[X86][CodeGen]Fix extract f16 from big vectors

[XinWang10]

Summary

When use llc -mattr=+avx512fp16, it will crash.

define half @test(<64 x half> %x, i64 %idx){
   %res = extractelement <64 x half> %x, i64 %idx
   ret half %res
}

The root cause is when we enable avx512fp16 we lose custom handler
for extract f16 from big vectors which is not loaded from pointer.

[XinWang10]

This comment was removed by XinWang10.

[XinWang10]

This comment was removed by XinWang10.

[XinWang10]

This comment was removed by XinWang10.

[XinWang10]

This comment was removed by XinWang10.

[XinWang10]

This comment was removed by XinWang10.

[XinWang10]

Fix extract f16 from big vectors

[pengfei]

llvm/lib/Target/X86/X86ISelLowering.cpp

34125

We usually use vNf16

34126–34128

Try to avoid defining variable used by assert only. Compiler may report used variable when do releast build.

34131–34132

You can use changeVectorElementTypeToInteger.

llvm/test/CodeGen/X86/avx512fp16-mov.ll

1352

Add nounwind.

[XinWang10]

Updating D141348: [X86][CodeGen]Fix extract f16 from big vectors

[pengfei]

llvm/lib/Target/X86/X86ISelLowering.cpp

34125

Sorry, I want to say vXf16.

[XinWang10]

Change tests

[XinWang10]

Update comments

[XinWang10]

llvm/lib/Target/X86/X86ISelLowering.cpp

34125

Thanks.

[pengfei]

LGTM.

llvm/lib/Target/X86/X86ISelLowering.cpp

     Results.push_back(Res.getValue(1));
     return;
   }
   case ISD::ADDRSPACECAST: {
     SDValue V = LowerADDRSPACECAST(SDValue(N,0), DAG);
     Results.push_back(V);
     return;
   }
-  case ISD::BITREVERSE:
+  case ISD::BITREVERSE: {
     assert(N->getValueType(0) == MVT::i64 && "Unexpected VT!");
     assert(Subtarget.hasXOP() && "Expected XOP");
     // We can use VPPERM by copying to a vector register and back. We'll need
     // to move the scalar in two i32 pieces.
     Results.push_back(LowerBITREVERSE(SDValue(N, 0), Subtarget, DAG));
     return;
   }
+  case ISD::EXTRACT_VECTOR_ELT: {
+    // f16 = extract vXf16 %vec, i64 %idx
+    assert(N->getSimpleValueType(0) == MVT::f16 &&
+           "Unexpected Value type of EXTRACT_VECTOR_ELT!");
+    assert(Subtarget.hasFP16() && "Expected FP16");
+    SDValue VecOp = N->getOperand(0);
+    EVT ExtVT = VecOp.getValueType().changeVectorElementTypeToInteger();
+    SDValue Split = DAG.getBitcast(ExtVT, N->getOperand(0));
+    Split = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Split,
+                        N->getOperand(1));
+    Split = DAG.getBitcast(MVT::f16, Split);
+    Results.push_back(Split);
+    return;
+  }
+  }
 }
 
 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   switch ((X86ISD::NodeType)Opcode) {
   case X86ISD::FIRST_NUMBER:       break;
 #define NODE_NAME_CASE(NODE) case X86ISD::NODE: return "X86ISD::" #NODE;
   NODE_NAME_CASE(BSF)
   NODE_NAME_CASE(BSR)

llvm/test/CodeGen/X86/avx512fp16-mov.ll

 ; CHECK-LABEL: extract_f16_7:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
 ; CHECK-NEXT:    ret{{[l|q]}}
    %res = extractelement <8 x half> %x, i32 7
    ret half %res
 }
 
+define half @extract_f16_8(<32 x half> %x, i64 %idx) nounwind {
+; X64-LABEL: extract_f16_8:
+; X64:       # %bb.0:
+; X64-NEXT:    pushq %rbp
+; X64-NEXT:    movq %rsp, %rbp
+; X64-NEXT:    andq $-64, %rsp
+; X64-NEXT:    subq $128, %rsp
+; X64-NEXT:    andl $31, %edi
+; X64-NEXT:    vmovaps %zmm0, (%rsp)
+; X64-NEXT:    vmovsh (%rsp,%rdi,2), %xmm0
+; X64-NEXT:    movq %rbp, %rsp
+; X64-NEXT:    popq %rbp
+; X64-NEXT:    vzeroupper
+; X64-NEXT:    retq
+;
+; X86-LABEL: extract_f16_8:
+; X86:       # %bb.0:
+; X86-NEXT:    pushl %ebp
+; X86-NEXT:    movl %esp, %ebp
+; X86-NEXT:    andl $-64, %esp
+; X86-NEXT:    subl $128, %esp
+; X86-NEXT:    movl 8(%ebp), %eax
+; X86-NEXT:    andl $31, %eax
+; X86-NEXT:    vmovaps %zmm0, (%esp)
+; X86-NEXT:    vmovsh (%esp,%eax,2), %xmm0
+; X86-NEXT:    movl %ebp, %esp
+; X86-NEXT:    popl %ebp
+; X86-NEXT:    vzeroupper
+; X86-NEXT:    retl
+   %res = extractelement <32 x half> %x, i64 %idx
+   ret half %res
+}
+
+define half @extract_f16_9(<64 x half> %x, i64 %idx) nounwind {
+; X64-LABEL: extract_f16_9:
+; X64:       # %bb.0:
+; X64-NEXT:    pushq %rbp
+; X64-NEXT:    movq %rsp, %rbp
+; X64-NEXT:    andq $-64, %rsp
+; X64-NEXT:    subq $192, %rsp
+; X64-NEXT:    andl $63, %edi
+; X64-NEXT:    vmovaps %zmm1, {{[0-9]+}}(%rsp)
+; X64-NEXT:    vmovaps %zmm0, (%rsp)
+; X64-NEXT:    vmovsh (%rsp,%rdi,2), %xmm0
+; X64-NEXT:    movq %rbp, %rsp
+; X64-NEXT:    popq %rbp
+; X64-NEXT:    vzeroupper
+; X64-NEXT:    retq
+;
+; X86-LABEL: extract_f16_9:
+; X86:       # %bb.0:
+; X86-NEXT:    pushl %ebp
+; X86-NEXT:    movl %esp, %ebp
+; X86-NEXT:    andl $-64, %esp
+; X86-NEXT:    subl $192, %esp
+; X86-NEXT:    movl 8(%ebp), %eax
+; X86-NEXT:    andl $63, %eax
+; X86-NEXT:    vmovaps %zmm1, {{[0-9]+}}(%esp)
+; X86-NEXT:    vmovaps %zmm0, (%esp)
+; X86-NEXT:    vmovsh (%esp,%eax,2), %xmm0
+; X86-NEXT:    movl %ebp, %esp
+; X86-NEXT:    popl %ebp
+; X86-NEXT:    vzeroupper
+; X86-NEXT:    retl
+   %res = extractelement <64 x half> %x, i64 %idx
+   ret half %res
+}
+
 define i16 @extract_i16_0(<8 x i16> %x) {
 ; CHECK-LABEL: extract_i16_0:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vmovw %xmm0, %eax
 ; CHECK-NEXT:    # kill: def $ax killed $ax killed $eax
 ; CHECK-NEXT:    ret{{[l|q]}}
    %res = extractelement <8 x i16> %x, i32 0
    ret i16 %res

 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    movsbl %dil, %eax
 ; X64-NEXT:    vmovw %eax, %xmm1
 ; X64-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; X64-NEXT:    vpcmpgtw %xmm2, %xmm1, %k1
 ; X64-NEXT:    vmovdqu16 %xmm0, %xmm0 {%k1} {z}
 ; X64-NEXT:    vmovw %xmm0, %eax
 ; X64-NEXT:    testw %ax, %ax
-; X64-NEXT:    je .LBB121_2
+; X64-NEXT:    je .LBB123_2
 ; X64-NEXT:  # %bb.1: # %for.body.preheader
 ; X64-NEXT:    movb $0, (%rsi)
-; X64-NEXT:  .LBB121_2: # %for.end
+; X64-NEXT:  .LBB123_2: # %for.end
 ; X64-NEXT:    retq
 ;
 ; X86-LABEL: pr52560:
 ; X86:       # %bb.0: # %entry
 ; X86-NEXT:    movsbl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    vmovw %eax, %xmm1
 ; X86-NEXT:    vpxor %xmm2, %xmm2, %xmm2
 ; X86-NEXT:    vpcmpgtw %xmm2, %xmm1, %k1
 ; X86-NEXT:    vmovdqu16 %xmm0, %xmm0 {%k1} {z}
 ; X86-NEXT:    vmovw %xmm0, %eax
 ; X86-NEXT:    testw %ax, %ax
-; X86-NEXT:    je .LBB121_2
+; X86-NEXT:    je .LBB123_2
 ; X86-NEXT:  # %bb.1: # %for.body.preheader
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    movb $0, (%eax)
-; X86-NEXT:  .LBB121_2: # %for.end
+; X86-NEXT:  .LBB123_2: # %for.end
 ; X86-NEXT:    retl
 entry:
   %conv = sext i8 %0 to i16
   %2 = insertelement <2 x i16> <i16 poison, i16 0>, i16 %conv, i32 0
   %3 = icmp sgt <2 x i16> %2, zeroinitializer
   %4 = select <2 x i1> %3, <2 x i16> %1, <2 x i16> <i16 0, i16 poison>
   %5 = extractelement <2 x i16> %4, i32 0
   %tobool.not14 = icmp eq i16 %5, 0