.
Adding Fma instruction can speed up the code. This was requested by
swiftshader, so they do not have to do this analysis themselves. It can
also help reduce the code size, and the work the ICD compilers have to
do.
When folding a vector shuffle feeding a vector shuffle, we do not
propagate an 0xFFFFFFFF, which has a special meaning, correctly. We
adjust the value making it lose it meaning as an undefined value.
Fixes#4581
* Fix extract with out-of-bounds index
When folding a OpCompositeExtract that is fed by an
OpCompositeConstruct, we handle and out of bounds
index, but only in the case where the result of the
OpCompostiteConstruct is a struct. This change
refactors that folding rule and then improves it to
handle an out-of-bounds access when the result of the
OpCompositeConstruct is a vector.
In SPIR-V, integers use 2s complement representation, so that signed
integer overflow and underflow is well defined. However, the constant
folder was causing overflow / underflow at the C++ level. This change
avoids such overflows by performing constant folding for IAdd, ISub and
IMul in the context of unsigned values, which works because signedness
is irrelevant according to the SPIR-V semantics for these instructions.
Fixes#4510.
Reject folding comparisons with unfoldable types.
Fixes#3343
When CCP is evaluating an instruction, it was trying to fold a
comparison with 64 bit integers. This was causing a fold failure later
since the folder still cannot deal with 64 bit integers.
* Implement constant folding for many transcendentals
This change adds support for folding of sin/cos/tan/asin/acos/atan,
exp/log/exp2/log2, sqrt, atan2 and pow.
The mechanism allows to use any C function to implement folding in the
future; for now I limited the actual additions to the most commonly used
intrinsics in the shaders.
Unary folder had to be tweaked to work with extended instructions - for
extended instructions, constants.size() == 2 and constants[0] ==
nullptr. This adjustment is similar to the one binary folder already
performs.
Fixes#1390.
* Fix Android build
On old versions of Android NDK, we don't get std::exp2/std::log2
because of partial C++11 support.
We do get ::exp2, but not ::log2 so we need to emulate that.
* Don't crash when folding construct of empty struct
An OpCompositeConstruct of an empty struct will be folded to a constant
under normal circumstances. However, if the id limit has been reached
and the constant cannot be generated, then other folding rules will be
tried.
These rules do not handle the case of an empty struct. We add allow it
to be handled.
Fixes http://crbug/1030194
* Changes based on the review.
Implements the following simplifications:
(a - b) + b => a
(a * b) + (a * c) => a * (b + c)
Also adds logic to simplification to handle rules that create new operations
that might need simplification, such as the second rule above.
Only perform the second simplification if the multiplies have the add as their
only use. Otherwise this is a deoptimization of size and performance.
* Handle extract with no indexes
It is possible that OpCompositeExtract instructions will not have any
indexes. This is not handled well by scalar replacement and instruction
folding.
Fixes https://crbug.com/1006435
* Fix typo.
* Refactor instruction folders
We want to refactor the instruction folder to allow different sets of
rules to be added to the instruction folder. We might want different
sets of rules in different circumstances.
We also need a way to add rules for extended instructions. Changes are
made to the FoldingRules class and ConstFoldingRules class to enable
that.
We added tests to check that we can fold extended instructions using the
new framework.
At the same time, I noticed that there were two tests that did not tests
what they were suppose to. They could not be easily salvaged. #2813 was
opened to track adding the new tests.
This fixes#2608.
The original test case had an out-of-bounds reference that ended up
folding into OpCompositeExtract that was indexing right outside the
constant composite.
The returned constant would then cause a segfault during constant
propagation.
* Fix OpDot folding of half float vectors.
The code that folds OpDot does not handle half floats correctly. After
trying to multiple the first components, we get a nullptr because we
don't fold half float values. This nullptr gets passed to the code that
does the addition, and causes an assert.
Fixes#2405.
* Remove use of deprecated googletest macro
INSTANTIATE_TEST_CASE_P has been deprecated. We need to use
INSTANTIATE_TEST_SUITE_P instead.
* Remove extra commas from test suites.
In C++, a bit shift of the same size as the type is undefined, but it is
defined in spir-v. When folding those cases, we have to be careful. We
cannot simply do the shift in C++.
Fixes https://crbug.com/917697.
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.
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.
* 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
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
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.
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.
