Issue 241788
Summary [SLP] suboptimal divrem vectorization
Labels llvm:SLPVectorizer, missed-optimization
Assignees
Reporter Andarwinux
    noticed from #241758
https://godbolt.org/z/8ofcT8xWE
```llvm
define void @urem_div_lane0(ptr %a, ptr %b, ptr %r, ptr %q) {
entry:
  %a0 = load i32, ptr %a, align 4
  %pa1 = getelementptr inbounds i32, ptr %a, i64 1
  %a1 = load i32, ptr %pa1, align 4
  %pa2 = getelementptr inbounds i32, ptr %a, i64 2
  %a2 = load i32, ptr %pa2, align 4
  %pa3 = getelementptr inbounds i32, ptr %a, i64 3
  %a3 = load i32, ptr %pa3, align 4
  %b0 = load i32, ptr %b, align 4
  %pb1 = getelementptr inbounds i32, ptr %b, i64 1
  %b1 = load i32, ptr %pb1, align 4
  %pb2 = getelementptr inbounds i32, ptr %b, i64 2
 %b2 = load i32, ptr %pb2, align 4
  %pb3 = getelementptr inbounds i32, ptr %b, i64 3
  %b3 = load i32, ptr %pb3, align 4
  %r0 = urem i32 %a0, %b0
 %r1 = urem i32 %a1, %b1
  %r2 = urem i32 %a2, %b2
  %r3 = urem i32 %a3, %b3
  %d0 = udiv i32 %a0, %b0
  store i32 %d0, ptr %q, align 4
  store i32 %r0, ptr %r, align 4
  %pr1 = getelementptr inbounds i32, ptr %r, i64 1
 store i32 %r1, ptr %pr1, align 4
  %pr2 = getelementptr inbounds i32, ptr %r, i64 2
  store i32 %r2, ptr %pr2, align 4
  %pr3 = getelementptr inbounds i32, ptr %r, i64 3
  store i32 %r3, ptr %pr3, align 4
  ret void
}
```
```llvm
define void @urem_div_lane0(ptr nofree readonly captures(none) %a, ptr nofree readonly captures(none) %b, ptr nofree writeonly captures(none) initializes((0, 16)) %r, ptr nofree writeonly captures(none) initializes((0, 4)) %q) local_unnamed_addr {
entry:
  %0 = load <4 x i32>, ptr %a, align 4
  %a0 = load i32, ptr %a, align 4
  %1 = load <4 x i32>, ptr %b, align 4
  %b0 = load i32, ptr %b, align 4
  %2 = urem <4 x i32> %0, %1
  %d0 = udiv i32 %a0, %b0
  store i32 %d0, ptr %q, align 4
  store <4 x i32> %2, ptr %r, align 4
  ret void
}
```
```asm
urem_div_lane0:
        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
```
SLP should rewrite rem vectorization candidates with paired divs to div+mul+sub like this one:
```llvm
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
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
```
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to