[X86] Do not lower INSERT_VECTOR_ELT to vselect for vXf16 without BWI

[pengfei]

Summary

We cannot handle i8/i16/f16 vselect without BWI.

Fixes #59980

[RKSimon]

LGTM with a few minor style suggestions

llvm/test/CodeGen/X86/pr59980.ll

8

(style) simplify these names / types and attributes (define void @PR59980) - strip noalias/nocapture/dereferenceable etc.

[skan]

You can use opt --metarenamer to do the renaming

[lebedev.ri]

llvm/test/CodeGen/X86/pr59980.ll

28

Please don't add new tests with typed pointers.

[pengfei]

In D141668#4050934, @skan wrote:

You can use opt --metarenamer to do the renaming

Thanks for the review! I used opt -passes=metarenamer,strip llvm/test/CodeGen/X86/pr59980.ll -S -o - --opaque-pointers. It reduces many manual work :)

llvm/lib/Target/X86/X86ISelLowering.cpp

SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
   auto *N2C = dyn_cast<ConstantSDNode>(N2);
 
   if (!N2C) {
     // Variable insertion indices, usually we're better off spilling to stack,
     // but AVX512 can use a variable compare+select by comparing against all
     // possible vector indices, and FP insertion has less gpr->simd traffic.
     if (!(Subtarget.hasBWI() ||
           (Subtarget.hasAVX512() && EltSizeInBits >= 32) ||
-          (Subtarget.hasSSE41() && VT.isFloatingPoint())))
+          (Subtarget.hasSSE41() && (EltVT == MVT::f32 || EltVT == MVT::f64))))
       return SDValue();
 
     MVT IdxSVT = MVT::getIntegerVT(EltSizeInBits);
     MVT IdxVT = MVT::getVectorVT(IdxSVT, NumElts);
     if (!isTypeLegal(IdxSVT) || !isTypeLegal(IdxVT))
       return SDValue();
 
     SDValue IdxExt = DAG.getZExtOrTrunc(N2, dl, IdxSVT);

llvm/test/CodeGen/X86/pr59980.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-apple-macosx10.15 | FileCheck %s
+
+%0 = type <{ half }>
+%1 = type <{ <16 x half> }>
+
+define void @foo(ptr %0, ptr %1, ptr %2) #0 {
+; CHECK-LABEL: foo:
+; CHECK:       ## %bb.0:
+; CHECK-NEXT:    pushq %rbp
+; CHECK-NEXT:    movq %rsp, %rbp
+; CHECK-NEXT:    andq $-32, %rsp
+; CHECK-NEXT:    subq $64, %rsp
+; CHECK-NEXT:    movl (%rdx), %eax
+; CHECK-NEXT:    andl $15, %eax
+; CHECK-NEXT:    vpinsrw $0, (%rdi), %xmm0, %xmm0
+; CHECK-NEXT:    vmovups (%rsi), %ymm1
+; CHECK-NEXT:    vmovaps %ymm1, (%rsp)
+; CHECK-NEXT:    vpextrw $0, %xmm0, (%rsp,%rax,2)
+; CHECK-NEXT:    vmovaps (%rsp), %ymm0
+; CHECK-NEXT:    vmovups %ymm0, (%rsi)
+; CHECK-NEXT:    movq %rbp, %rsp
+; CHECK-NEXT:    popq %rbp
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+  %4 = bitcast ptr %2 to ptr
+  %5 = load i64, ptr %4, align 8
+  %6 = getelementptr inbounds %0, ptr %0, i64 0, i32 0
+  %7 = load half, ptr %6, align 2
+  %8 = getelementptr inbounds %1, ptr %1, i64 0, i32 0
+  %9 = load <16 x half>, ptr %8, align 16
+  %10 = trunc i64 %5 to i32
+  %11 = insertelement <16 x half> %9, half %7, i32 %10
+  store <16 x half> %11, ptr %8, align 16
+  ret void
+}
+
+attributes #0 = { nounwind "target-features"="+f16c" }