[llvm] [InstCombine] Split div/rem by a select of constants with a variable dividend (PR #216908)

[Adam Scott via llvm-commits]

================
@@ -518,5 +518,85 @@ define <2 x i64> @sdivrem_identity_const(<2 x i1> %c, <2 x i64> %x) {
   %r = add <2 x i64> %div, %rem
   ret <2 x i64> %r
 }
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; CHECK-X64: {{.*}}
+
+define i32 @sdiv_select_of_constants(i32 %x, i1 %c) {
+; CHECK-X64-LABEL: sdiv_select_of_constants:
+; CHECK-X64:       # %bb.0:
+; CHECK-X64-NEXT:    # kill: def $edi killed $edi def $rdi
+; CHECK-X64-NEXT:    leal 3(%rdi), %ecx
+; CHECK-X64-NEXT:    testl %edi, %edi
+; CHECK-X64-NEXT:    cmovnsl %edi, %ecx
+; CHECK-X64-NEXT:    sarl $2, %ecx
+; CHECK-X64-NEXT:    movslq %edi, %rax
+; CHECK-X64-NEXT:    imulq $1717986919, %rax, %rax # imm = 0x66666667
+; CHECK-X64-NEXT:    movq %rax, %rdx
+; CHECK-X64-NEXT:    shrq $63, %rdx
+; CHECK-X64-NEXT:    sarq $33, %rax
+; CHECK-X64-NEXT:    addl %edx, %eax
+; CHECK-X64-NEXT:    testb $1, %sil
+; CHECK-X64-NEXT:    cmovnel %ecx, %eax
+; CHECK-X64-NEXT:    # kill: def $eax killed $eax killed $rax
+; CHECK-X64-NEXT:    retq
+  %a = sdiv i32 %x, 4
+  %b = sdiv i32 %x, 5
+  %r = select i1 %c, i32 %a, i32 %b
+  ret i32 %r
+}
+
+define i32 @urem_select_of_constants(i32 %x, i1 %c) {
+; CHECK-X64-V3-LABEL: urem_select_of_constants:
+; CHECK-X64-V3:       # %bb.0:
+; CHECK-X64-V3-NEXT:    movl %edi, %eax
+; CHECK-X64-V3-NEXT:    movl %edi, %edx
+; CHECK-X64-V3-NEXT:    movl $2863311531, %ecx # imm = 0xAAAAAAAB
+; CHECK-X64-V3-NEXT:    imulq %rdx, %rcx
+; CHECK-X64-V3-NEXT:    shrq $34, %rcx
+; CHECK-X64-V3-NEXT:    addl %ecx, %ecx
+; CHECK-X64-V3-NEXT:    leal (%rcx,%rcx,2), %ecx
+; CHECK-X64-V3-NEXT:    movabsq $2635249153617166336, %rdi # imm = 0x24924924A0000000
+; CHECK-X64-V3-NEXT:    mulxq %rdi, %rdx, %rdx
+; CHECK-X64-V3-NEXT:    leal (,%rdx,8), %edi
+; CHECK-X64-V3-NEXT:    subl %edx, %edi
+; CHECK-X64-V3-NEXT:    testb $1, %sil
+; CHECK-X64-V3-NEXT:    cmovnel %ecx, %edi
+; CHECK-X64-V3-NEXT:    subl %edi, %eax
+; CHECK-X64-V3-NEXT:    retq
+;
+; CHECK-X64-V4-LABEL: urem_select_of_constants:
----------------
as4230 wrote:

With `uint16_t a / (12 + b)` it seems the branchless split that this PR does still wins by RT of 2.0 instead of 3.0 https://godbolt.org/z/d689b5c65 .
The already vector cases go the other way though as you already figured in the newer issue you made: https://github.com/llvm/llvm-project/issues/217613



https://github.com/llvm/llvm-project/pull/216908