[DAG] FoldConstantArithmetic - fold intop(bitcast(buildvector(c1)),bitcast(buildvector(c1))) -> bitcast(intop(buildvector(c1'),buildvector(c2')))

[RKSimon]

Summary

Enable FoldConstantArithmetic to constant fold bitcasted constant build vectors. These have typically been bitcasted for type legalization purposes.

By extracting the raw constant bit data, performing the constant fold, and then casting the constant bit data back to the (legalized) type, we can perform constant folding on integer types after legalization.

This in particular helps 32-bit targets which need to handle vXi64 build vectors - during legalization the (unsupported) i64 elements are split to create a bitcasted v2Xi32 build vector.

Addresses some regressions in D113192.

[lebedev.ri]

Looks good to me, but please wait for one more review.

[pengfei]

This tests look good.

llvm/test/CodeGen/X86/packss.ll

200–203

Nit, I think this value is not necessary now.

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

if (NumOps == 2) {
     }
 
     // fold (add Sym, c) -> Sym+c
     if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Ops[0]))
       return FoldSymbolOffset(Opcode, VT, GA, Ops[1].getNode());
     if (TLI->isCommutativeBinOp(Opcode))
       if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Ops[1]))
         return FoldSymbolOffset(Opcode, VT, GA, Ops[0].getNode());
-
-    // If this is a bitwise logic opcode see if we can fold bitcasted ops.
-    // TODO: Can we generalize this and fold any bitcasted constant data?
-    if (ISD::isBitwiseLogicOp(Opcode) && Ops[0].getOpcode() == ISD::BITCAST &&
-        Ops[1].getOpcode() == ISD::BITCAST) {
-      SDValue InnerN1 = peekThroughBitcasts(Ops[0].getOperand(0));
-      SDValue InnerN2 = peekThroughBitcasts(Ops[1].getOperand(0));
-      EVT InnerVT = InnerN1.getValueType();
-      if (InnerVT == InnerN2.getValueType() && InnerVT.isInteger())
-        if (SDValue C =
-                FoldConstantArithmetic(Opcode, DL, InnerVT, {InnerN1, InnerN2}))
-          return getBitcast(VT, C);
-    }
   }
 
   // This is for vector folding only from here on.
   if (!VT.isVector())
     return SDValue();
 
   ElementCount NumElts = VT.getVectorElementCount();
 
+  // See if we can fold through bitcasted integer ops.
+  // TODO: Can we handle undef elements?
+  if (NumOps == 2 && VT.isFixedLengthVector() && VT.isInteger() &&
+      Ops[0].getValueType() == VT && Ops[1].getValueType() == VT &&
+      Ops[0].getOpcode() == ISD::BITCAST &&
+      Ops[1].getOpcode() == ISD::BITCAST) {
+    SDValue N1 = peekThroughBitcasts(Ops[0]);
+    SDValue N2 = peekThroughBitcasts(Ops[1]);
+    auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
+    auto *BV2 = dyn_cast<BuildVectorSDNode>(N2);
+    EVT BVVT = N1.getValueType();
+    if (BV1 && BV2 && BVVT.isInteger() && BVVT == N2.getValueType()) {
+      bool IsLE = getDataLayout().isLittleEndian();
+      unsigned EltBits = VT.getScalarSizeInBits();
+      SmallVector<APInt> RawBits1, RawBits2;
+      BitVector UndefElts1, UndefElts2;
+      if (BV1->getConstantRawBits(IsLE, EltBits, RawBits1, UndefElts1) &&
+          BV2->getConstantRawBits(IsLE, EltBits, RawBits2, UndefElts2) &&
+          UndefElts1.none() && UndefElts2.none()) {
+        SmallVector<APInt> RawBits;
+        for (unsigned I = 0, E = NumElts.getFixedValue(); I != E; ++I) {
+          Optional<APInt> Fold = FoldValue(Opcode, RawBits1[I], RawBits2[I]);
+          if (!Fold)
+            break;
+          RawBits.push_back(Fold.getValue());
+        }
+        if (RawBits.size() == NumElts.getFixedValue()) {
+          // We have constant folded, but we need to cast this again back to
+          // the original (possibly legalized) type.
+          SmallVector<APInt> DstBits;
+          BitVector DstUndefs;
+          BuildVectorSDNode::recastRawBits(IsLE, BVVT.getScalarSizeInBits(),
+                                           DstBits, RawBits, DstUndefs,
+                                           BitVector(RawBits.size(), false));
+          EVT BVEltVT = BV1->getOperand(0).getValueType();
+          unsigned BVEltBits = BVEltVT.getSizeInBits();
+          SmallVector<SDValue> Ops(DstBits.size(), getUNDEF(BVEltVT));
+          for (unsigned I = 0, E = DstBits.size(); I != E; ++I) {
+            if (DstUndefs[I])
+              continue;
+            Ops[I] = getConstant(DstBits[I].sextOrSelf(BVEltBits), DL, BVEltVT);
+          }
+          return getBitcast(VT, getBuildVector(BVVT, DL, Ops));
+        }
+      }
+    }
+  }
+
   auto IsScalarOrSameVectorSize = [NumElts](const SDValue &Op) {
     return !Op.getValueType().isVector() ||
            Op.getValueType().getVectorElementCount() == NumElts;
   };
 
   auto IsBuildVectorSplatVectorOrUndef = [](const SDValue &Op) {
     return Op.isUndef() || Op.getOpcode() == ISD::CONDCODE ||
            Op.getOpcode() == ISD::BUILD_VECTOR ||

llvm/test/CodeGen/X86/packss.ll

 }
 
 define <8 x i16> @trunc_ashr_v4i64_demandedelts(<4 x i64> %a0) {
 ; X86-SSE-LABEL: trunc_ashr_v4i64_demandedelts:
 ; X86-SSE:       # %bb.0:
 ; X86-SSE-NEXT:    psllq $63, %xmm1
 ; X86-SSE-NEXT:    psllq $63, %xmm0
 ; X86-SSE-NEXT:    psrlq $63, %xmm0
-; X86-SSE-NEXT:    movdqa {{.*#+}} xmm2 = [1,0,0,0]
+; X86-SSE-NEXT:    movdqa {{.*#+}} xmm2 = [1,0,0,2147483648]
 ; X86-SSE-NEXT:    pxor %xmm2, %xmm0
 ; X86-SSE-NEXT:    psubq %xmm2, %xmm0
 ; X86-SSE-NEXT:    psrlq $63, %xmm1
 ; X86-SSE-NEXT:    pxor %xmm2, %xmm1
 ; X86-SSE-NEXT:    psubq %xmm2, %xmm1
 ; X86-SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,0,0]
 ; X86-SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
 ; X86-SSE-NEXT:    packssdw %xmm1, %xmm0

 ; X86-AVX1-NEXT:    vzeroupper
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: trunc_ashr_v4i64_demandedelts:
 ; X86-AVX2:       # %bb.0:
 ; X86-AVX2-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [63,0,0,0,63,0,0,0]
 ; X86-AVX2-NEXT:    # ymm1 = mem[0,1,0,1]
 ; X86-AVX2-NEXT:    vpsllvq %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,2147483648,0,2147483648,0,2147483648,0,2147483648]
-; X86-AVX2-NEXT:    vpsrlvq %ymm1, %ymm2, %ymm2
 ; X86-AVX2-NEXT:    vpsrlvq %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpxor %ymm2, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsubq %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [1,0,0,2147483648,1,0,0,2147483648]
+; X86-AVX2-NEXT:    # ymm1 = mem[0,1,0,1]
+; X86-AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpsubq %ymm1, %ymm0, %ymm0
 ; X86-AVX2-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,0,0,0,4,4,4,4]
 ; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
 ; X86-AVX2-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
 ; X86-AVX2-NEXT:    vzeroupper
 ; X86-AVX2-NEXT:    retl
 ;
 ; X64-SSE-LABEL: trunc_ashr_v4i64_demandedelts:
 ; X64-SSE:       # %bb.0:

llvm/test/CodeGen/X86/vector-fshl-rot-128.ll

 ;
 ; XOP-LABEL: constant_funnnel_v2i64:
 ; XOP:       # %bb.0:
 ; XOP-NEXT:    vprotq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
 ; XOP-NEXT:    retq
 ;
 ; X86-SSE2-LABEL: constant_funnnel_v2i64:
 ; X86-SSE2:       # %bb.0:
-; X86-SSE2-NEXT:    pxor %xmm1, %xmm1
-; X86-SSE2-NEXT:    psubq {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
-; X86-SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
+; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT:    psrlq $60, %xmm1
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psrlq %xmm1, %xmm2
-; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm3
-; X86-SSE2-NEXT:    psrlq %xmm1, %xmm3
-; X86-SSE2-NEXT:    movsd {{.*#+}} xmm3 = xmm2[0],xmm3[1]
+; X86-SSE2-NEXT:    psrlq $50, %xmm2
+; X86-SSE2-NEXT:    movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psllq $4, %xmm1
 ; X86-SSE2-NEXT:    psllq $14, %xmm0
 ; X86-SSE2-NEXT:    movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
-; X86-SSE2-NEXT:    orpd %xmm3, %xmm0
+; X86-SSE2-NEXT:    orpd %xmm2, %xmm0
 ; X86-SSE2-NEXT:    retl
   %res = call <2 x i64> @llvm.fshl.v2i64(<2 x i64> %x, <2 x i64> %x, <2 x i64> <i64 4, i64 14>)
   ret <2 x i64> %res
 }
 
 define <4 x i32> @constant_funnnel_v4i32(<4 x i32> %x) nounwind {
 ; SSE2-LABEL: constant_funnnel_v4i32:
 ; SSE2:       # %bb.0:

 ; XOP-LABEL: splatconstant_funnnel_v2i64:
 ; XOP:       # %bb.0:
 ; XOP-NEXT:    vprotq $14, %xmm0, %xmm0
 ; XOP-NEXT:    retq
 ;
 ; X86-SSE2-LABEL: splatconstant_funnnel_v2i64:
 ; X86-SSE2:       # %bb.0:
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE2-NEXT:    psllq $14, %xmm1
-; X86-SSE2-NEXT:    psrlq $50, %xmm0
-; X86-SSE2-NEXT:    movsd {{.*#+}} xmm0 = xmm0[0,1]
-; X86-SSE2-NEXT:    orpd %xmm1, %xmm0
+; X86-SSE2-NEXT:    psrlq $50, %xmm1
+; X86-SSE2-NEXT:    psllq $14, %xmm0
+; X86-SSE2-NEXT:    por %xmm1, %xmm0
 ; X86-SSE2-NEXT:    retl
   %res = call <2 x i64> @llvm.fshl.v2i64(<2 x i64> %x, <2 x i64> %x, <2 x i64> <i64 14, i64 14>)
   ret <2 x i64> %res
 }
 
 define <4 x i32> @splatconstant_funnnel_v4i32(<4 x i32> %x) nounwind {
 ; SSE-LABEL: splatconstant_funnnel_v4i32:
 ; SSE:       # %bb.0:

llvm/test/CodeGen/X86/vector-fshr-rot-128.ll

 ;
 ; XOP-LABEL: constant_funnnel_v2i64:
 ; XOP:       # %bb.0:
 ; XOP-NEXT:    vprotq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
 ; XOP-NEXT:    retq
 ;
 ; X86-SSE2-LABEL: constant_funnnel_v2i64:
 ; X86-SSE2:       # %bb.0:
-; X86-SSE2-NEXT:    pxor %xmm1, %xmm1
-; X86-SSE2-NEXT:    psubq {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
-; X86-SSE2-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
+; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT:    psllq $60, %xmm1
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
-; X86-SSE2-NEXT:    psllq %xmm1, %xmm2
-; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
-; X86-SSE2-NEXT:    movdqa %xmm0, %xmm3
-; X86-SSE2-NEXT:    psllq %xmm1, %xmm3
-; X86-SSE2-NEXT:    movsd {{.*#+}} xmm3 = xmm2[0],xmm3[1]
+; X86-SSE2-NEXT:    psllq $50, %xmm2
+; X86-SSE2-NEXT:    movsd {{.*#+}} xmm2 = xmm1[0],xmm2[1]
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE2-NEXT:    psrlq $4, %xmm1
 ; X86-SSE2-NEXT:    psrlq $14, %xmm0
 ; X86-SSE2-NEXT:    movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
-; X86-SSE2-NEXT:    orpd %xmm3, %xmm0
+; X86-SSE2-NEXT:    orpd %xmm2, %xmm0
 ; X86-SSE2-NEXT:    retl
   %res = call <2 x i64> @llvm.fshr.v2i64(<2 x i64> %x, <2 x i64> %x, <2 x i64> <i64 4, i64 14>)
   ret <2 x i64> %res
 }
 
 define <4 x i32> @constant_funnnel_v4i32(<4 x i32> %x) nounwind {
 ; SSE2-LABEL: constant_funnnel_v4i32:
 ; SSE2:       # %bb.0:

 ; XOP-LABEL: splatconstant_funnnel_v2i64:
 ; XOP:       # %bb.0:
 ; XOP-NEXT:    vprotq $50, %xmm0, %xmm0
 ; XOP-NEXT:    retq
 ;
 ; X86-SSE2-LABEL: splatconstant_funnnel_v2i64:
 ; X86-SSE2:       # %bb.0:
 ; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
-; X86-SSE2-NEXT:    psrlq $14, %xmm1
-; X86-SSE2-NEXT:    psllq $50, %xmm0
-; X86-SSE2-NEXT:    movsd {{.*#+}} xmm0 = xmm0[0,1]
-; X86-SSE2-NEXT:    orpd %xmm1, %xmm0
+; X86-SSE2-NEXT:    psllq $50, %xmm1
+; X86-SSE2-NEXT:    psrlq $14, %xmm0
+; X86-SSE2-NEXT:    por %xmm1, %xmm0
 ; X86-SSE2-NEXT:    retl
   %res = call <2 x i64> @llvm.fshr.v2i64(<2 x i64> %x, <2 x i64> %x, <2 x i64> <i64 14, i64 14>)
   ret <2 x i64> %res
 }
 
 define <4 x i32> @splatconstant_funnnel_v4i32(<4 x i32> %x) nounwind {
 ; SSE-LABEL: splatconstant_funnnel_v4i32:
 ; SSE:       # %bb.0:

llvm/test/CodeGen/X86/vector-shift-ashr-128.ll

 ; AVX512VL-NEXT:    retq
 ;
 ; X86-SSE-LABEL: constant_shift_v2i64:
 ; X86-SSE:       # %bb.0:
 ; X86-SSE-NEXT:    movdqa %xmm0, %xmm1
 ; X86-SSE-NEXT:    psrlq $1, %xmm1
 ; X86-SSE-NEXT:    psrlq $7, %xmm0
 ; X86-SSE-NEXT:    movsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
-; X86-SSE-NEXT:    movapd {{.*#+}} xmm1 = [2.0E+0,7.2911220195563975E-304]
+; X86-SSE-NEXT:    movapd {{.*#+}} xmm1 = [0,1073741824,0,16777216]
 ; X86-SSE-NEXT:    xorpd %xmm1, %xmm0
 ; X86-SSE-NEXT:    psubq %xmm1, %xmm0
 ; X86-SSE-NEXT:    retl
   %shift = ashr <2 x i64> %a, <i64 1, i64 7>
   ret <2 x i64> %shift
 }
 
 define <4 x i32> @constant_shift_v4i32(<4 x i32> %a) nounwind {

llvm/test/CodeGen/X86/vector-shift-ashr-256.ll

 ; X86-AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,1073741824,0,16777216]
 ; X86-AVX1-NEXT:    vpxor %xmm2, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vpsubq %xmm2, %xmm0, %xmm0
 ; X86-AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; X86-AVX1-NEXT:    retl
 ;
 ; X86-AVX2-LABEL: constant_shift_v4i64:
 ; X86-AVX2:       # %bb.0:
-; X86-AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [1,0,7,0,31,0,62,0]
-; X86-AVX2-NEXT:    vmovdqa {{.*#+}} ymm2 = [0,2147483648,0,2147483648,0,2147483648,0,2147483648]
-; X86-AVX2-NEXT:    vpsrlvq %ymm1, %ymm2, %ymm2
-; X86-AVX2-NEXT:    vpsrlvq %ymm1, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpxor %ymm2, %ymm0, %ymm0
-; X86-AVX2-NEXT:    vpsubq %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpsrlvq {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [0,1073741824,0,16777216,0,1,2,0]
+; X86-AVX2-NEXT:    vpxor %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT:    vpsubq %ymm1, %ymm0, %ymm0
 ; X86-AVX2-NEXT:    retl
   %shift = ashr <4 x i64> %a, <i64 1, i64 7, i64 31, i64 62>
   ret <4 x i64> %shift
 }
 
 define <8 x i32> @constant_shift_v8i32(<8 x i32> %a) nounwind {
 ; AVX1-LABEL: constant_shift_v8i32:
 ; AVX1:       # %bb.0: