1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
|
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN %s
; The type promotion for the vector loads v3i32/v3f32 into v4i32/v4f32 is enabled
; only when the alignment is 8-byte or higher.
; Otherwise, split the load into two separate loads (dwordx2 + dword).
; This type promotion on smaller aligned loads can cause a page fault error
; while accessing one extra dword beyond the buffer.
define protected amdgpu_kernel void @load_v3i32_align4(ptr addrspace(1) %arg) #0 {
; GCN-LABEL: load_v3i32_align4:
; GCN: ; %bb.0:
; GCN: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_load_dwordx2 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x0
; GCN-NEXT: s_load_dword s{{[0-9]+}}, s[0:1], 0x8
%vec = load <3 x i32>, ptr addrspace(1) %arg, align 4
store <3 x i32> %vec, ptr addrspace(1) undef, align 4
ret void
}
define protected amdgpu_kernel void @load_v3i32_align8(ptr addrspace(1) %arg) #0 {
; GCN-LABEL: load_v3i32_align8:
; GCN: ; %bb.0:
; GCN: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_load_dwordx4 s{{\[[0-9]:[0-9]+\]}}, s[4:5], 0x0
%vec = load <3 x i32>, ptr addrspace(1) %arg, align 8
store <3 x i32> %vec, ptr addrspace(1) undef, align 8
ret void
}
define protected amdgpu_kernel void @load_v3i32_align16(ptr addrspace(1) %arg) #0 {
; GCN-LABEL: load_v3i32_align16:
; GCN: ; %bb.0:
; GCN: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_load_dwordx4 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x0
%vec = load <3 x i32>, ptr addrspace(1) %arg, align 16
store <3 x i32> %vec, ptr addrspace(1) undef, align 16
ret void
}
define protected amdgpu_kernel void @load_v3f32_align4(ptr addrspace(1) %arg) #0 {
; GCN-LABEL: load_v3f32_align4:
; GCN: ; %bb.0:
; GCN: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_load_dwordx2 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x0
; GCN-NEXT: s_load_dword s{{[0-9]+}}, s[0:1], 0x8
%vec = load <3 x float>, ptr addrspace(1) %arg, align 4
store <3 x float> %vec, ptr addrspace(1) undef, align 4
ret void
}
define protected amdgpu_kernel void @load_v3f32_align8(ptr addrspace(1) %arg) #0 {
; GCN-LABEL: load_v3f32_align8:
; GCN: ; %bb.0:
; GCN: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_load_dwordx4 s{{\[[0-9]:[0-9]+\]}}, s[4:5], 0x0
%vec = load <3 x float>, ptr addrspace(1) %arg, align 8
store <3 x float> %vec, ptr addrspace(1) undef, align 8
ret void
}
define protected amdgpu_kernel void @load_v3f32_align16(ptr addrspace(1) %arg) #0 {
; GCN-LABEL: load_v3f32_align16:
; GCN: ; %bb.0:
; GCN: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_load_dwordx4 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x0
%vec = load <3 x float>, ptr addrspace(1) %arg, align 16
store <3 x float> %vec, ptr addrspace(1) undef, align 16
ret void
}
attributes #0 = { nounwind noinline }
|
