llvm/test/CodeGen/NVPTX/fast-math.ll
Justin Holewinski 3a8ee4ffd7 [NVPTX] Use approximate FP ops when unsafe-fp-math is used, and append
.ftz to instructions if the nvptx-f32ftz attribute is set to "true"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@186820 91177308-0d34-0410-b5e6-96231b3b80d8
2013-07-22 12:18:04 +00:00

44 lines
866 B
LLVM

; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
declare float @llvm.nvvm.sqrt.f(float)
; CHECK: sqrt_div
; CHECK: sqrt.rn.f32
; CHECK: div.rn.f32
define float @sqrt_div(float %a, float %b) {
%t1 = tail call float @llvm.nvvm.sqrt.f(float %a)
%t2 = fdiv float %t1, %b
ret float %t2
}
; CHECK: sqrt_div_fast
; CHECK: sqrt.approx.f32
; CHECK: div.approx.f32
define float @sqrt_div_fast(float %a, float %b) #0 {
%t1 = tail call float @llvm.nvvm.sqrt.f(float %a)
%t2 = fdiv float %t1, %b
ret float %t2
}
; CHECK: fadd
; CHECK: add.f32
define float @fadd(float %a, float %b) {
%t1 = fadd float %a, %b
ret float %t1
}
; CHECK: fadd_ftz
; CHECK: add.ftz.f32
define float @fadd_ftz(float %a, float %b) #1 {
%t1 = fadd float %a, %b
ret float %t1
}
attributes #0 = { "unsafe-fp-math" = "true" }
attributes #1 = { "nvptx-f32ftz" = "true" }