[llvm-branch-commits] [llvm] [SLP] Pass a lane as the vector cost context only if all lanes agree (PR #241758)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Thu Nov 12 02:25:41 PST 2026
================
----------------
Andarwinux wrote:
This looks like a regression.
https://godbolt.org/z/3cnha5G1h
For this specific example, I think best to vectorize it as 4 × udiv + mul + sub.
```llvm
define void @src(ptr noalias %a, ptr noalias %b, ptr noalias %r, ptr noalias %q) {
entry:
%0 = load <4 x i32>, ptr %a, align 4
%1 = load i32, ptr %a, align 4
%2 = load <4 x i32>, ptr %b, align 4
%3 = load i32, ptr %b, align 4
%4 = urem <4 x i32> %0, %2
store <4 x i32> %4, ptr %r, align 4
%div = udiv i32 %1, %3
store i32 %div, ptr %q, align 4
ret void
}
define void @tgt(ptr noalias %a, ptr noalias %b, ptr noalias %r, ptr noalias %q) {
entry:
%0 = load <4 x i32>, ptr %a, align 4
%1 = load <4 x i32>, ptr %b, align 4
%div = udiv <4 x i32> %0, %1
%mul = mul <4 x i32> %div, %1
%sub = sub <4 x i32> %0, %mul
store <4 x i32> %sub, ptr %r, align 4
%vecext22 = extractelement <4 x i32> %div, i64 0
store i32 %vecext22, ptr %q, align 4
ret void
}
```
```asm
src:
vmovdqu xmm0, xmmword ptr [rdi]
vmovdqu xmm1, xmmword ptr [rsi]
mov eax, dword ptr [rdi]
mov r8, rdx
xor edx, edx
div dword ptr [rsi]
vcvtudq2pd ymm2, xmm1
vcvtudq2pd ymm3, xmm0
vdivpd ymm2, ymm3, ymm2
mov dword ptr [rcx], eax
vcvttpd2udq xmm2, ymm2
vpmulld xmm1, xmm2, xmm1
vpsubd xmm0, xmm0, xmm1
vmovdqu xmmword ptr [r8], xmm0
vzeroupper
ret
tgt:
vmovdqu xmm0, xmmword ptr [rdi]
vmovdqu xmm1, xmmword ptr [rsi]
vcvtudq2pd ymm2, xmm1
vcvtudq2pd ymm3, xmm0
vdivpd ymm2, ymm3, ymm2
vcvttpd2udq xmm2, ymm2
vpmulld xmm1, xmm2, xmm1
vmovd dword ptr [rcx], xmm2
vpsubd xmm0, xmm0, xmm1
vmovdqu xmmword ptr [rdx], xmm0
vzeroupper
ret
```
https://github.com/llvm/llvm-project/pull/241758
More information about the llvm-branch-commits
mailing list