mirror of
https://github.com/RPCS3/llvm-mirror.git
synced 2024-11-23 03:02:36 +01:00
f3ab4849fb
Now that we have fneg, prefer using it over "fsub -0.0, ...". This helps in particular with strict FP tests, as fneg does not raise any exceptions.
45 lines
1.2 KiB
LLVM
45 lines
1.2 KiB
LLVM
; Test negated floating-point absolute.
|
|
;
|
|
; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 | FileCheck %s
|
|
; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z13 | FileCheck %s
|
|
|
|
; Test f32.
|
|
declare float @llvm.fabs.f32(float %f)
|
|
define float @f1(float %f) {
|
|
; CHECK-LABEL: f1:
|
|
; CHECK: lndfr %f0, %f0
|
|
; CHECK: br %r14
|
|
%abs = call float @llvm.fabs.f32(float %f)
|
|
%res = fneg float %abs
|
|
ret float %res
|
|
}
|
|
|
|
; Test f64.
|
|
declare double @llvm.fabs.f64(double %f)
|
|
define double @f2(double %f) {
|
|
; CHECK-LABEL: f2:
|
|
; CHECK: lndfr %f0, %f0
|
|
; CHECK: br %r14
|
|
%abs = call double @llvm.fabs.f64(double %f)
|
|
%res = fneg double %abs
|
|
ret double %res
|
|
}
|
|
|
|
; Test f128. With the loads and stores, a pure negative-absolute would
|
|
; probably be better implemented using an OI on the upper byte. Do some
|
|
; extra processing so that using FPRs is unequivocally better.
|
|
declare fp128 @llvm.fabs.f128(fp128 %f)
|
|
define void @f3(fp128 *%ptr, fp128 *%ptr2) {
|
|
; CHECK-LABEL: f3:
|
|
; CHECK: lnxbr
|
|
; CHECK: dxbr
|
|
; CHECK: br %r14
|
|
%orig = load fp128, fp128 *%ptr
|
|
%abs = call fp128 @llvm.fabs.f128(fp128 %orig)
|
|
%negabs = fneg fp128 %abs
|
|
%op2 = load fp128, fp128 *%ptr2
|
|
%res = fdiv fp128 %negabs, %op2
|
|
store fp128 %res, fp128 *%ptr
|
|
ret void
|
|
}
|