diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -20222,7 +20222,7 @@ // 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); diff --git a/llvm/test/CodeGen/X86/pr59980.ll b/llvm/test/CodeGen/X86/pr59980.ll new file mode 100644 --- /dev/null +++ b/llvm/test/CodeGen/X86/pr59980.ll @@ -0,0 +1,39 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py +; RUN: llc < %s -mtriple=x86_64-apple-macosx10.15 | FileCheck %s + +%Ts7Float16V = type <{ half }> +%Ts7Float16V13SIMD16StorageV = type <{ <16 x half> }> + +define swiftcc void @"$ss7Float16V13SIMD16StorageVyABSicipADTk"(%Ts7Float16V* noalias nocapture dereferenceable(2) %0, %Ts7Float16V13SIMD16StorageV* nocapture dereferenceable(32) %1, i8* %2) #0 { +; CHECK-LABEL: $ss7Float16V13SIMD16StorageVyABSicipADTk: +; CHECK: ## %bb.0: ## %entry +; 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 +entry: + %._value = bitcast i8* %2 to i64* + %3 = load i64, i64* %._value, align 8 + %._value1 = getelementptr inbounds %Ts7Float16V, %Ts7Float16V* %0, i64 0, i32 0 + %4 = load half, half* %._value1, align 2 + %._value2 = getelementptr inbounds %Ts7Float16V13SIMD16StorageV, %Ts7Float16V13SIMD16StorageV* %1, i64 0, i32 0 + %5 = load <16 x half>, <16 x half>* %._value2, align 16 + %6 = trunc i64 %3 to i32 + %7 = insertelement <16 x half> %5, half %4, i32 %6 + store <16 x half> %7, <16 x half>* %._value2, align 16 + ret void +} + +attributes #0 = { nounwind "target-features"="+f16c" }