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
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
|
# RUN: llc -march=amdgcn -verify-machineinstrs -run-pass si-load-store-opt -o - %s | FileCheck %s
# Check that SILoadStoreOptimizer honors memory dependencies between moved
# instructions.
#
# The following IR snippet would usually be optimized by the peephole optimizer.
# However, an equivalent situation can occur with buffer instructions as well.
# CHECK-LABEL: name: mem_dependency
# CHECK: DS_READ2_B32 %0, 0, 1,
# CHECK: DS_WRITE_B32 %0, killed %1, 64,
# CHECK: DS_READ2_B32 %0, 16, 17,
# CHECK: DS_WRITE_B32 killed %0, %5, 0
--- |
define amdgpu_kernel void @mem_dependency(i32 addrspace(3)* %ptr.0) nounwind {
%ptr.4 = getelementptr i32, i32 addrspace(3)* %ptr.0, i32 1
%ptr.64 = getelementptr i32, i32 addrspace(3)* %ptr.0, i32 16
%1 = load i32, i32 addrspace(3)* %ptr.0
store i32 %1, i32 addrspace(3)* %ptr.64
%2 = load i32, i32 addrspace(3)* %ptr.64
%3 = load i32, i32 addrspace(3)* %ptr.4
%4 = add i32 %2, %3
store i32 %4, i32 addrspace(3)* %ptr.0
ret void
}
@lds0 = external dso_local unnamed_addr addrspace(3) global [256 x i32], align 4
@lds1 = external dso_local unnamed_addr addrspace(3) global [256 x i32], align 4
@lds2 = external dso_local unnamed_addr addrspace(3) global [256 x i32], align 4
@lds3 = external dso_local unnamed_addr addrspace(3) global [256 x i32], align 4
define void @asm_defines_address() #0 {
bb:
%tmp1 = load i32, i32 addrspace(3)* getelementptr inbounds ([256 x i32], [256 x i32] addrspace(3)* @lds0, i32 0, i32 0), align 4
%0 = and i32 %tmp1, 255
%tmp3 = load i32, i32 addrspace(3)* getelementptr ([256 x i32], [256 x i32] addrspace(3)* @lds1, i32 0, i32 undef), align 4
%tmp6 = load i32, i32 addrspace(3)* getelementptr ([256 x i32], [256 x i32] addrspace(3)* @lds3, i32 0, i32 undef), align 4
%tmp7 = tail call i32 asm "v_or_b32 $0, 0, $1", "=v,v"(i32 %tmp6) #1
%tmp10 = lshr i32 %tmp7, 16
%tmp11 = and i32 %tmp10, 255
%tmp12 = getelementptr inbounds [256 x i32], [256 x i32] addrspace(3)* @lds1, i32 0, i32 %tmp11
%tmp13 = load i32, i32 addrspace(3)* %tmp12, align 4
%tmp14 = xor i32 %tmp3, %tmp13
%tmp15 = lshr i32 %tmp14, 8
%tmp16 = and i32 %tmp15, 16711680
%tmp19 = lshr i32 %tmp16, 16
%tmp20 = and i32 %tmp19, 255
%tmp21 = getelementptr inbounds [256 x i32], [256 x i32] addrspace(3)* @lds1, i32 0, i32 %tmp20
%tmp22 = load i32, i32 addrspace(3)* %tmp21, align 4
%tmp24 = load i32, i32 addrspace(3)* getelementptr ([256 x i32], [256 x i32] addrspace(3)* @lds2, i32 0, i32 undef), align 4
%tmp25 = xor i32 %tmp22, %tmp24
%tmp26 = and i32 %tmp25, -16777216
%tmp28 = or i32 %0, %tmp26
store volatile i32 %tmp28, i32 addrspace(1)* undef
ret void
}
attributes #0 = { convergent nounwind }
attributes #1 = { convergent nounwind readnone }
define amdgpu_kernel void @move_waw_hazards() #0 {
ret void
}
attributes #0 = { convergent nounwind }
define amdgpu_kernel void @merge_mmos(i32 addrspace(1)* %ptr_addr1) { ret void }
...
---
name: mem_dependency
alignment: 0
exposesReturnsTwice: false
legalized: false
regBankSelected: false
selected: false
tracksRegLiveness: true
liveins:
- { reg: '$vgpr0', virtual-reg: '%1' }
frameInfo:
isFrameAddressTaken: false
isReturnAddressTaken: false
hasStackMap: false
hasPatchPoint: false
stackSize: 0
offsetAdjustment: 0
maxAlignment: 0
adjustsStack: false
hasCalls: false
maxCallFrameSize: 0
hasOpaqueSPAdjustment: false
hasVAStart: false
hasMustTailInVarArgFunc: false
body: |
bb.0:
liveins: $vgpr0
%1:vgpr_32 = COPY $vgpr0
$m0 = S_MOV_B32 -1
%2:vgpr_32 = DS_READ_B32 %1, 0, 0, implicit $m0, implicit $exec :: (load 4 from %ir.ptr.0)
DS_WRITE_B32 %1, killed %2, 64, 0, implicit $m0, implicit $exec :: (store 4 into %ir.ptr.64)
; Make this load unmergeable, to tempt SILoadStoreOptimizer into merging the
; other two loads.
%6:vreg_64 = DS_READ2_B32 %1, 16, 17, 0, implicit $m0, implicit $exec :: (load 8 from %ir.ptr.64, align 4)
%3:vgpr_32 = COPY %6.sub0
%4:vgpr_32 = DS_READ_B32 %1, 4, 0, implicit $m0, implicit $exec :: (load 4 from %ir.ptr.4)
%5:vgpr_32 = V_ADD_I32_e32 killed %3, killed %4, implicit-def $vcc, implicit $exec
DS_WRITE_B32 killed %1, %5, 0, 0, implicit killed $m0, implicit $exec :: (store 4 into %ir.ptr.0)
S_ENDPGM 0
...
---
# Make sure the asm def isn't moved after the point where it's used for
# the address.
# CHECK-LABEL: name: asm_defines_address
# CHECK: DS_READ2ST64_B32
# CHECK: DS_READ2ST64_B32
# CHECK: INLINEASM
# CHECK: DS_READ_B32
# CHECK: DS_READ_B32
name: asm_defines_address
tracksRegLiveness: true
registers:
- { id: 0, class: vgpr_32, preferred-register: '' }
body: |
bb.0:
%1:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
%2:vgpr_32 = DS_READ_B32 %1, 3072, 0, implicit $m0, implicit $exec :: (dereferenceable load 4 from `i32 addrspace(3)* getelementptr inbounds ([256 x i32], [256 x i32] addrspace(3)* @lds0, i32 0, i32 0)`, addrspace 3)
%3:vgpr_32 = DS_READ_B32 %1, 2048, 0, implicit $m0, implicit $exec :: (load 4 from `i32 addrspace(3)* getelementptr ([256 x i32], [256 x i32] addrspace(3)* @lds1, i32 0, i32 undef)`, addrspace 3)
%4:vgpr_32 = DS_READ_B32 %1, 1024, 0, implicit $m0, implicit $exec :: (load 4 from `i32 addrspace(3)* getelementptr ([256 x i32], [256 x i32] addrspace(3)* @lds3, i32 0, i32 undef)`, addrspace 3)
INLINEASM &"v_or_b32 $0, 0, $1", 32, 327690, def %0, 327689, %4
%5:vgpr_32 = DS_READ_B32 %0, 2048, 0, implicit $m0, implicit $exec :: (load 4 from %ir.tmp12, addrspace 3)
%6:vgpr_32 = DS_READ_B32 %5, 2048, 0, implicit $m0, implicit $exec :: (load 4 from %ir.tmp21, addrspace 3)
%7:vgpr_32 = DS_READ_B32 %1, 0, 0, implicit $m0, implicit $exec :: (load 4 from `i32 addrspace(3)* getelementptr ([256 x i32], [256 x i32] addrspace(3)* @lds2, i32 0, i32 undef)`, addrspace 3)
S_SETPC_B64_return undef $sgpr30_sgpr31, implicit %6, implicit %7
...
---
# Make sure Write-after-Write hazards are correctly detected and the
# instructions moved accordingly.
# operations.
# CHECK-LABEL: name: move_waw_hazards
# CHECK: S_AND_B64
# CHECK: S_CMP_EQ_U32
name: move_waw_hazards
tracksRegLiveness: true
body: |
bb.0:
liveins: $sgpr0_sgpr1
%3:sgpr_64 = COPY $sgpr0_sgpr1
%6:sreg_32_xm0_xexec = S_MOV_B32 0
%7:sreg_32_xm0 = S_MOV_B32 0
%8:sreg_64_xexec = REG_SEQUENCE killed %6, %subreg.sub0, %7, %subreg.sub1
%9:sreg_128 = S_LOAD_DWORDX4_IMM killed %8, 0, 0, 0 :: (invariant load 16, addrspace 6)
%31:sreg_64_xexec = S_BUFFER_LOAD_DWORDX2_IMM %9, 0, 0, 0 :: (dereferenceable invariant load 4)
%10:sreg_32_xm0_xexec = COPY %31.sub0
%11:sreg_32_xm0_xexec = COPY killed %31.sub1
%12:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %9, 2, 0, 0 :: (dereferenceable invariant load 4)
%13:sreg_64 = V_CMP_NE_U32_e64 killed %11, 0, implicit $exec
%15:sreg_64 = V_CMP_NE_U32_e64 killed %12, 0, implicit $exec
%17:sreg_64_xexec = S_AND_B64 killed %13, killed %15, implicit-def dead $scc
S_CMP_EQ_U32 killed %10, 0, implicit-def $scc
%18:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %9, 3, 0, 0 :: (dereferenceable invariant load 4)
S_ENDPGM 0
...
---
# CHECK-LABEL: merge_mmos
# CHECK: S_BUFFER_LOAD_DWORDX2_IMM %0, 0, 0, 0 :: (dereferenceable invariant load 8, align 4)
# CHECK: BUFFER_LOAD_DWORDX2_OFFSET %0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8, align 4)
# CHECK: BUFFER_STORE_DWORDX2_OFFSET_exact killed %{{[0-9]+}}, %0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8, align 4)
# CHECK: BUFFER_LOAD_DWORDX2_OFFSET %0, 0, 64, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8 from %ir.ptr_addr1 + 64, align 4
# CHECK: BUFFER_STORE_DWORDX2_OFFSET_exact killed %{{[0-9]+}}, %0, 0, 64, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into %ir.ptr_addr1 + 64, align 4
name: merge_mmos
tracksRegLiveness: true
body: |
bb.0:
liveins: $sgpr0_sgpr1_sgpr2_sgpr3
%0:sreg_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
%1:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0, 0, 0, 0 :: (dereferenceable invariant load 4)
%2:sreg_32_xm0_xexec = S_BUFFER_LOAD_DWORD_IMM %0, 1, 0, 0 :: (dereferenceable invariant load 4)
%3:vgpr_32 = BUFFER_LOAD_DWORD_OFFSET %0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4)
%4:vgpr_32 = BUFFER_LOAD_DWORD_OFFSET %0, 0, 4, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4)
BUFFER_STORE_DWORD_OFFSET_exact %3, %0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4)
BUFFER_STORE_DWORD_OFFSET_exact %4, %0, 0, 4, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4)
%5:vgpr_32 = BUFFER_LOAD_DWORD_OFFSET %0, 0, 64, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from %ir.ptr_addr1 + 64)
%6:vgpr_32 = BUFFER_LOAD_DWORD_OFFSET %0, 0, 68, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from %ir.ptr_addr1 + 68)
BUFFER_STORE_DWORD_OFFSET_exact %5, %0, 0, 64, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.ptr_addr1 + 64)
BUFFER_STORE_DWORD_OFFSET_exact %6, %0, 0, 68, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.ptr_addr1 + 68)
S_ENDPGM 0
...
|
