1 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
3 ; Check that we properly realign the stack. While 4-byte access is all
4 ; that is ever needed, some transformations rely on the known bits from the alignment of the pointer (e.g.
8 ; 4 byte emergency stack slot
9 ; = 144 bytes with padding between them
11 ; GCN-LABEL: {{^}}needs_align16_default_stack_align:
12 ; GCN-DAG: v_lshlrev_b32_e32 [[SCALED_IDX:v[0-9]+]], 4, v0
13 ; GCN-DAG: v_lshrrev_b32_e64 [[FRAMEDIFF:v[0-9]+]], 6, s32
14 ; GCN: v_add_u32_e32 [[FI:v[0-9]+]], vcc, [[FRAMEDIFF]], [[SCALED_IDX]]
18 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
19 ; GCN: v_or_b32_e32 v{{[0-9]+}}, 12
20 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
21 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
22 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
26 ; GCN: ; ScratchSize: 144
27 define void @needs_align16_default_stack_align(i32 %idx) #0 {
28 %alloca.align16 = alloca [8 x <4 x i32>], align 16, addrspace(5)
29 %gep0 = getelementptr inbounds [8 x <4 x i32>], [8 x <4 x i32>] addrspace(5)* %alloca.align16, i32 0, i32 %idx
30 store volatile <4 x i32> <i32 1, i32 2, i32 3, i32 4>, <4 x i32> addrspace(5)* %gep0, align 16
34 ; GCN-LABEL: {{^}}needs_align16_stack_align4:
35 ; GCN: s_add_i32 [[SCRATCH_REG:s[0-9]+]], s32, 0x3c0{{$}}
36 ; GCN: s_and_b32 s33, [[SCRATCH_REG]], 0xfffffc00
38 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
39 ; GCN: v_or_b32_e32 v{{[0-9]+}}, 12
40 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
41 ; GCN: s_addk_i32 s32, 0x2800{{$}}
42 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
43 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
45 ; GCN: s_addk_i32 s32, 0xd800
47 ; GCN: ; ScratchSize: 160
48 define void @needs_align16_stack_align4(i32 %idx) #2 {
49 %alloca.align16 = alloca [8 x <4 x i32>], align 16, addrspace(5)
50 %gep0 = getelementptr inbounds [8 x <4 x i32>], [8 x <4 x i32>] addrspace(5)* %alloca.align16, i32 0, i32 %idx
51 store volatile <4 x i32> <i32 1, i32 2, i32 3, i32 4>, <4 x i32> addrspace(5)* %gep0, align 16
55 ; GCN-LABEL: {{^}}needs_align32:
56 ; GCN: s_add_i32 [[SCRATCH_REG:s[0-9]+]], s32, 0x7c0{{$}}
57 ; GCN: s_and_b32 s33, [[SCRATCH_REG]], 0xfffff800
59 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
60 ; GCN: v_or_b32_e32 v{{[0-9]+}}, 12
61 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
62 ; GCN: s_addk_i32 s32, 0x3000{{$}}
63 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
64 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
66 ; GCN: s_addk_i32 s32, 0xd000
68 ; GCN: ; ScratchSize: 192
69 define void @needs_align32(i32 %idx) #0 {
70 %alloca.align16 = alloca [8 x <4 x i32>], align 32, addrspace(5)
71 %gep0 = getelementptr inbounds [8 x <4 x i32>], [8 x <4 x i32>] addrspace(5)* %alloca.align16, i32 0, i32 %idx
72 store volatile <4 x i32> <i32 1, i32 2, i32 3, i32 4>, <4 x i32> addrspace(5)* %gep0, align 32
76 ; GCN-LABEL: {{^}}force_realign4:
77 ; GCN: s_add_i32 [[SCRATCH_REG:s[0-9]+]], s32, 0xc0{{$}}
78 ; GCN: s_and_b32 s33, [[SCRATCH_REG]], 0xffffff00
79 ; GCN: s_addk_i32 s32, 0xd00{{$}}
81 ; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], 0 offen
82 ; GCN: s_addk_i32 s32, 0xf300
84 ; GCN: ; ScratchSize: 52
85 define void @force_realign4(i32 %idx) #1 {
86 %alloca.align16 = alloca [8 x i32], align 4, addrspace(5)
87 %gep0 = getelementptr inbounds [8 x i32], [8 x i32] addrspace(5)* %alloca.align16, i32 0, i32 %idx
88 store volatile i32 3, i32 addrspace(5)* %gep0, align 4
92 ; GCN-LABEL: {{^}}kernel_call_align16_from_8:
93 ; GCN: s_movk_i32 s32, 0x400{{$}}
96 define amdgpu_kernel void @kernel_call_align16_from_8() #0 {
97 %alloca = alloca i32, align 4, addrspace(5)
98 store volatile i32 2, i32 addrspace(5)* %alloca
99 call void @needs_align16_default_stack_align(i32 1)
103 ; The call sequence should keep the stack on call aligned to 4
104 ; GCN-LABEL: {{^}}kernel_call_align16_from_5:
105 ; GCN: s_movk_i32 s32, 0x400
107 define amdgpu_kernel void @kernel_call_align16_from_5() {
108 %alloca0 = alloca i8, align 1, addrspace(5)
109 store volatile i8 2, i8 addrspace(5)* %alloca0
111 call void @needs_align16_default_stack_align(i32 1)
115 ; GCN-LABEL: {{^}}kernel_call_align4_from_5:
116 ; GCN: s_movk_i32 s32, 0x400
118 define amdgpu_kernel void @kernel_call_align4_from_5() {
119 %alloca0 = alloca i8, align 1, addrspace(5)
120 store volatile i8 2, i8 addrspace(5)* %alloca0
122 call void @needs_align16_stack_align4(i32 1)
126 ; GCN-LABEL: {{^}}default_realign_align128:
127 ; GCN: s_mov_b32 [[FP_COPY:s[0-9]+]], s33
128 ; GCN-NEXT: s_add_i32 s33, s32, 0x1fc0
129 ; GCN-NEXT: s_and_b32 s33, s33, 0xffffe000
130 ; GCN-NEXT: s_addk_i32 s32, 0x4000
132 ; GCN: buffer_store_dword v0, off, s[0:3], s33{{$}}
133 ; GCN: s_addk_i32 s32, 0xc000
134 ; GCN: s_mov_b32 s33, [[FP_COPY]]
135 define void @default_realign_align128(i32 %idx) #0 {
136 %alloca.align = alloca i32, align 128, addrspace(5)
137 store volatile i32 9, i32 addrspace(5)* %alloca.align, align 128
141 ; GCN-LABEL: {{^}}disable_realign_align128:
143 ; GCN: buffer_store_dword v0, off, s[0:3], s32{{$}}
145 define void @disable_realign_align128(i32 %idx) #3 {
146 %alloca.align = alloca i32, align 128, addrspace(5)
147 store volatile i32 9, i32 addrspace(5)* %alloca.align, align 128
151 declare void @extern_func(<32 x i32>, i32) #0
152 define void @func_call_align1024_bp_gets_vgpr_spill(<32 x i32> %a, i32 %b) #0 {
153 ; The test forces the stack to be realigned to a new boundary
154 ; since there is a local object with an alignment of 1024.
155 ; Should use BP to access the incoming stack arguments.
156 ; The BP value is saved/restored with a VGPR spill.
158 ; GCN-LABEL: func_call_align1024_bp_gets_vgpr_spill:
159 ; GCN: buffer_store_dword [[VGPR_REG:v[0-9]+]], off, s[0:3], s32 offset:1028 ; 4-byte Folded Spill
160 ; GCN-NEXT: s_mov_b64 exec, s[16:17]
161 ; GCN-NEXT: v_writelane_b32 [[VGPR_REG]], s33, 2
162 ; GCN-DAG: s_add_i32 [[SCRATCH_REG:s[0-9]+]], s32, 0xffc0
163 ; GCN: s_and_b32 s33, [[SCRATCH_REG]], 0xffff0000
164 ; GCN: v_mov_b32_e32 v32, 0
165 ; GCN-DAG: v_writelane_b32 [[VGPR_REG]], s34, 3
166 ; GCN: s_mov_b32 s34, s32
167 ; GCN: buffer_store_dword v32, off, s[0:3], s33 offset:1024
168 ; GCN-NEXT: s_waitcnt vmcnt(0)
169 ; GCN-NEXT: buffer_load_dword v{{[0-9]+}}, off, s[0:3], s34
170 ; GCN-DAG: s_add_i32 s32, s32, 0x30000
171 ; GCN: buffer_store_dword v{{[0-9]+}}, off, s[0:3], s32
172 ; GCN: s_swappc_b64 s[30:31],
174 ; GCN: v_readlane_b32 s31, [[VGPR_REG]], 1
175 ; GCN: v_readlane_b32 s30, [[VGPR_REG]], 0
176 ; GCN: s_add_i32 s32, s32, 0xfffd0000
177 ; GCN-NEXT: v_readlane_b32 s33, [[VGPR_REG]], 2
178 ; GCN-NEXT: v_readlane_b32 s34, [[VGPR_REG]], 3
179 ; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
180 ; GCN-NEXT: buffer_load_dword [[VGPR_REG]], off, s[0:3], s32 offset:1028 ; 4-byte Folded Reload
181 ; GCN-NEXT: s_mov_b64 exec, s[4:5]
182 ; GCN: s_setpc_b64 s[30:31]
183 %temp = alloca i32, align 1024, addrspace(5)
184 store volatile i32 0, i32 addrspace(5)* %temp, align 1024
185 call void @extern_func(<32 x i32> %a, i32 %b)
189 %struct.Data = type { [9 x i32] }
190 define i32 @needs_align1024_stack_args_used_inside_loop(%struct.Data addrspace(5)* nocapture readonly byval(%struct.Data) align 8 %arg) local_unnamed_addr #4 {
191 ; The local object allocation needed an alignment of 1024.
192 ; Since the function argument is accessed in a loop with an
193 ; index variable, the base pointer first get loaded into a VGPR
194 ; and that value should be further referenced to load the incoming values.
195 ; The BP value will get saved/restored in an SGPR at the prolgoue/epilogue.
197 ; GCN-LABEL: needs_align1024_stack_args_used_inside_loop:
198 ; GCN: s_mov_b32 [[FP_COPY:s[0-9]+]], s33
199 ; GCN-NEXT: s_mov_b32 [[BP_COPY:s[0-9]+]], s34
200 ; GCN-NEXT: s_add_i32 s33, s32, 0xffc0
201 ; GCN-NEXT: s_mov_b32 s34, s32
202 ; GCN-NEXT: s_and_b32 s33, s33, 0xffff0000
203 ; GCN-NEXT: v_lshrrev_b32_e64 [[VGPR_REG:v[0-9]+]], 6, s34
204 ; GCN-NEXT: v_mov_b32_e32 v{{[0-9]+}}, 0
205 ; GCN: s_add_i32 s32, s32, 0x30000
206 ; GCN: buffer_store_dword v{{[0-9]+}}, off, s[0:3], s33 offset:1024
207 ; GCN: buffer_load_dword v{{[0-9]+}}, [[VGPR_REG]], s[0:3], 0 offen
208 ; GCN: v_add_u32_e32 [[VGPR_REG]], vcc, 4, [[VGPR_REG]]
209 ; GCN: s_add_i32 s32, s32, 0xfffd0000
210 ; GCN-NEXT: s_mov_b32 s33, [[FP_COPY]]
211 ; GCN-NEXT: s_mov_b32 s34, [[BP_COPY]]
212 ; GCN-NEXT: s_setpc_b64 s[30:31]
214 %local_var = alloca i32, align 1024, addrspace(5)
215 store volatile i32 0, i32 addrspace(5)* %local_var, align 1024
218 loop_end: ; preds = %loop_body
219 %idx_next = add nuw nsw i32 %lp_idx, 1
220 %lp_exit_cond = icmp eq i32 %idx_next, 9
221 br i1 %lp_exit_cond, label %exit, label %loop_body
223 loop_body: ; preds = %loop_end, %begin
224 %lp_idx = phi i32 [ 0, %begin ], [ %idx_next, %loop_end ]
225 %ptr = getelementptr inbounds %struct.Data, %struct.Data addrspace(5)* %arg, i32 0, i32 0, i32 %lp_idx
226 %val = load i32, i32 addrspace(5)* %ptr, align 8
227 %lp_cond = icmp eq i32 %val, %lp_idx
228 br i1 %lp_cond, label %loop_end, label %exit
230 exit: ; preds = %loop_end, %loop_body
231 %out = phi i32 [ 0, %loop_body ], [ 1, %loop_end ]
235 define void @no_free_scratch_sgpr_for_bp_copy(<32 x i32> %a, i32 %b) #0 {
236 ; GCN-LABEL: no_free_scratch_sgpr_for_bp_copy:
238 ; GCN: v_writelane_b32 [[VGPR_REG:v[0-9]+]], s34, 0
239 ; GCN-NEXT: s_mov_b32 s34, s32
240 ; GCN-NEXT: buffer_load_dword v{{[0-9]+}}, off, s[0:3], s34
241 ; GCN: v_readlane_b32 s34, [[VGPR_REG:v[0-9]+]], 0
242 ; GCN: buffer_store_dword v{{[0-9]+}}, off, s[0:3], s33 offset:128
243 ; GCN-NEXT: s_waitcnt vmcnt(0)
244 ; GCN-NEXT: ;;#ASMSTART
245 ; GCN-NEXT: ;;#ASMEND
246 ; GCN: s_setpc_b64 s[30:31]
247 %local_val = alloca i32, align 128, addrspace(5)
248 store volatile i32 %b, i32 addrspace(5)* %local_val, align 128
249 ; Use all clobberable registers, so BP has to spill to a VGPR.
250 call void asm sideeffect "",
251 "~{s0},~{s1},~{s2},~{s3},~{s4},~{s5},~{s6},~{s7},~{s8},~{s9}
252 ,~{s10},~{s11},~{s12},~{s13},~{s14},~{s15},~{s16},~{s17},~{s18},~{s19}
253 ,~{s20},~{s21},~{s22},~{s23},~{s24},~{s25},~{s26},~{s27},~{s28},~{s29}
258 define void @no_free_regs_spill_bp_to_memory(<32 x i32> %a, i32 %b) #5 {
259 ; If there are no free SGPRs or VGPRs available we must spill the BP to memory.
261 ; GCN-LABEL: no_free_regs_spill_bp_to_mem
262 ; GCN: s_or_saveexec_b64 s[4:5], -1
263 ; GCN: v_mov_b32_e32 v0, s33
264 ; GCN: buffer_store_dword v0, off, s[0:3], s32
265 ; GCN: v_mov_b32_e32 v0, s34
266 ; GCN-DAG: buffer_store_dword v0, off, s[0:3], s32
267 %local_val = alloca i32, align 128, addrspace(5)
268 store volatile i32 %b, i32 addrspace(5)* %local_val, align 128
270 call void asm sideeffect "; clobber nonpreserved SGPRs and 64 CSRs",
271 "~{s4},~{s5},~{s6},~{s7},~{s8},~{s9}
272 ,~{s10},~{s11},~{s12},~{s13},~{s14},~{s15},~{s16},~{s17},~{s18},~{s19}
273 ,~{s20},~{s21},~{s22},~{s23},~{s24},~{s25},~{s26},~{s27},~{s28},~{s29}
274 ,~{s40},~{s41},~{s42},~{s43},~{s44},~{s45},~{s46},~{s47},~{s48},~{s49}
275 ,~{s50},~{s51},~{s52},~{s53},~{s54},~{s55},~{s56},~{s57},~{s58},~{s59}
276 ,~{s60},~{s61},~{s62},~{s63},~{s64},~{s65},~{s66},~{s67},~{s68},~{s69}
277 ,~{s70},~{s71},~{s72},~{s73},~{s74},~{s75},~{s76},~{s77},~{s78},~{s79}
278 ,~{s80},~{s81},~{s82},~{s83},~{s84},~{s85},~{s86},~{s87},~{s88},~{s89}
279 ,~{s90},~{s91},~{s92},~{s93},~{s94},~{s95},~{s96},~{s97},~{s98},~{s99}
280 ,~{s100},~{s101},~{s102},~{s39},~{vcc}"() #0
282 call void asm sideeffect "; clobber all VGPRs",
283 "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7},~{v8},~{v9}
284 ,~{v10},~{v11},~{v12},~{v13},~{v14},~{v15},~{v16},~{v17},~{v18},~{v19}
285 ,~{v20},~{v21},~{v22},~{v23},~{v24},~{v25},~{v26},~{v27},~{v28},~{v29}
286 ,~{v30},~{v31},~{v32},~{v33},~{v34},~{v35},~{v36},~{v37},~{v38}" () #0
290 define void @spill_bp_to_memory_scratch_reg_needed_mubuf_offset(<32 x i32> %a, i32 %b, [4096 x i8] addrspace(5)* byval([4096 x i8]) align 4 %arg) #5 {
291 ; If the size of the offset exceeds the MUBUF offset field we need another
292 ; scratch VGPR to hold the offset.
294 ; GCN-LABEL: spill_bp_to_memory_scratch_reg_needed_mubuf_offset
295 ; GCN: s_or_saveexec_b64 s[4:5], -1
296 ; GCN-NEXT: s_add_i32 s6, s32, 0x42100
297 ; GCN-NEXT: buffer_store_dword v39, off, s[0:3], s6 ; 4-byte Folded Spill
298 ; GCN-NEXT: s_mov_b64 exec, s[4:5]
299 ; GCN-NEXT: v_mov_b32_e32 v0, s33
300 ; GCN-NOT: v_mov_b32_e32 v0, 0x1088
301 ; GCN-NEXT: s_add_i32 s6, s32, 0x42200
302 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s6 ; 4-byte Folded Spill
303 ; GCN-NEXT: v_mov_b32_e32 v0, s34
304 ; GCN-NOT: v_mov_b32_e32 v0, 0x108c
305 ; GCN-NEXT: s_add_i32 s6, s32, 0x42300
306 ; GCN-NEXT: s_mov_b32 s34, s32
307 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s6 ; 4-byte Folded Spill
308 %local_val = alloca i32, align 128, addrspace(5)
309 store volatile i32 %b, i32 addrspace(5)* %local_val, align 128
311 call void asm sideeffect "; clobber nonpreserved SGPRs and 64 CSRs",
312 "~{s4},~{s5},~{s6},~{s7},~{s8},~{s9}
313 ,~{s10},~{s11},~{s12},~{s13},~{s14},~{s15},~{s16},~{s17},~{s18},~{s19}
314 ,~{s20},~{s21},~{s22},~{s23},~{s24},~{s25},~{s26},~{s27},~{s28},~{s29}
315 ,~{s40},~{s41},~{s42},~{s43},~{s44},~{s45},~{s46},~{s47},~{s48},~{s49}
316 ,~{s50},~{s51},~{s52},~{s53},~{s54},~{s55},~{s56},~{s57},~{s58},~{s59}
317 ,~{s60},~{s61},~{s62},~{s63},~{s64},~{s65},~{s66},~{s67},~{s68},~{s69}
318 ,~{s70},~{s71},~{s72},~{s73},~{s74},~{s75},~{s76},~{s77},~{s78},~{s79}
319 ,~{s80},~{s81},~{s82},~{s83},~{s84},~{s85},~{s86},~{s87},~{s88},~{s89}
320 ,~{s90},~{s91},~{s92},~{s93},~{s94},~{s95},~{s96},~{s97},~{s98},~{s99}
321 ,~{s100},~{s101},~{s102},~{s39},~{vcc}"() #0
323 call void asm sideeffect "; clobber all VGPRs",
324 "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7},~{v8},~{v9}
325 ,~{v10},~{v11},~{v12},~{v13},~{v14},~{v15},~{v16},~{v17},~{v18},~{v19}
326 ,~{v20},~{v21},~{v22},~{v23},~{v24},~{v25},~{v26},~{v27},~{v28},~{v29}
327 ,~{v30},~{v31},~{v32},~{v33},~{v34},~{v35},~{v36},~{v37},~{v38}"() #0
331 attributes #0 = { noinline nounwind }
332 attributes #1 = { noinline nounwind "stackrealign" }
333 attributes #2 = { noinline nounwind alignstack=4 }
334 attributes #3 = { noinline nounwind "no-realign-stack" }
335 attributes #4 = { noinline nounwind "frame-pointer"="all"}
336 attributes #5 = { noinline nounwind "amdgpu-waves-per-eu"="6,6" }