[llvm-bugs] [Bug 46599] New: Commutable subtractions

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=46599

            Bug ID: 46599
           Summary: Commutable subtractions
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Common Code Generator Code
          Assignee: unassignedbugs at nondot.org
          Reporter: llvm-dev at redking.me.uk
                CC: craig.topper at gmail.com, efriedma at quicinc.com,
                    lebedev.ri at gmail.com, llvm-bugs at lists.llvm.org,
                    spatel+llvm at rotateright.com

If a subtraction result is only ever used for certain instructions (f)abs,
(f)mul etc. we might be able to tag the subtraction as "commutable".

This is particularly interesting for SSE memory folding (either direct or via
folding stack operations):
https://godbolt.org/z/zkELgP

#include <x86intrin.h>

auto dp_offset(__m128 *x, __m128 *y, __m128 z, __m128 w) {
    __m128 xx = _mm_sub_ps(*x, z);
    __m128 yy = _mm_sub_ps(*y, w);
    return _mm_add_ps(
        _mm_mul_ps(xx, xx),
        _mm_mul_ps(yy, yy)
    );
}
auto dp_offset_commute(__m128 *x, __m128 *y, __m128 z, __m128 w) {
    __m128 xx = _mm_sub_ps(z, *x);
    __m128 yy = _mm_sub_ps(w, *y);
    return _mm_add_ps(
        _mm_mul_ps(xx, xx),
        _mm_mul_ps(yy, yy)
    );
}

clang -g0 -O3 -march=btver2

dp_offset:
  vmovaps (%rdi), %xmm2
  vmovaps (%rsi), %xmm3
  vsubps %xmm0, %xmm2, %xmm0
  vsubps %xmm1, %xmm3, %xmm1
  vmulps %xmm0, %xmm0, %xmm0
  vmulps %xmm1, %xmm1, %xmm1
  vaddps %xmm1, %xmm0, %xmm0
  retq
dp_offset_commute:
  vsubps (%rdi), %xmm0, %xmm0
  vsubps (%rsi), %xmm1, %xmm1
  vmulps %xmm0, %xmm0, %xmm0
  vmulps %xmm1, %xmm1, %xmm1
  vaddps %xmm1, %xmm0, %xmm0
  retq

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20200706/afb7d622/attachment.html>