|
120
|
1 ; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -amdgpu-promote-alloca < %s | FileCheck %s
|
|
|
2
|
|
|
3 ; CHECK-LABEL: @volatile_load(
|
|
121
|
4 ; CHECK: alloca [4 x i32]
|
|
|
5 ; CHECK: load volatile i32, i32*
|
|
|
6 define amdgpu_kernel void @volatile_load(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
|
|
120
|
7 entry:
|
|
121
|
8 %stack = alloca [4 x i32], align 4
|
|
120
|
9 %tmp = load i32, i32 addrspace(1)* %in, align 4
|
|
121
|
10 %arrayidx1 = getelementptr inbounds [4 x i32], [4 x i32]* %stack, i32 0, i32 %tmp
|
|
120
|
11 %load = load volatile i32, i32* %arrayidx1
|
|
|
12 store i32 %load, i32 addrspace(1)* %out
|
|
|
13 ret void
|
|
|
14 }
|
|
|
15
|
|
|
16 ; CHECK-LABEL: @volatile_store(
|
|
121
|
17 ; CHECK: alloca [4 x i32]
|
|
|
18 ; CHECK: store volatile i32 %tmp, i32*
|
|
|
19 define amdgpu_kernel void @volatile_store(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
|
|
120
|
20 entry:
|
|
121
|
21 %stack = alloca [4 x i32], align 4
|
|
120
|
22 %tmp = load i32, i32 addrspace(1)* %in, align 4
|
|
121
|
23 %arrayidx1 = getelementptr inbounds [4 x i32], [4 x i32]* %stack, i32 0, i32 %tmp
|
|
120
|
24 store volatile i32 %tmp, i32* %arrayidx1
|
|
|
25 ret void
|
|
|
26 }
|
|
|
27
|
|
|
28 ; Has on OK non-volatile user but also a volatile user
|
|
|
29 ; CHECK-LABEL: @volatile_and_non_volatile_load(
|
|
|
30 ; CHECK: alloca double
|
|
|
31 ; CHECK: load double
|
|
|
32 ; CHECK: load volatile double
|
|
121
|
33 define amdgpu_kernel void @volatile_and_non_volatile_load(double addrspace(1)* nocapture %arg, i32 %arg1) #0 {
|
|
120
|
34 bb:
|
|
|
35 %tmp = alloca double, align 8
|
|
|
36 store double 0.000000e+00, double* %tmp, align 8
|
|
|
37
|
|
|
38 %tmp4 = load double, double* %tmp, align 8
|
|
|
39 %tmp5 = load volatile double, double* %tmp, align 8
|
|
|
40
|
|
|
41 store double %tmp4, double addrspace(1)* %arg
|
|
|
42 ret void
|
|
|
43 }
|
|
|
44
|
|
|
45 attributes #0 = { nounwind }
|
