[SDAG][X86] Expand pow2 mulo using shifts

[nikic]

Summary

Expand MULO with constant power of two operand to a shift. The overflow is checked with (x << shift) >> shift == x.

[nikic]

llvm/test/CodeGen/X86/mulo-pow2.ll

30

This being optimized is a nice side-effect, but we really should have a fold for mulo(x, 1).

109

This code is generated by the mulo(x, 2) -> addo(x, x) fold and seems worse than what we'd get from the shift expansion. Not sure what to do about it.

[nikic]

llvm/test/CodeGen/X86/mulo-pow2.ll

133

This one isn't right ... INT_MIN can be multiplied by 0 and 1, not by 0 and -1.

[nikic]

Ensure the optimization is not applied to signed min.

[nikic]

Add changes to an ARM test.

[nikic]

llvm/test/CodeGen/ARM/umulo-32.ll

41 ↗

(On Diff #189987)

We're doing a select of setcc here, both get expanded and the setcc is converted into this construction.

[efriedma]

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

5529

(!isSigned || C.isNonNegative()) is specifically supposed to handle llvm.smul.with.overflow.i32(%x, INT_MIN)? Can't you just convert that to llvm.umul.with.overflow.i32(%x, INT_MIN) and lower it using an unsigned shift?

[nikic]

Handle smulo with signed_min like umulo via logical shift.

[nikic]

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

5529

Yes, you're right. I've implemented this now.

[efriedma]

LGTM

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

TargetLowering::expandFixedPointMul(SDNode *Node, SelectionDAG &DAG) const {
   return DAG.getNode(ISD::FSHR, dl, VT, Hi, Lo,
                      DAG.getConstant(Scale, dl, ShiftTy));
 }
 
 bool TargetLowering::expandMULO(SDNode *Node, SDValue &Result,
                                 SDValue &Overflow, SelectionDAG &DAG) const {
   SDLoc dl(Node);
   EVT VT = Node->getValueType(0);
+  EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
+  SDValue LHS = Node->getOperand(0);
+  SDValue RHS = Node->getOperand(1);
+  bool isSigned = Node->getOpcode() == ISD::SMULO;
+
+  // For power-of-two multiplications we can use a simpler shift expansion.
+  if (ConstantSDNode *RHSC = isConstOrConstSplat(RHS)) {
+    const APInt &C = RHSC->getAPIntValue();
+    // mulo(X, 1 << S) -> { X << S, (X << S) >> S != X }
+    if (C.isPowerOf2() && (!isSigned || C.isNonNegative())) {
+      EVT ShiftAmtTy = getShiftAmountTy(VT, DAG.getDataLayout());
+      SDValue ShiftAmt = DAG.getConstant(C.logBase2(), dl, ShiftAmtTy);
+      Result = DAG.getNode(ISD::SHL, dl, VT, LHS, ShiftAmt);
+      Overflow = DAG.getSetCC(dl, SetCCVT,
+          DAG.getNode(isSigned ? ISD::SRA : ISD::SRL, dl, VT, Result, ShiftAmt),
+          LHS, ISD::SETNE);
+      return true;
+    }
+  }
+
   EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getScalarSizeInBits() * 2);
   if (VT.isVector())
     WideVT = EVT::getVectorVT(*DAG.getContext(), WideVT,
                               VT.getVectorNumElements());
 
-  SDValue LHS = Node->getOperand(0);
-  SDValue RHS = Node->getOperand(1);
   SDValue BottomHalf;
   SDValue TopHalf;
   static const unsigned Ops[2][3] =
       { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
         { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
-  bool isSigned = Node->getOpcode() == ISD::SMULO;
   if (isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
     BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
     TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
   } else if (isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
     BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
                              RHS);
     TopHalf = BottomHalf.getValue(1);
   } else if (isTypeLegal(WideVT)) {
if (DAG.getDataLayout().isLittleEndian()) {
       BottomHalf = Ret.getOperand(0);
       TopHalf = Ret.getOperand(1);
     } else {
       BottomHalf = Ret.getOperand(1);
       TopHalf = Ret.getOperand(0);
     }
   }
 
-  EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
   Result = BottomHalf;
   if (isSigned) {
     SDValue ShiftAmt = DAG.getConstant(
         VT.getScalarSizeInBits() - 1, dl,
         getShiftAmountTy(BottomHalf.getValueType(), DAG.getDataLayout()));
     SDValue Sign = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, ShiftAmt);
     Overflow = DAG.getSetCC(dl, SetCCVT, TopHalf, Sign, ISD::SETNE);
   } else {

llvm/test/CodeGen/X86/mulo-pow2.ll

; AVX-NEXT:    retq
     %z = extractvalue { <4 x i32>, <4 x i1> } %x, 1
     %u = select <4 x i1> %z, <4 x i32> %b, <4 x i32> %y
     ret <4 x i32> %u
 }
 
 define <4 x i32> @umul_v4i32_1(<4 x i32> %a, <4 x i32> %b) nounwind {
 ; AVX-LABEL: umul_v4i32_1:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; AVX-NEXT:    vmovdqa {{.*#+}} xmm3 = [1,1,1,1]
-; AVX-NEXT:    vpmuludq %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vpmuludq %xmm3, %xmm0, %xmm3
-; AVX-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
-; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5],xmm2[6,7]
-; AVX-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX-NEXT:    vpcmpeqd %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
     %x = call { <4 x i32>, <4 x i1> } @llvm.umul.with.overflow.v4i32(<4 x i32> %a, <4 x i32> <i32 1, i32 1, i32 1, i32 1>)
     %y = extractvalue { <4 x i32>, <4 x i1> } %x, 0
     %z = extractvalue { <4 x i32>, <4 x i1> } %x, 1
     %u = select <4 x i1> %z, <4 x i32> %b, <4 x i32> %y
     ret <4 x i32> %u
 }
 
; AVX-NEXT:    retq
     %z = extractvalue { <4 x i32>, <4 x i1> } %x, 1
     %u = select <4 x i1> %z, <4 x i32> %b, <4 x i32> %y
     ret <4 x i32> %u
 }
 
 define <4 x i32> @umul_v4i32_8(<4 x i32> %a, <4 x i32> %b) nounwind {
 ; AVX-LABEL: umul_v4i32_8:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; AVX-NEXT:    vmovdqa {{.*#+}} xmm3 = [8,8,8,8]
-; AVX-NEXT:    vpmuludq %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vpmuludq %xmm3, %xmm0, %xmm3
-; AVX-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
-; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5],xmm2[6,7]
-; AVX-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX-NEXT:    vpcmpeqd %xmm3, %xmm2, %xmm2
+; AVX-NEXT:    vpand {{.*}}(%rip), %xmm0, %xmm2
+; AVX-NEXT:    vpcmpeqd %xmm0, %xmm2, %xmm2
 ; AVX-NEXT:    vpslld $3, %xmm0, %xmm0
 ; AVX-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
     %x = call { <4 x i32>, <4 x i1> } @llvm.umul.with.overflow.v4i32(<4 x i32> %a, <4 x i32> <i32 8, i32 8, i32 8, i32 8>)
     %y = extractvalue { <4 x i32>, <4 x i1> } %x, 0
     %z = extractvalue { <4 x i32>, <4 x i1> } %x, 1
     %u = select <4 x i1> %z, <4 x i32> %b, <4 x i32> %y
     ret <4 x i32> %u
 }
 
 define <4 x i32> @umul_v4i32_2pow31(<4 x i32> %a, <4 x i32> %b) nounwind {
 ; AVX-LABEL: umul_v4i32_2pow31:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; AVX-NEXT:    vmovdqa {{.*#+}} xmm3 = [2147483648,2147483648,2147483648,2147483648]
-; AVX-NEXT:    vpmuludq %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vpmuludq %xmm3, %xmm0, %xmm3
-; AVX-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
-; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5],xmm2[6,7]
-; AVX-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; AVX-NEXT:    vpcmpeqd %xmm3, %xmm2, %xmm2
+; AVX-NEXT:    vpand {{.*}}(%rip), %xmm0, %xmm2
+; AVX-NEXT:    vpcmpeqd %xmm0, %xmm2, %xmm2
 ; AVX-NEXT:    vpslld $31, %xmm0, %xmm0
 ; AVX-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
     %x = call { <4 x i32>, <4 x i1> } @llvm.umul.with.overflow.v4i32(<4 x i32> %a, <4 x i32> <i32 2147483648, i32 2147483648, i32 2147483648, i32 2147483648>)
     %y = extractvalue { <4 x i32>, <4 x i1> } %x, 0
     %z = extractvalue { <4 x i32>, <4 x i1> } %x, 1
     %u = select <4 x i1> %z, <4 x i32> %b, <4 x i32> %y
     ret <4 x i32> %u
; AVX-NEXT:    retq
     %z = extractvalue { <4 x i32>, <4 x i1> } %x, 1
     %u = select <4 x i1> %z, <4 x i32> %b, <4 x i32> %y
     ret <4 x i32> %u
 }
 
 define <4 x i32> @smul_v4i32_1(<4 x i32> %a, <4 x i32> %b) nounwind {
 ; AVX-LABEL: smul_v4i32_1:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; AVX-NEXT:    vmovdqa {{.*#+}} xmm3 = [1,1,1,1]
-; AVX-NEXT:    vpmuldq %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vpmuldq %xmm3, %xmm0, %xmm3
-; AVX-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
-; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5],xmm2[6,7]
-; AVX-NEXT:    vpsrad $31, %xmm0, %xmm3
-; AVX-NEXT:    vpcmpeqd %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
     %x = call { <4 x i32>, <4 x i1> } @llvm.smul.with.overflow.v4i32(<4 x i32> %a, <4 x i32> <i32 1, i32 1, i32 1, i32 1>)
     %y = extractvalue { <4 x i32>, <4 x i1> } %x, 0
     %z = extractvalue { <4 x i32>, <4 x i1> } %x, 1
     %u = select <4 x i1> %z, <4 x i32> %b, <4 x i32> %y
     ret <4 x i32> %u
 }
 
 define <4 x i32> @smul_v4i32_2(<4 x i32> %a, <4 x i32> %b) nounwind {
 ; AVX-LABEL: smul_v4i32_2:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; AVX-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm3
 ; AVX-NEXT:    vpcmpeqd %xmm4, %xmm4, %xmm4
 ; AVX-NEXT:    vpxor %xmm4, %xmm3, %xmm3
 ; AVX-NEXT:    vpaddd %xmm0, %xmm0, %xmm0
 ; AVX-NEXT:    vpcmpgtd %xmm0, %xmm2, %xmm2
 ; AVX-NEXT:    vpxor %xmm4, %xmm2, %xmm2
 ; AVX-NEXT:    vpcmpeqd %xmm2, %xmm3, %xmm2
 ; AVX-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
     %x = call { <4 x i32>, <4 x i1> } @llvm.smul.with.overflow.v4i32(<4 x i32> %a, <4 x i32> <i32 2, i32 2, i32 2, i32 2>)
     %y = extractvalue { <4 x i32>, <4 x i1> } %x, 0
     %z = extractvalue { <4 x i32>, <4 x i1> } %x, 1
     %u = select <4 x i1> %z, <4 x i32> %b, <4 x i32> %y
     ret <4 x i32> %u
 }
 
 define <4 x i32> @smul_v4i32_8(<4 x i32> %a, <4 x i32> %b) nounwind {
 ; AVX-LABEL: smul_v4i32_8:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; AVX-NEXT:    vmovdqa {{.*#+}} xmm3 = [8,8,8,8]
-; AVX-NEXT:    vpmuldq %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vpmuldq %xmm3, %xmm0, %xmm3
-; AVX-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
-; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5],xmm2[6,7]
-; AVX-NEXT:    vpslld $3, %xmm0, %xmm0
-; AVX-NEXT:    vpsrad $31, %xmm0, %xmm3
-; AVX-NEXT:    vpcmpeqd %xmm3, %xmm2, %xmm2
-; AVX-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; AVX-NEXT:    vpslld $3, %xmm0, %xmm2
+; AVX-NEXT:    vpsrad $3, %xmm2, %xmm3
+; AVX-NEXT:    vpcmpeqd %xmm0, %xmm3, %xmm0
+; AVX-NEXT:    vblendvps %xmm0, %xmm2, %xmm1, %xmm0
 ; AVX-NEXT:    retq
     %x = call { <4 x i32>, <4 x i1> } @llvm.smul.with.overflow.v4i32(<4 x i32> %a, <4 x i32> <i32 8, i32 8, i32 8, i32 8>)
     %y = extractvalue { <4 x i32>, <4 x i1> } %x, 0
     %z = extractvalue { <4 x i32>, <4 x i1> } %x, 1
     %u = select <4 x i1> %z, <4 x i32> %b, <4 x i32> %y
     ret <4 x i32> %u
 }
 
 define <4 x i32> @smul_v4i32_2pow31(<4 x i32> %a, <4 x i32> %b) nounwind {
 ; AVX-LABEL: smul_v4i32_2pow31:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vpshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
 ; AVX-NEXT:    vmovdqa {{.*#+}} xmm3 = [2147483648,2147483648,2147483648,2147483648]
 ; AVX-NEXT:    vpmuldq %xmm3, %xmm2, %xmm2
 ; AVX-NEXT:    vpmuldq %xmm3, %xmm0, %xmm3
 ; AVX-NEXT:    vpshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
 ; AVX-NEXT:    vpblendw {{.*#+}} xmm2 = xmm3[0,1],xmm2[2,3],xmm3[4,5],xmm2[6,7]