Implement `llvm.coro.await.suspend` intrinsics, to deal with performance
regression after prohibiting `.await_suspend` inlining, as suggested in
#64945.
Actually, there are three new intrinsics, which directly correspond to
each of three forms of `await_suspend`:
```
void llvm.coro.await.suspend.void(ptr %awaiter, ptr %frame, ptr @wrapperFunction)
i1 llvm.coro.await.suspend.bool(ptr %awaiter, ptr %frame, ptr @wrapperFunction)
ptr llvm.coro.await.suspend.handle(ptr %awaiter, ptr %frame, ptr @wrapperFunction)
```
There are three different versions instead of one, because in `bool`
case it's result is used for resuming via a branch, and in
`coroutine_handle` case exceptions from `await_suspend` are handled in
the coroutine, and exceptions from the subsequent `.resume()` are
propagated to the caller.
Await-suspend block is simplified down to intrinsic calls only, for
example for symmetric transfer:
```
%id = call token @llvm.coro.save(ptr null)
%handle = call ptr @llvm.coro.await.suspend.handle(ptr %awaiter, ptr %frame, ptr @wrapperFunction)
call void @llvm.coro.resume(%handle)
%result = call i8 @llvm.coro.suspend(token %id, i1 false)
switch i8 %result, ...
```
All await-suspend logic is moved out into a wrapper function, generated
for each suspension point.
The signature of the function is `<type> wrapperFunction(ptr %awaiter,
ptr %frame)` where `<type>` is one of `void` `i1` or `ptr`, depending on
the return type of `await_suspend`.
Intrinsic calls are lowered during `CoroSplit` pass, right after the
split.
Because I'm new to LLVM, I'm not sure if the helper function generation,
calls to them and lowering are implemented in the right way, especially
with regard to various metadata and attributes, i. e. for TBAA. All
things that seemed questionable are marked with `FIXME` comments.
There is another detail: in case of symmetric transfer raw pointer to
the frame of coroutine, that should be resumed, is returned from the
helper function and a direct call to `@llvm.coro.resume` is generated.
C++ standard demands, that `.resume()` method is evaluated. Not sure how
important is this, because code has been generated in the same way
before, sans helper function.
This patch fixes the issues introduced in
bb5c3899d1.
I moved the check for the instruction to be div before I check for the
fast math flags which resolves the crash in
```
float a, b;
double sqrt();
void c() { b = a / sqrt(a); }
```
---------
Co-authored-by: Matt Arsenault <arsenm2@gmail.com>
Instead of taking the sign of the cast operation as the required since
for the transform, only force a sign if an operation is maybe
negative.
This gives us more flexability when checking if the floats are safely
converable to integers.
Closes#84389
This improves overall analysis for minbitwidth in SLP. It allows to
analyze the trees with store/insertelement root nodes. Also, instead of
using single minbitwidth, detected from the very first analysis stage,
it tries to detect the best one for each trunc/ext subtree in the graph
and use it for the subtree.
Results in better code and less vector register pressure.
Metric: size..text
Program size..text
results results0 diff
test-suite :: SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant.test 92549.00 92609.00 0.1%
test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test 663381.00 663493.00 0.0%
test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test 663381.00 663493.00 0.0%
test-suite :: MultiSource/Benchmarks/Bullet/bullet.test 307182.00 307214.00 0.0%
test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 2040257.00 2040273.00 0.0%
test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 12396098.00 12395858.00 -0.0%
test-suite :: External/SPEC/CINT2006/445.gobmk/445.gobmk.test 909944.00 909768.00 -0.0%
SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant - 4 scalar
instructions remain scalar (good).
Spec2017/x264 - the whole function idct4x4dc is vectorized using <16
x i16> instead of <16 x i32>, also zext/trunc are removed. In other
places last vector zext/sext removed and replaced by
extractelement + scalar zext/sext pair.
MultiSource/Benchmarks/Bullet/bullet - reduce or <4 x i32> replaced by
reduce or <4 x i8>
Spec2017/imagick - Removed extra zext from 2 packs of the operations.
Spec2017/parest - Removed extra zext, replaced by extractelement+scalar
zext
Spec2017/blender - the whole bunch of vector zext/sext replaced by
extractelement+scalar zext/sext, some extra code vectorized in smaller
types.
Spec2006/gobmk - fixed cost estimation, some small code remains scalar.
Original Pull Request: https://github.com/llvm/llvm-project/pull/84334
The patch has the same functionality (no test changes, no changes in
benchmarks) as the original patch, just has some compile time
improvements + fixes for xxhash unittest, discovered earlier in the
previous version of the patch.
Reviewers:
Pull Request: https://github.com/llvm/llvm-project/pull/84536
We're going from a single use to two independent uses, we need these two
to see consistent values for undef. As an example, consider x = 0x2 when
y = 0b00u1. If the sub use picks 0b0001 and the cmp use picks 0b0011,
that would be incorrect.
UBSAN checks can be too expensive to be used
in release binaries. However not all code affect
performace in the same way. Removing small
number of checks in hot code we can performance
loss, preserving most of the checks.
Targets the dynamic realignment pattern of `(Ptr + Align - 1) & -Align;`
as implemented by gep then ptrmask.
Specifically, when the pointer already has alignment information,
dynamically realigning it to less than is already known should be a
no-op. Discovered while writing test cases for another patch.
For the zero low bits of a known aligned pointer, adding the gep index
then removing it with a mask is a no-op. Folding the ptrmask effect
entirely into the gep is the ideal result as that unblocks other
optimisations that are not aware of ptrmask.
In some other cases the gep is known to be dead and is removed without
changing the ptrmask.
In the least effective case, this transform creates a new gep with a
rounded-down index and still leaves the ptrmask unchanged. That
simplified gep is still a minor improvement, geps are cheap and ptrmask
occurs in address calculation contexts so I don't think it's worth
special casing to avoid the extra instruction.
This improves overall analysis for minbitwidth in SLP. It allows to
analyze the trees with store/insertelement root nodes. Also, instead of
using single minbitwidth, detected from the very first analysis stage,
it tries to detect the best one for each trunc/ext subtree in the graph
and use it for the subtree.
Results in better code and less vector register pressure.
Metric: size..text
Program size..text
results results0 diff
test-suite :: SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant.test 92549.00 92609.00 0.1%
test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test 663381.00 663493.00 0.0%
test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test 663381.00 663493.00 0.0%
test-suite :: MultiSource/Benchmarks/Bullet/bullet.test 307182.00 307214.00 0.0%
test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 2040257.00 2040273.00 0.0%
test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 12396098.00 12395858.00 -0.0%
test-suite :: External/SPEC/CINT2006/445.gobmk/445.gobmk.test 909944.00 909768.00 -0.0%
SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant - 4 scalar
instructions remain scalar (good).
Spec2017/x264 - the whole function idct4x4dc is vectorized using <16
x i16> instead of <16 x i32>, also zext/trunc are removed. In other
places last vector zext/sext removed and replaced by
extractelement + scalar zext/sext pair.
MultiSource/Benchmarks/Bullet/bullet - reduce or <4 x i32> replaced by
reduce or <4 x i8>
Spec2017/imagick - Removed extra zext from 2 packs of the operations.
Spec2017/parest - Removed extra zext, replaced by extractelement+scalar
zext
Spec2017/blender - the whole bunch of vector zext/sext replaced by
extractelement+scalar zext/sext, some extra code vectorized in smaller
types.
Spec2006/gobmk - fixed cost estimation, some small code remains scalar.
Reviewers: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/84334
pr73894.ll is failing on a number of non-AArch64 buildbots. I'm not
certain that this is a proper fix, but I think it's best to move the
test to the test/Transforms/LoopVectorize/AArch64/ directory and replace
the triple with one commonly used in that directory.
llvm#73894
This extends the work from 7755c26 to all of the different backend
taken count kinds that we print for the scev analysis printer. As
before, the goal is to cut down on confusion as i4 -1 is a very
different (unsigned) value from i32 -1.
Instead of generating a <1 x i1> active lane mask intrinsic, generate
the equivalent scalar ICMP instead. This allows us to avoid
unnecessarily extracting the scalar part from the vector mask.
Fixes llvm#73894.
The full fold is one of the following:
1) `(fp_binop ({s|u}itofp x), ({s|u}itofp y))`
-> `({s|u}itofp (int_binop x, y))`
2) `(fp_binop ({s|u}itofp x), FpC)`
-> `({s|u}itofp (int_binop x, (fpto{s|u}i FpC)))`
And support the following binops:
`fmul` -> `mul`
`fadd` -> `add`
`fsub` -> `sub`
Proofs: https://alive2.llvm.org/ce/z/zuacA8
The proofs timeout, so they must be reproduced locally.
Closes#82555
When printing the result of SCEV's analysis, we can avoid printing
the predicated backedge taken count and the predicates if the predicates
are empty and no new information is provided. This helps to reduce the
verbosity of the output.
When printing the result of the analysis, i8 -1 and i64 -1 are quite
different in terms of analysis quality. In a recent conversion with
a new contributor, we ran into exactly this confusion.
Adding the type for constant scevs more globally seems worthwhile, but
introduces a much larger test diff. I'm splitting this off first since
it addresses the immediate need, and then going to do some further
changes to clarify a few related bits of analysis result output.
Instructions in unreachable basic blocks are removed, but terminators
are not. In this case, even instructions that are only referenced by
a terminator, such as a return instruction, cannot be processed
properly.
This patch changes the operand of a return instruction in an
unreachable basic block to poison if it refers to the instruction,
allowing the instruction to be properly processed.
Fixes#65107.
When replacing with a non-constant, it's possible that the result of the
simplification is actually more complicated than the original, and may
result in an infinite combine loop.
Mitigate the issue by requiring that either the replacement or
simplification result is constant, which should ensure that it's
simpler. While this check is crude, it does not appear to cause
optimization regressions in real-world code in practice.
Fixes https://github.com/llvm/llvm-project/issues/83127.
So that we can benefit from some instcombine optimizations.
This PR contains two commits: the first is for adding tests and the
second is for the optimization.
In case of loads/stores from an immediate address, avoid rematerializing
the constant for every block and allow consthoist to hoist it to the
entry block.
This improves overall analysis for minbitwidth in SLP. It allows to
analyze the trees with store/insertelement root nodes. Also, instead of
using single minbitwidth, detected from the very first analysis stage,
it tries to detect the best one for each trunc/ext subtree in the graph
and use it for the subtree.
Results in better code and less vector register pressure.
Metric: size..text
Program size..text
results results0 diff
test-suite :: SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant.test 92549.00 92609.00 0.1%
test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test 663381.00 663493.00 0.0%
test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test 663381.00 663493.00 0.0%
test-suite :: MultiSource/Benchmarks/Bullet/bullet.test 307182.00 307214.00 0.0%
test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 2040257.00 2040273.00 0.0%
test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 12396098.00 12395858.00 -0.0%
test-suite :: External/SPEC/CINT2006/445.gobmk/445.gobmk.test 909944.00 909768.00 -0.0%
SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant - 4 scalar
instructions remain scalar (good).
Spec2017/x264 - the whole function idct4x4dc is vectorized using <16
x i16> instead of <16 x i32>, also zext/trunc are removed. In other
places last vector zext/sext removed and replaced by
extractelement + scalar zext/sext pair.
MultiSource/Benchmarks/Bullet/bullet - reduce or <4 x i32> replaced by
reduce or <4 x i8>
Spec2017/imagick - Removed extra zext from 2 packs of the operations.
Spec2017/parest - Removed extra zext, replaced by extractelement+scalar
zext
Spec2017/blender - the whole bunch of vector zext/sext replaced by
extractelement+scalar zext/sext, some extra code vectorized in smaller
types.
Spec2006/gobmk - fixed cost estimation, some small code remains scalar.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/78976
It's common to delete some instructions after using SCEVExpander,
while it is still live (but will not be used afterwards). In that
case, the AssertingVH may trigger. Replace it with a PoisoningVH
so that we only detect the case where the SCEVExpander actually is
used in a problematic fashion after the instruction removal.
The alternative would be to add clear() calls to more code paths.
Fixes https://github.com/llvm/llvm-project/issues/83404.
762f762504/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp (L394-L407)
Comment from @topperc:
> This transforms assumes the mask is a non-zero splat. We only know its
a splat and not provably all 0s. The mask is a constexpr that includes
the address of the global variable. We can't resolve the constant
expression to an exact value.
Fixes#83947.
When inlining across functions with different target features, we
perform roughly two checks:
1. The caller features must be a superset of the callee features.
2. Calls in the callee cannot use types where the target features would
change the call ABI (e.g. by changing whether something is passed in a
zmm or two ymm registers). The latter check is very crude right now.
The latter check currently also catches inline asm "calls". I believe
that inline asm should be excluded from this check, as it is independent
from the usual call ABI, and instead governed by the inline asm
constraint string.
Fixes https://github.com/llvm/llvm-project/issues/67054.
This patch adds support for canonicalization of icmp with a scalable
splat. Some optimizations assume that `icmp pred X, APInt C` is in
canonical form.
Fixes https://github.com/llvm/llvm-project/issues/83931.
