[X86] Only call combineBitcastToBoolVector after legalization

[RKSimon]

Prevents infinite loop between combineBitcastToBoolVector and hoistLogicOpWithSameOpcodeHands, which only performs the "logicop(bitcast(A),bitcast(B)) -> bitcast(logicop(A,B))" upto type legalization.

combineBitcastToBoolVector doesn't care much as its mainly for AVX512 cleanup that X86DomainReassignment can't handle for us.

Fixes #123333

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

Changes

Prevents infinite loop between combineBitcastToBoolVector and hoistLogicOpWithSameOpcodeHands, which only performs the "logicop(bitcast(A),bitcast(B)) -> bitcast(logicop(A,B))" upto type legalization.

combineBitcastToBoolVector doesn't care much as its mainly for AVX512 cleanup that X86DomainReassignment can't handle for us.

Fixes #123333

---

Full diff: https://github.com/llvm/llvm-project/pull/123386.diff

2 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+1-1)
• (added) llvm/test/CodeGen/X86/pr123333.ll (+34)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 84736f18011a9d..b74a2d7d2e8480 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -45053,7 +45053,7 @@ static SDValue combineBitcast(SDNode *N, SelectionDAG &DAG,
       N0 = DAG.getBitcast(MVT::i8, N0);
       return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
     }
-  } else {
+  } else if (DCI.isAfterLegalizeDAG()) {
     // If we're bitcasting from iX to vXi1, see if the integer originally
     // began as a vXi1 and whether we can remove the bitcast entirely.
     if (VT.isVector() && VT.getScalarType() == MVT::i1 &&
diff --git a/llvm/test/CodeGen/X86/pr123333.ll b/llvm/test/CodeGen/X86/pr123333.ll
new file mode 100644
index 00000000000000..3d2ac257f3b9a6
--- /dev/null
+++ b/llvm/test/CodeGen/X86/pr123333.ll
@@ -0,0 +1,34 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=x86_64-- -mcpu=x86-64-v4 | FileCheck %s
+
+define void @PR123333(ptr %in0, ptr %in1, ptr %in2, ptr %out0, ptr %out1) {
+; CHECK-LABEL: PR123333:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    kmovq (%rsi), %k0
+; CHECK-NEXT:    kmovq (%rdi), %k1
+; CHECK-NEXT:    korq %k1, %k0, %k2
+; CHECK-NEXT:    vmovdqu8 (%rdx), %zmm0 {%k2} {z}
+; CHECK-NEXT:    vmovdqu64 %zmm0, (%rcx)
+; CHECK-NEXT:    kandnq %k0, %k1, %k0
+; CHECK-NEXT:    knotq %k0, %k1
+; CHECK-NEXT:    vmovdqu8 {{.*#+}} zmm0 {%k1} {z} = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; CHECK-NEXT:    vmovdqu64 %zmm0, (%r8)
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+  %i0 = load i64, ptr %in0, align 8
+  %i1 = load i64, ptr %in1, align 8
+  %i2 = load <64 x i8>, ptr %in2, align 1
+  %i7 = bitcast i64 %i1 to <64 x i1>
+  %i9 = bitcast i64 %i0 to <64 x i1>
+  %i12 = or i64 %i1, %i0
+  %i13 = bitcast i64 %i12 to <64 x i1>
+  %i15 = or <64 x i1> %i13, %i9
+  %i17 = select <64 x i1> %i15, <64 x i8> %i2, <64 x i8> zeroinitializer
+  store <64 x i8> %i17, ptr %out0, align 16
+  %i21 = icmp eq <64 x i8> %i2, splat (i8 4)
+  %i22 = xor <64 x i1> %i9, splat (i1 true)
+  %i23 = and <64 x i1> %i22, %i7
+  %i25 = select <64 x i1> %i23, <64 x i8> zeroinitializer, <64 x i8> splat (i8 15)
+  store <64 x i8> %i25, ptr %out1, align 16
+  ret void
+}