[DAGCombiner] Allow divide by constant optimization on opaque constants.

[craig.topper]

Summary

I believe this restores the behavior we had before r339147.

Fixes PR38622.

[chandlerc]

Nice! I like it when the fix is simple.

test/CodeGen/X86/divide-by-constant.ll

334 ↗

(On Diff #161369)

If you add nounwind to this, the asm will get a lot less verbose while still showing the relevant bits.

[chandlerc]

This time actually marking it as LGTM.

[xbolva00]

test/CodeGen/X86/divide-by-constant.ll

404 ↗

(On Diff #161369)

another test for sdiv?

[RKSimon]

Thanks @craig.topper !

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (!N->getFlags().hasExact() && ISD::matchUnaryPredicate(N1, IsPowerOfTwo)) {
     SDValue Res = DAG.getSelect(DL, VT, IsNeg, Sub, Sra);
     return Res;
   }
 
   // If integer divide is expensive and we satisfy the requirements, emit an
   // alternate sequence.  Targets may check function attributes for size/speed
   // trade-offs.
   AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
-  if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) &&
+  if (isConstantOrConstantVector(N1) &&
       !TLI.isIntDivCheap(N->getValueType(0), Attr))
     if (SDValue Op = BuildSDIV(N))
       return Op;
 
   return SDValue();
 }
 
 SDValue DAGCombiner::visitUDIV(SDNode *N) {
if (isConstantOrConstantVector(N10, /*NoOpaques*/ true) &&
       SDValue Add = DAG.getNode(ISD::ADD, DL, ADDVT, N1.getOperand(1), Trunc);
       AddToWorklist(Add.getNode());
       return DAG.getNode(ISD::SRL, DL, VT, N0, Add);
     }
   }
 
   // fold (udiv x, c) -> alternate
   AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
-  if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) &&
+  if (isConstantOrConstantVector(N1) &&
       !TLI.isIntDivCheap(N->getValueType(0), Attr))
     if (SDValue Op = BuildUDIV(N))
       return Op;
 
   return SDValue();
 }
 
 // handles ISD::SREM and ISD::UREM

llvm/trunk/test/CodeGen/X86/divide-by-constant.ll

 ; X64-NEXT:    leaq (%rcx,%rdx), %rax
 ; X64-NEXT:    shrq $2, %rax
 ; X64-NEXT:    retq
 entry:
   %rem = urem i64 %x, 12345
   %div = udiv i64 %rem, 7
   ret i64 %div
 }
+
+define { i64, i32 } @PR38622(i64) {
+; X32-LABEL: PR38622:
+; X32:       # %bb.0:
+; X32-NEXT:    pushl %ebp
+; X32-NEXT:    .cfi_def_cfa_offset 8
+; X32-NEXT:    pushl %ebx
+; X32-NEXT:    .cfi_def_cfa_offset 12
+; X32-NEXT:    pushl %edi
+; X32-NEXT:    .cfi_def_cfa_offset 16
+; X32-NEXT:    pushl %esi
+; X32-NEXT:    .cfi_def_cfa_offset 20
+; X32-NEXT:    subl $12, %esp
+; X32-NEXT:    .cfi_def_cfa_offset 32
+; X32-NEXT:    .cfi_offset %esi, -20
+; X32-NEXT:    .cfi_offset %edi, -16
+; X32-NEXT:    .cfi_offset %ebx, -12
+; X32-NEXT:    .cfi_offset %ebp, -8
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %ebx
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %ebp
+; X32-NEXT:    pushl $0
+; X32-NEXT:    .cfi_adjust_cfa_offset 4
+; X32-NEXT:    pushl $-294967296 # imm = 0xEE6B2800
+; X32-NEXT:    .cfi_adjust_cfa_offset 4
+; X32-NEXT:    pushl %ebp
+; X32-NEXT:    .cfi_adjust_cfa_offset 4
+; X32-NEXT:    pushl %ebx
+; X32-NEXT:    .cfi_adjust_cfa_offset 4
+; X32-NEXT:    calll __udivdi3
+; X32-NEXT:    addl $16, %esp
+; X32-NEXT:    .cfi_adjust_cfa_offset -16
+; X32-NEXT:    movl %eax, %esi
+; X32-NEXT:    movl %edx, %edi
+; X32-NEXT:    pushl $0
+; X32-NEXT:    .cfi_adjust_cfa_offset 4
+; X32-NEXT:    pushl $-294967296 # imm = 0xEE6B2800
+; X32-NEXT:    .cfi_adjust_cfa_offset 4
+; X32-NEXT:    pushl %ebp
+; X32-NEXT:    .cfi_adjust_cfa_offset 4
+; X32-NEXT:    pushl %ebx
+; X32-NEXT:    .cfi_adjust_cfa_offset 4
+; X32-NEXT:    calll __umoddi3
+; X32-NEXT:    addl $16, %esp
+; X32-NEXT:    .cfi_adjust_cfa_offset -16
+; X32-NEXT:    movl %eax, %ecx
+; X32-NEXT:    movl %esi, %eax
+; X32-NEXT:    movl %edi, %edx
+; X32-NEXT:    addl $12, %esp
+; X32-NEXT:    .cfi_def_cfa_offset 20
+; X32-NEXT:    popl %esi
+; X32-NEXT:    .cfi_def_cfa_offset 16
+; X32-NEXT:    popl %edi
+; X32-NEXT:    .cfi_def_cfa_offset 12
+; X32-NEXT:    popl %ebx
+; X32-NEXT:    .cfi_def_cfa_offset 8
+; X32-NEXT:    popl %ebp
+; X32-NEXT:    .cfi_def_cfa_offset 4
+; X32-NEXT:    retl
+;
+; X64-LABEL: PR38622:
+; X64:       # %bb.0:
+; X64-NEXT:    movq %rdi, %rax
+; X64-NEXT:    shrq $11, %rax
+; X64-NEXT:    movabsq $4835703278458517, %rcx # imm = 0x112E0BE826D695
+; X64-NEXT:    mulq %rcx
+; X64-NEXT:    shrq $9, %rdx
+; X64-NEXT:    imull $-294967296, %edx, %eax # imm = 0xEE6B2800
+; X64-NEXT:    subl %eax, %edi
+; X64-NEXT:    movq %rdx, %rax
+; X64-NEXT:    movl %edi, %edx
+; X64-NEXT:    retq
+  %2 = udiv i64 %0, 4000000000
+  %3 = urem i64 %0, 4000000000
+  %4 = trunc i64 %3 to i32
+  %5 = insertvalue { i64, i32 } undef, i64 %2, 0
+  %6 = insertvalue { i64, i32 } %5, i32 %4, 1
+  ret { i64, i32 } %6
+}