[PATCH] D63391: [CodeGen] [SelectionDAG] More efficient code for X % C == 0 (UREM case) (try 2)

[Roman Lebedev via Phabricator via llvm-commits]

lebedev.ri added a comment.

So the first thought about `multiplicativeInverse()` being broken does **not** appear to be correct
(although that interface is extremely susceptible to misuse),

The miscompiles in test-suite happened for `urem` by `100` (and `10`, did not check for more cases).
This code produces:

  define i32 @test_urem_100(i32 %X) nounwind readnone {
  ; X86-LABEL: test_urem_100:
  ; X86:       # %bb.0:
  ; X86-NEXT:    imull $-1030792151, {{[0-9]+}}(%esp), %ecx # imm = 0xC28F5C29
  ; X86-NEXT:    rorl $2, %ecx
  ; X86-NEXT:    xorl %eax, %eax
  ; X86-NEXT:    cmpl $171798692, %ecx # imm = 0xA3D70A4
  ; X86-NEXT:    setb %al
  ; X86-NEXT:    retl
  ;
  ; X64-LABEL: test_urem_100:
  ; X64:       # %bb.0:
  ; X64-NEXT:    imull $-1030792151, %edi, %ecx # imm = 0xC28F5C29
  ; X64-NEXT:    rorl $2, %ecx
  ; X64-NEXT:    xorl %eax, %eax
  ; X64-NEXT:    cmpl $171798692, %ecx # imm = 0xA3D70A4
  ; X64-NEXT:    setb %al
  ; X64-NEXT:    retq
    %urem = urem i32 %X, 100
    %cmp = icmp eq i32 %urem, 0
    %ret = zext i1 %cmp to i32
    ret i32 %ret
  }

GCC: https://godbolt.org/z/ks6Wm-

  imull   $-1030792151, %edi, %edi
  rorl    $2, %edi
  cmpl    $42949672, %edi
  setbe   %al

The difference is in the constant we compare with: `171798692` (patch) vs `42949672` (GCC).


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D63391/new/

https://reviews.llvm.org/D63391