deb84db5b4
We need to apply DXIL op attributes to the functions themselves, and all DXIL ops should have the `unwind` attribute. This matches the DXC behaviour and what consumers like warp's GPU-based validation expect. Fixes #193620
37 lines
1.4 KiB
LLVM
37 lines
1.4 KiB
LLVM
; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
|
|
|
|
; Make sure dxil operation function calls for fmax are generated for half/float/double.
|
|
|
|
; CHECK-LABEL:test_fmax_half
|
|
define noundef half @test_fmax_half(half noundef %a, half noundef %b) {
|
|
entry:
|
|
; CHECK: call half @dx.op.binary.f16(i32 35, half %{{.*}}, half %{{.*}})
|
|
%0 = call half @llvm.maxnum.f16(half %a, half %b)
|
|
ret half %0
|
|
}
|
|
|
|
; CHECK-LABEL:test_fmax_float
|
|
define noundef float @test_fmax_float(float noundef %a, float noundef %b) {
|
|
entry:
|
|
; CHECK: call float @dx.op.binary.f32(i32 35, float %{{.*}}, float %{{.*}})
|
|
%0 = call float @llvm.maxnum.f32(float %a, float %b)
|
|
ret float %0
|
|
}
|
|
|
|
; CHECK-LABEL:test_fmax_double
|
|
define noundef double @test_fmax_double(double noundef %a, double noundef %b) {
|
|
entry:
|
|
; CHECK: call double @dx.op.binary.f64(i32 35, double %{{.*}}, double %{{.*}})
|
|
%0 = call double @llvm.maxnum.f64(double %a, double %b)
|
|
ret double %0
|
|
}
|
|
|
|
; CHECK-DAG: declare half @dx.op.binary.f16(i32, half, half) #[[#ATTR0:]]
|
|
; CHECK-DAG: declare float @dx.op.binary.f32(i32, float, float) #[[#ATTR0]]
|
|
; CHECK-DAG: declare double @dx.op.binary.f64(i32, double, double) #[[#ATTR0]]
|
|
; CHECK: attributes #[[#ATTR0]] = { nounwind memory(none) }
|
|
|
|
declare half @llvm.maxnum.f16(half, half)
|
|
declare float @llvm.maxnum.f32(float, float)
|
|
declare double @llvm.maxnum.f64(double, double)
|