[X86] Immediately call LowerShift from lowerBuildVectorToBitOp.

[craig.topper]

Summary

If we don't immediately lower the vector shift, the splat
constant vector we created may get turned into a constant pool
load before we get around to lowering the shift. This makes it
a lot more difficult to create a shift by constant. Sometimes we
fail to see through the constant pool at all and end up trying
to lower as if it was a variable shift. This requires custom
handling and may create an unsupported vselect on pre-sse-4.1
targets. Since we're after LegalizeVectorOps we are unable to
legalize the unsupported vselect as that code is in LegalizeDAG.

So calling LowerShift immediately ensures that we get see the
splat constant.

Fixes PR46527.

[RKSimon]

LGTM - cheers

llvm/lib/Target/X86/X86ISelLowering.cpp

if (VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
     bool isUndefHI = NumUndefsHI == Half;
     return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
                                  isUndefLO, isUndefHI);
   }
 
   return SDValue();
 }
 
+static SDValue LowerShift(SDValue Op, const X86Subtarget &Subtarget,
+                          SelectionDAG &DAG);
+
 /// If a BUILD_VECTOR's source elements all apply the same bit operation and
 /// one of their operands is constant, lower to a pair of BUILD_VECTOR and
 /// just apply the bit to the vectors.
 /// NOTE: Its not in our interest to start make a general purpose vectorizer
 /// from this, but enough scalar bit operations are created from the later
 /// legalization + scalarization stages to need basic support.
 static SDValue lowerBuildVectorToBitOp(BuildVectorSDNode *Op,
+                                       const X86Subtarget &Subtarget,
                                        SelectionDAG &DAG) {
   SDLoc DL(Op);
   MVT VT = Op->getSimpleValueType(0);
   unsigned NumElems = VT.getVectorNumElements();
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
   // Check that all elements have the same opcode.
   // TODO: Should we allow UNDEFS and if so how many?
static SDValue lowerBuildVectorToBitOp(BuildVectorSDNode *Op,
   // Limit to shifts by uniform immediates.
   // TODO: Only accept vXi8/vXi64 special cases?
   // TODO: Permit non-uniform XOP/AVX2/MULLO cases?
   if (IsShift && any_of(RHSElts, [&](SDValue V) { return RHSElts[0] != V; }))
     return SDValue();
 
   SDValue LHS = DAG.getBuildVector(VT, DL, LHSElts);
   SDValue RHS = DAG.getBuildVector(VT, DL, RHSElts);
-  return DAG.getNode(Opcode, DL, VT, LHS, RHS);
+  SDValue Res = DAG.getNode(Opcode, DL, VT, LHS, RHS);
+
+  if (!IsShift)
+    return Res;
+
+  // Immediately lower the shift to ensure the constant build vector doesn't
+  // get converted to a constant pool before the shift is lowered.
+  return LowerShift(Res, Subtarget, DAG);
 }
 
 /// Create a vector constant without a load. SSE/AVX provide the bare minimum
 /// functionality to do this, so it's all zeros, all ones, or some derivation
 /// that is cheap to calculate.
 static SDValue materializeVectorConstant(SDValue Op, SelectionDAG &DAG,
                                          const X86Subtarget &Subtarget) {
   SDLoc DL(Op);
X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
 
   BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
   if (SDValue AddSub = lowerToAddSubOrFMAddSub(BV, Subtarget, DAG))
     return AddSub;
   if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
     return HorizontalOp;
   if (SDValue Broadcast = lowerBuildVectorAsBroadcast(BV, Subtarget, DAG))
     return Broadcast;
-  if (SDValue BitOp = lowerBuildVectorToBitOp(BV, DAG))
+  if (SDValue BitOp = lowerBuildVectorToBitOp(BV, Subtarget, DAG))
     return BitOp;
 
   unsigned EVTBits = EltVT.getSizeInBits();
 
   unsigned NumZero  = 0;
   unsigned NumNonZero = 0;
   uint64_t NonZeros = 0;
   bool IsAllConstants = true;

llvm/test/CodeGen/X86/pr46527.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+;RUN: llc < %s -mtriple=i686-unknown -mattr=sse2 -relocation-model=pic | FileCheck %s
+
+define void @f(<16 x i8>* %out, <16 x i8> %in, i1 %flag) {
+; CHECK-LABEL: f:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    calll .L0$pb
+; CHECK-NEXT:    .cfi_adjust_cfa_offset 4
+; CHECK-NEXT:  .L0$pb:
+; CHECK-NEXT:    popl %eax
+; CHECK-NEXT:    .cfi_adjust_cfa_offset -4
+; CHECK-NEXT:  .Ltmp0:
+; CHECK-NEXT:    addl $_GLOBAL_OFFSET_TABLE_+(.Ltmp0-.L0$pb), %eax
+; CHECK-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; CHECK-NEXT:    movb {{[0-9]+}}(%esp), %dl
+; CHECK-NEXT:    notb %dl
+; CHECK-NEXT:    andb $1, %dl
+; CHECK-NEXT:    movzbl %dl, %edx
+; CHECK-NEXT:    movd %edx, %xmm1
+; CHECK-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; CHECK-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[0,0,2,3,4,5,6,7]
+; CHECK-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[0,0,0,0]
+; CHECK-NEXT:    paddb %xmm1, %xmm1
+; CHECK-NEXT:    pxor %xmm0, %xmm1
+; CHECK-NEXT:    pxor {{\.LCPI.*}}@GOTOFF(%eax), %xmm1
+; CHECK-NEXT:    movdqa %xmm1, (%ecx)
+; CHECK-NEXT:    retl
+entry:
+  %0 = select i1 %flag, i8 0, i8 2
+  %1 = insertelement <16 x i8> undef, i8 %0, i32 0
+  %2 = shufflevector <16 x i8> %1, <16 x i8> undef, <16 x i32> zeroinitializer
+  %3 = xor <16 x i8> %2, %in
+  %4 = xor <16 x i8> %3, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
+  store <16 x i8> %4, <16 x i8>* %out, align 16
+  ret void
+}