[DAG] Failure to fold select(x, sub(x, c), m) -> sub(x, and(c,m))

[RKSimon]

https://godbolt.org/z/a1PczEM8a

If we're selecting a subtraction with a non-constant we fold the select into an and:

#include <x86intrin.h>
auto masked_select(__m128i a, __m128i b, __m128i x, __m128i y) {
    return _mm_blendv_epi8(a, _mm_sub_epi32(a, b), _mm_cmpgt_epi32(x,y));
}

masked_select(long long __vector(2), long long __vector(2), long long __vector(2), long long __vector(2)): # @masked_select(long long __vector(2), long long __vector(2), long long __vector(2), long long __vector(2))
  pcmpgtd %xmm3, %xmm2
  pand %xmm1, %xmm2
  psubd %xmm2, %xmm0
  retq

But for constants this fails, which on x86 can result in a BLENDV instruction, which is never faster than an AND

#include <x86intrin.h>
auto masked_select_const(__m128i a, __m128i x, __m128i y) {
    __m128i b = _mm_set1_epi32(24);
    return _mm_blendv_epi8(a, _mm_sub_epi32(a, b), _mm_cmpgt_epi32(x,y));
}

masked_select_const(long long __vector(2), long long __vector(2), long long __vector(2)): # @masked_select_const(long long __vector(2), long long __vector(2), long long __vector(2))
  movdqa %xmm0, %xmm3
  movdqa .LCPI3_0(%rip), %xmm4 # xmm4 = [4294967272,4294967272,4294967272,4294967272]
  paddd %xmm0, %xmm4
  pcmpgtd %xmm2, %xmm1
  movdqa %xmm1, %xmm0
  blendvps %xmm0, %xmm4, %xmm3
  movaps %xmm3, %xmm0
  retq

[llvmbot]

@llvm/issue-subscribers-backend-x86

https://godbolt.org/z/a1PczEM8a

If we're selecting a subtracting a non-constant we fold the select into a and:

#include <x86intrin.h>
auto masked_select(__m128i a, __m128i b, __m128i x, __m128i y) {
    return _mm_blendv_epi8(a, _mm_sub_epi32(a, b), _mm_cmpgt_epi32(x,y));
}

masked_select(long long __vector(2), long long __vector(2), long long __vector(2), long long __vector(2)): # @masked_select(long long __vector(2), long long __vector(2), long long __vector(2), long long __vector(2))
  pcmpgtd %xmm3, %xmm2
  pand %xmm1, %xmm2
  psubd %xmm2, %xmm0
  retq

But for constants this fails, which on x86 can result in a BLENDV instruction, which is never faster than a AND

#include <x86intrin.h>
auto masked_select_const(__m128i a, __m128i x, __m128i y) {
    __m128i b = _mm_set1_epi32(24);
    return _mm_blendv_epi8(a, _mm_sub_epi32(a, b), _mm_cmpgt_epi32(x,y));
}

masked_select_const(long long __vector(2), long long __vector(2), long long __vector(2)): # @masked_select_const(long long __vector(2), long long __vector(2), long long __vector(2))
  movdqa %xmm0, %xmm3
  movdqa .LCPI3_0(%rip), %xmm4 # xmm4 = [4294967272,4294967272,4294967272,4294967272]
  paddd %xmm0, %xmm4
  pcmpgtd %xmm2, %xmm1
  movdqa %xmm1, %xmm0
  blendvps %xmm0, %xmm4, %xmm3
  movaps %xmm3, %xmm0
  retq

[RKSimon]

CC @elhewaty

[elhewaty]

assign me, please.

[elhewaty]

@RKSimon Is there any source I can use to understand DAG internals.

[RKSimon]

I'd start by seeing whats the difference between the IR being fed to DAG from masked_select vs masked_select_const - you will probably need to remove a lot of unnecessary bitcasts. Then step through the DAGCombine stages of running llc in a debugger - add breakpoints to the start of visitADD/visitSUB/visitVSELECT and see whats happening.

You can also use "llc --debug" (using a debug assertion build) to dump out everything llc has done: https://rust.godbolt.org/z/szYv5G8n9

[elhewaty]

Hello @RKSimon.

// select X, sub(X, C), m --> sub (X, and(C, m))
  if (N1.getOpcode() == ISD::SUB && N1.getOperand(0) == N0 && N1.hasOneUse()) {
    if (dyn_cast<ConstantSDNode>(N1.getOperand(1)))
      return DAG.getNode(ISD::SUB, DL, N0.getValueType(), N0,
                         DAG.getNode(ISD::AND, DL, N2.getValueType(),
                                     N1.getOperand(1), N2));
  }

Here's what reached so far, I tried to match a pattern in visitSELECT function.
is this logic correct?

[RKSimon]

Yes, that looks about right - you should use isConstantIntBuildVectorOrConstantInt instead of dyn_cast<ConstantSDNode> so it can match vector constant as well

[RKSimon]

Also, you need to sort out argument order (sorry when I reported this I was thinking _mm_blendv_epi8 order not select IR order)