[PATCH] D147996: [X86] combineConcatVectorOps - remove FADD/FSUB/FMUL handling (2-1)
Xiang Zhang via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Tue Apr 11 00:35:01 PDT 2023
xiangzhangllvm created this revision.
xiangzhangllvm added reviewers: RKSimon, LuoYuanke, pengfei.
Herald added a subscriber: hiraditya.
Herald added a project: All.
xiangzhangllvm requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.
Due to VADD, VSUB, VMUL can executed on more ports than VINSERT.
We tend to remove FADD/FSUB/FMUL handling from combineConcatVectorOps.
More details pls refer to fb91f0 <https://reviews.llvm.org/rGfb91f0a298492ed7db8928e18cda35228474007e>
**Problems**
**1** This patch cause affected tests generated worse code. We need to fix it.
PS: After [[ https://reviews.llvm.org/rGfb91f0a298492ed7db8928e18cda35228474007e | fb91f0 ]] there is another [[ https://reviews.llvm.org/rG649b14928a67e016f3e01ac46499aaf1824c2d09 | 649b14 ]] optimize these tests
**2** We may also need to root cause why middle end vector optimization didn't work
for there cases.
For example, llvm/test/CodeGen/X86/widen_fadd.ll function widen_fadd_v2f32_v8f32
was not be optimized (in avx mode)
**from**
define void @widen_fadd_v2f32_v8f32(ptr %a0, ptr %b0, ptr %c0) {
%a2 = getelementptr inbounds i8, ptr %a0, i64 8
%b2 = getelementptr inbounds i8, ptr %b0, i64 8
%c2 = getelementptr inbounds i8, ptr %c0, i64 8
%a4 = getelementptr inbounds i8, ptr %a0, i64 16
%b4 = getelementptr inbounds i8, ptr %b0, i64 16
%c4 = getelementptr inbounds i8, ptr %c0, i64 16
%a6 = getelementptr inbounds i8, ptr %a0, i64 24
%b6 = getelementptr inbounds i8, ptr %b0, i64 24
%c6 = getelementptr inbounds i8, ptr %c0, i64 24
%va0 = load <2 x float>, ptr %a0, align 4
%vb0 = load <2 x float>, ptr %b0, align 4
%va2 = load <2 x float>, ptr %a2, align 4
%vb2 = load <2 x float>, ptr %b2, align 4
%va4 = load <2 x float>, ptr %a4, align 4
%vb4 = load <2 x float>, ptr %b4, align 4
%va6 = load <2 x float>, ptr %a6, align 4
%vb6 = load <2 x float>, ptr %b6, align 4
%vc0 = fadd <2 x float> %va0, %vb0
%vc2 = fadd <2 x float> %va2, %vb2
%vc4 = fadd <2 x float> %va4, %vb4
%vc6 = fadd <2 x float> %va6, %vb6
store <2 x float> %vc0, ptr %c0, align 4
store <2 x float> %vc2, ptr %c2, align 4
store <2 x float> %vc4, ptr %c4, align 4
store <2 x float> %vc6, ptr %c6, align 4
ret void
}
**to**
define void @widen_fadd_v2f32_v8f32(ptr %a0, ptr %b0, ptr %c0) {
%va0 = load <8 x float>, ptr %a0, align 4
%vb0 = load <8 x float>, ptr %b0, align 4
%vc0 = fadd <8 x float> %va0, %vb0
store <8 x float> %vc0, ptr %c0, align 4
ret void
}
https://reviews.llvm.org/D147996
Files:
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/widen_fadd.ll
llvm/test/CodeGen/X86/widen_fmul.ll
llvm/test/CodeGen/X86/widen_fsub.ll
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D147996.512351.patch
Type: text/x-patch
Size: 17868 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230411/d62a3bca/attachment.bin>
More information about the llvm-commits
mailing list