[X86][FP16] Do not create VBROADCAST_LOAD for f16 without AVX2 (#91125)

AVX doesn't provide 16-bit BROADCAST instruction.

Fixes #91005

Author: phoebewang

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -7295,7 +7295,7 @@ static SDValue lowerBuildVectorAsBroadcast(BuildVectorSDNode *BVOp,
     // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
     if (ScalarSize == 32 ||
         (ScalarSize == 64 && (IsGE256 || Subtarget.hasVLX())) ||
-        CVT == MVT::f16 ||
+        (CVT == MVT::f16 && Subtarget.hasAVX2()) ||
         (OptForSize && (ScalarSize == 64 || Subtarget.hasAVX2()))) {
       const Constant *C = nullptr;
       if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))

llvm/test/CodeGen/X86/pr91005.ll

@@ -0,0 +1,40 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=+f16c < %s | FileCheck %s
+
+define void @PR91005(ptr %0) minsize {
+; CHECK-LABEL: PR91005:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    xorl %eax, %eax
+; CHECK-NEXT:    testb %al, %al
+; CHECK-NEXT:    je .LBB0_2
+; CHECK-NEXT:  # %bb.1:
+; CHECK-NEXT:    vpcmpeqw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; CHECK-NEXT:    vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; CHECK-NEXT:    vpextrw $0, %xmm0, %eax
+; CHECK-NEXT:    movzwl %ax, %eax
+; CHECK-NEXT:    vmovd %eax, %xmm0
+; CHECK-NEXT:    vcvtph2ps %xmm0, %xmm0
+; CHECK-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; CHECK-NEXT:    vmulss %xmm1, %xmm0, %xmm0
+; CHECK-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
+; CHECK-NEXT:    vmovd %xmm0, %eax
+; CHECK-NEXT:    movw %ax, (%rdi)
+; CHECK-NEXT:  .LBB0_2: # %common.ret
+; CHECK-NEXT:    retq
+  %2 = bitcast <2 x half> poison to <2 x i16>
+  %3 = icmp eq <2 x i16> %2, <i16 31744, i16 31744>
+  br i1 poison, label %4, label %common.ret
+
+common.ret:                                       ; preds = %4, %1
+  ret void
+
+4:                                                ; preds = %1
+  %5 = select <2 x i1> %3, <2 x half> <half 0xH3C00, half 0xH3C00>, <2 x half> zeroinitializer
+  %6 = fmul <2 x half> %5, zeroinitializer
+  %7 = fsub <2 x half> %6, zeroinitializer
+  %8 = extractelement <2 x half> %7, i64 0
+  store half %8, ptr %0, align 2
+  br label %common.ret
+}
+
+declare <2 x half> @llvm.fabs.v2f16(<2 x half>)