diff options
author | Amara Emerson <amara@apple.com> | 2021-08-18 00:19:58 -0700 |
---|---|---|
committer | Amara Emerson <amara@apple.com> | 2021-08-19 16:38:52 -0700 |
commit | 95ac3d15e9fe86d9b51b51d02cb3c1640bf30dee (patch) | |
tree | eb928600c2d369838959eb0ada71055ff07e0dcb /llvm/lib/Target | |
parent | fbb8e772ec501a1b71643db90e9c6445e17d7cac (diff) | |
download | llvm-95ac3d15e9fe86d9b51b51d02cb3c1640bf30dee.zip llvm-95ac3d15e9fe86d9b51b51d02cb3c1640bf30dee.tar.gz llvm-95ac3d15e9fe86d9b51b51d02cb3c1640bf30dee.tar.bz2 |
[AArch64][GlobalISel] Add G_VECREDUCE fewerElements support for full scalarization.
For some reductions like G_VECREDUCE_OR on AArch64, we need to scalarize
completely if the source is <= 64b. This change adds support for that in
the legalizer. If the source has a pow-2 num elements, then we can do
a tree reduction using the scalar operation in the individual elements.
Otherwise, we just create a sequential chain of operations.
For AArch64, we only need to scalarize if the input is <64b. If it's great than
64b then we can first do a fewElements step to 64b, taking advantage of vector
instructions until we reach the point of scalarization.
I also had to relax the verifier checks for reductions because the intrinsics
support <1 x EltTy> types, which we lower to scalars for GlobalISel.
Differential Revision: https://reviews.llvm.org/D108276
Diffstat (limited to 'llvm/lib/Target')
-rw-r--r-- | llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp | 21 |
1 files changed, 21 insertions, 0 deletions
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp index 5320f44..d68467b 100644 --- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp +++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp @@ -691,6 +691,27 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST) .clampMaxNumElements(1, s32, 4) .lower(); + getActionDefinitionsBuilder(G_VECREDUCE_OR) + // Try to break down into smaller vectors as long as they're at least 64 + // bits. This lets us use vector operations for some parts of the + // reduction. + .fewerElementsIf( + [=](const LegalityQuery &Q) { + LLT SrcTy = Q.Types[1]; + if (SrcTy.isScalar()) + return false; + if (!isPowerOf2_32(SrcTy.getNumElements())) + return false; + // We can usually perform 64b vector operations. + return SrcTy.getSizeInBits() > 64; + }, + [=](const LegalityQuery &Q) { + LLT SrcTy = Q.Types[1]; + return std::make_pair(1, SrcTy.divide(2)); + }) + .scalarize(1) + .lower(); + getActionDefinitionsBuilder({G_UADDSAT, G_USUBSAT}) .lowerIf([=](const LegalityQuery &Q) { return Q.Types[0].isScalar(); }); |