72 lines
2.6 KiB
LLVM
72 lines
2.6 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
; RUN: llc -mtriple=x86_64-apple-macosx %s -o - | FileCheck %s --check-prefix=CHECK-SOFT
|
|
; RUN: llc -mtriple=x86_64-apple-macosx -mattr=+f16c %s -o - | FileCheck %s --check-prefix=CHECK-F16C
|
|
|
|
define void @extendhfxf(ptr %outptr, ptr %inptr) nounwind {
|
|
; CHECK-SOFT-LABEL: extendhfxf:
|
|
; CHECK-SOFT: ## %bb.0:
|
|
; CHECK-SOFT-NEXT: pushq %rbx
|
|
; CHECK-SOFT-NEXT: subq $16, %rsp
|
|
; CHECK-SOFT-NEXT: movq %rdi, %rbx
|
|
; CHECK-SOFT-NEXT: movzwl (%rsi), %edi
|
|
; CHECK-SOFT-NEXT: callq ___extendhfsf2
|
|
; CHECK-SOFT-NEXT: movss %xmm0, {{[0-9]+}}(%rsp)
|
|
; CHECK-SOFT-NEXT: flds {{[0-9]+}}(%rsp)
|
|
; CHECK-SOFT-NEXT: fstpt (%rbx)
|
|
; CHECK-SOFT-NEXT: addq $16, %rsp
|
|
; CHECK-SOFT-NEXT: popq %rbx
|
|
; CHECK-SOFT-NEXT: retq
|
|
;
|
|
; CHECK-F16C-LABEL: extendhfxf:
|
|
; CHECK-F16C: ## %bb.0:
|
|
; CHECK-F16C-NEXT: movzwl (%rsi), %eax
|
|
; CHECK-F16C-NEXT: vmovd %eax, %xmm0
|
|
; CHECK-F16C-NEXT: vcvtph2ps %xmm0, %xmm0
|
|
; CHECK-F16C-NEXT: vmovss %xmm0, -{{[0-9]+}}(%rsp)
|
|
; CHECK-F16C-NEXT: flds -{{[0-9]+}}(%rsp)
|
|
; CHECK-F16C-NEXT: fstpt (%rdi)
|
|
; CHECK-F16C-NEXT: retq
|
|
%in = load half, ptr %inptr
|
|
%fp80 = fpext half %in to x86_fp80
|
|
store x86_fp80 %fp80, ptr %outptr
|
|
ret void
|
|
}
|
|
|
|
; FIXME: We don't currently provide __truncxfhf2, but we can't lower this as
|
|
; successive fptruncs (like we do fpext) because of double rounding.
|
|
; We also don't currently soft-float this call, like we do e.g., __truncsfhf2:
|
|
; the latter long predates the fp16 parameter passing ABI, so can't change.
|
|
; If we ever add a __truncxfhf2, we're not bound by existing ABI.
|
|
define void @truncxfhf(ptr %outptr, ptr %inptr) nounwind {
|
|
; CHECK-SOFT-LABEL: truncxfhf:
|
|
; CHECK-SOFT: ## %bb.0:
|
|
; CHECK-SOFT-NEXT: pushq %rbx
|
|
; CHECK-SOFT-NEXT: subq $16, %rsp
|
|
; CHECK-SOFT-NEXT: movq %rdi, %rbx
|
|
; CHECK-SOFT-NEXT: fldt (%rsi)
|
|
; CHECK-SOFT-NEXT: fstpt (%rsp)
|
|
; CHECK-SOFT-NEXT: callq ___truncxfhf2
|
|
; CHECK-SOFT-NEXT: pextrw $0, %xmm0, %eax
|
|
; CHECK-SOFT-NEXT: movw %ax, (%rbx)
|
|
; CHECK-SOFT-NEXT: addq $16, %rsp
|
|
; CHECK-SOFT-NEXT: popq %rbx
|
|
; CHECK-SOFT-NEXT: retq
|
|
;
|
|
; CHECK-F16C-LABEL: truncxfhf:
|
|
; CHECK-F16C: ## %bb.0:
|
|
; CHECK-F16C-NEXT: pushq %rbx
|
|
; CHECK-F16C-NEXT: subq $16, %rsp
|
|
; CHECK-F16C-NEXT: movq %rdi, %rbx
|
|
; CHECK-F16C-NEXT: fldt (%rsi)
|
|
; CHECK-F16C-NEXT: fstpt (%rsp)
|
|
; CHECK-F16C-NEXT: callq ___truncxfhf2
|
|
; CHECK-F16C-NEXT: vpextrw $0, %xmm0, (%rbx)
|
|
; CHECK-F16C-NEXT: addq $16, %rsp
|
|
; CHECK-F16C-NEXT: popq %rbx
|
|
; CHECK-F16C-NEXT: retq
|
|
%in = load x86_fp80, ptr %inptr
|
|
%half = fptrunc x86_fp80 %in to half
|
|
store half %half, ptr %outptr
|
|
ret void
|
|
}
|