[X86][BF16] Fix 2 crashes with vector broadcast

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D151808

(cherry picked from commit 801dd88)

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -2195,6 +2195,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
       setOperationAction(ISD::FMUL, VT, Expand);
       setOperationAction(ISD::FDIV, VT, Expand);
       setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
+      setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
     }
     addLegalFPImmediate(APFloat::getZero(APFloat::BFloat()));
   }
@@ -2207,6 +2208,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::FMUL, MVT::v32bf16, Expand);
     setOperationAction(ISD::FDIV, MVT::v32bf16, Expand);
     setOperationAction(ISD::BUILD_VECTOR, MVT::v32bf16, Custom);
+    setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32bf16, Custom);
   }
 
   if (!Subtarget.useSoftFloat() && Subtarget.hasVLX()) {
@@ -18773,11 +18775,11 @@ static SDValue lower256BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
     return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
   }
 
-  if (VT == MVT::v16f16) {
-    V1 = DAG.getBitcast(MVT::v16i16, V1);
-    V2 = DAG.getBitcast(MVT::v16i16, V2);
-    return DAG.getBitcast(MVT::v16f16,
-                          DAG.getVectorShuffle(MVT::v16i16, DL, V1, V2, Mask));
+  if (VT == MVT::v16f16 || VT.getVectorElementType() == MVT::bf16) {
+    MVT IVT = VT.changeVectorElementTypeToInteger();
+    V1 = DAG.getBitcast(IVT, V1);
+    V2 = DAG.getBitcast(IVT, V2);
+    return DAG.getBitcast(VT, DAG.getVectorShuffle(IVT, DL, V1, V2, Mask));
   }
 
   switch (VT.SimpleTy) {

llvm/lib/Target/X86/X86InstrAVX512.td

@@ -12969,6 +12969,27 @@ let Predicates = [HasBF16, HasVLX] in {
             (VCVTNEPS2BF16Z256rr VR256X:$src)>;
   def : Pat<(v8bf16 (int_x86_vcvtneps2bf16256 (loadv8f32 addr:$src))),
             (VCVTNEPS2BF16Z256rm addr:$src)>;
+
+  def : Pat<(v8bf16 (X86VBroadcastld16 addr:$src)),
+            (VPBROADCASTWZ128rm addr:$src)>;
+  def : Pat<(v16bf16 (X86VBroadcastld16 addr:$src)),
+            (VPBROADCASTWZ256rm addr:$src)>;
+
+  def : Pat<(v8bf16 (X86VBroadcast (v8bf16 VR128X:$src))),
+            (VPBROADCASTWZ128rr VR128X:$src)>;
+  def : Pat<(v16bf16 (X86VBroadcast (v8bf16 VR128X:$src))),
+            (VPBROADCASTWZ256rr VR128X:$src)>;
+
+  // TODO: No scalar broadcast due to we don't support legal scalar bf16 so far.
+}
+
+let Predicates = [HasBF16] in {
+  def : Pat<(v32bf16 (X86VBroadcastld16 addr:$src)),
+            (VPBROADCASTWZrm addr:$src)>;
+
+  def : Pat<(v32bf16 (X86VBroadcast (v8bf16 VR128X:$src))),
+            (VPBROADCASTWZrr VR128X:$src)>;
+  // TODO: No scalar broadcast due to we don't support legal scalar bf16 so far.
 }
 
 let Constraints = "$src1 = $dst" in {

llvm/test/CodeGen/X86/avx512bf16-vl-intrinsics.ll

@@ -356,3 +356,49 @@ entry:
   %2 = select <4 x i1> %1, <4 x float> %0, <4 x float> %E
   ret <4 x float> %2
 }
+
+define <16 x i16> @test_no_vbroadcast1() {
+; CHECK-LABEL: test_no_vbroadcast1:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vcvtneps2bf16 %xmm0, %xmm0 # encoding: [0x62,0xf2,0x7e,0x08,0x72,0xc0]
+; CHECK-NEXT:    vpbroadcastw %xmm0, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x79,0xc0]
+; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
+entry:
+  %0 = tail call <8 x bfloat> @llvm.x86.avx512bf16.mask.cvtneps2bf16.128(<4 x float> poison, <8 x bfloat> zeroinitializer, <4 x i1> <i1 true, i1 true, i1 true, i1 true>)
+  %1 = bitcast <8 x bfloat> %0 to <8 x i16>
+  %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <16 x i32> zeroinitializer
+  ret <16 x i16> %2
+}
+
+;; FIXME: This should generate the same output as above, but let's fix the crash first.
+define <16 x bfloat> @test_no_vbroadcast2() nounwind {
+; X86-LABEL: test_no_vbroadcast2:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    pushl %ebp # encoding: [0x55]
+; X86-NEXT:    movl %esp, %ebp # encoding: [0x89,0xe5]
+; X86-NEXT:    andl $-32, %esp # encoding: [0x83,0xe4,0xe0]
+; X86-NEXT:    subl $64, %esp # encoding: [0x83,0xec,0x40]
+; X86-NEXT:    vcvtneps2bf16 %xmm0, %xmm0 # encoding: [0x62,0xf2,0x7e,0x08,0x72,0xc0]
+; X86-NEXT:    vmovaps %xmm0, (%esp) # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x29,0x04,0x24]
+; X86-NEXT:    vpbroadcastw (%esp), %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x79,0x04,0x24]
+; X86-NEXT:    movl %ebp, %esp # encoding: [0x89,0xec]
+; X86-NEXT:    popl %ebp # encoding: [0x5d]
+; X86-NEXT:    retl # encoding: [0xc3]
+;
+; X64-LABEL: test_no_vbroadcast2:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    pushq %rbp # encoding: [0x55]
+; X64-NEXT:    movq %rsp, %rbp # encoding: [0x48,0x89,0xe5]
+; X64-NEXT:    andq $-32, %rsp # encoding: [0x48,0x83,0xe4,0xe0]
+; X64-NEXT:    subq $64, %rsp # encoding: [0x48,0x83,0xec,0x40]
+; X64-NEXT:    vcvtneps2bf16 %xmm0, %xmm0 # encoding: [0x62,0xf2,0x7e,0x08,0x72,0xc0]
+; X64-NEXT:    vmovaps %xmm0, (%rsp) # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x29,0x04,0x24]
+; X64-NEXT:    vpbroadcastw (%rsp), %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x79,0x04,0x24]
+; X64-NEXT:    movq %rbp, %rsp # encoding: [0x48,0x89,0xec]
+; X64-NEXT:    popq %rbp # encoding: [0x5d]
+; X64-NEXT:    retq # encoding: [0xc3]
+entry:
+  %0 = tail call <8 x bfloat> @llvm.x86.avx512bf16.mask.cvtneps2bf16.128(<4 x float> poison, <8 x bfloat> zeroinitializer, <4 x i1> <i1 true, i1 true, i1 true, i1 true>)
+  %1 = shufflevector <8 x bfloat> %0, <8 x bfloat> undef, <16 x i32> zeroinitializer
+  ret <16 x bfloat> %1
+}