xref: /llvm-project/llvm/test/CodeGen/AArch64/stack-guard-sve.ll (revision 1ee315ae7964c8433b772e0b5d667834994ba753)
1; RUN: llc -mtriple=aarch64--linux-gnu -mattr=+sve < %s | FileCheck %s
2
3declare dso_local void @val_fn(<vscale x 4 x float>)
4declare dso_local void @ptr_fn(ptr)
5
6; An alloca of a scalable vector shouldn't trigger stack protection.
7
8; CHECK-LABEL: call_value:
9; CHECK-NOT: mov x19, sp
10; CHECK: addvl sp, sp, #-1
11; CHECK-NOT: __stack_chk_guard
12; CHECK: st1w { {{z[0-9]+.s}} }, {{p[0-9]+}}, [x29, #-1, mul vl]
13define void @call_value() #0 {
14entry:
15  %x = alloca <vscale x 4 x float>, align 16
16  store <vscale x 4 x float> zeroinitializer, ptr %x, align 16
17  %0 = load <vscale x 4 x float>, ptr %x, align 16
18  call void @val_fn(<vscale x 4 x float> %0)
19  ret void
20}
21
22; CHECK-LABEL: call_value_strong:
23; CHECK-NOT: mov x19, sp
24; CHECK: addvl sp, sp, #-1
25; CHECK-NOT: __stack_chk_guard
26; CHECK: st1w { {{z[0-9]+.s}} }, {{p[0-9]+}}, [x29, #-1, mul vl]
27define void @call_value_strong() #1 {
28entry:
29  %x = alloca <vscale x 4 x float>, align 16
30  store <vscale x 4 x float> zeroinitializer, ptr %x, align 16
31  %0 = load <vscale x 4 x float>, ptr %x, align 16
32  call void @val_fn(<vscale x 4 x float> %0)
33  ret void
34}
35
36; Address-taking of a scalable vector should trigger stack protection only with
37; sspstrong, and the scalable vector should be be placed below the stack guard.
38
39; CHECK-LABEL: call_ptr:
40; CHECK-NOT: mov x19, sp
41; CHECK: addvl sp, sp, #-1
42; CHECK-NOT: __stack_chk_guard
43; CHECK: addvl x0, x29, #-1
44; CHECK: bl ptr_fn
45define void @call_ptr() #0 {
46entry:
47  %x = alloca <vscale x 4 x float>, align 16
48  call void @ptr_fn(ptr %x)
49  ret void
50}
51
52; CHECK-LABEL: call_ptr_strong:
53; CHECK: mov x29, sp
54; CHECK: addvl sp, sp, #-2
55; CHECK-DAG: addvl [[ADDR:x[0-9]+]], x29, #-1
56; CHECK-DAG: ldr [[VAL:x[0-9]+]], [{{x[0-9]+}}, :lo12:__stack_chk_guard]
57; CHECK-DAG: str [[VAL]], [[[ADDR]]]
58; CHECK-DAG: addvl x0, x29, #-2
59; CHECK: bl ptr_fn
60define void @call_ptr_strong() #1 {
61entry:
62  %x = alloca <vscale x 4 x float>, align 16
63  call void @ptr_fn(ptr %x)
64  ret void
65}
66
67; Check that both variables are addressed in the same way
68
69; CHECK-LABEL: call_both:
70; CHECK: mov x29, sp
71; CHECK: addvl sp, sp, #-2
72; CHECK-NOT: __stack_chk_guard
73; CHECK: st1w { {{z[0-9]+.s}} }, {{p[0-9]+}}, [x29, #-1, mul vl]
74; CHECK: bl val_fn
75; CHECK: addvl x0, x29, #-2
76; CHECK: bl ptr_fn
77define void @call_both() #0 {
78entry:
79  %x = alloca <vscale x 4 x float>, align 16
80  %y = alloca <vscale x 4 x float>, align 16
81  store <vscale x 4 x float> zeroinitializer, ptr %x, align 16
82  %0 = load <vscale x 4 x float>, ptr %x, align 16
83  call void @val_fn(<vscale x 4 x float> %0)
84  call void @ptr_fn(ptr %y)
85  ret void
86}
87
88; CHECK-LABEL: call_both_strong:
89; CHECK: mov x29, sp
90; CHECK: addvl sp, sp, #-3
91; CHECK-DAG: addvl [[ADDR:x[0-9]+]], x29, #-1
92; CHECK-DAG: ldr [[VAL:x[0-9]+]], [{{x[0-9]+}}, :lo12:__stack_chk_guard]
93; CHECK-DAG: str [[VAL]], [[[ADDR]]]
94; CHECK-DAG: st1w { {{z[0-9]+.s}} }, {{p[0-9]+}}, [x29, #-2, mul vl]
95; CHECK: bl val_fn
96; CHECK: addvl x0, x29, #-3
97; CHECK: bl ptr_fn
98define void @call_both_strong() #1 {
99entry:
100  %x = alloca <vscale x 4 x float>, align 16
101  %y = alloca <vscale x 4 x float>, align 16
102  store <vscale x 4 x float> zeroinitializer, ptr %x, align 16
103  %0 = load <vscale x 4 x float>, ptr %x, align 16
104  call void @val_fn(<vscale x 4 x float> %0)
105  call void @ptr_fn(ptr %y)
106  ret void
107}
108
109; Pushed callee-saved regs should be above the stack guard
110
111; CHECK-LABEL: callee_save:
112; CHECK: mov x29, sp
113; CHECK: addvl sp, sp, #-18
114; CHECK: str {{z[0-9]+}}, [sp, #{{[0-9]+}}, mul vl]
115; CHECK-NOT: mov x29, sp
116; CHECK: addvl sp, sp, #-1
117; CHECK-NOT: __stack_chk_guard
118; CHECK: addvl [[REG:x[0-9]+]], x29, #-11
119; CHECK: st1w { {{z[0-9]+.s}} }, {{p[0-9]+}}, [[[REG]], #-8, mul vl]
120define void @callee_save(<vscale x 4 x float> %x) #0 {
121entry:
122  %x.addr = alloca <vscale x 4 x float>, align 16
123  store <vscale x 4 x float> %x, ptr %x.addr, align 16
124  call void @ptr_fn(ptr %x.addr)
125  ret void
126}
127
128; CHECK-LABEL: callee_save_strong:
129; CHECK: mov x29, sp
130; CHECK: addvl sp, sp, #-18
131; CHECK: str {{z[0-9]+}}, [sp, #{{[0-9]+}}, mul vl]
132; CHECK: addvl sp, sp, #-2
133; CHECK-DAG: addvl [[ADDR:x[0-9]+]], x29, #-19
134; CHECK-DAG: ldr [[VAL:x[0-9]+]], [{{x[0-9]+}}, :lo12:__stack_chk_guard]
135; CHECK-DAG: str [[VAL]], [[[ADDR]]]
136; CHECK-DAG: addvl [[ADDR2:x[0-9]+]], x29, #-12
137; CHECK-DAG: st1w { z0.s }, p0, [[[ADDR2]], #-8, mul vl]
138define void @callee_save_strong(<vscale x 4 x float> %x) #1 {
139entry:
140  %x.addr = alloca <vscale x 4 x float>, align 16
141  store <vscale x 4 x float> %x, ptr %x.addr, align 16
142  call void @ptr_fn(ptr %x.addr)
143  ret void
144}
145
146; Check that local stack allocation works correctly both when we have a stack
147; guard but no vulnerable SVE objects, and when we do have such objects.
148
149; CHECK-LABEL: local_stack_alloc:
150; CHECK: mov x29, sp
151; CHECK: sub sp, sp, #16, lsl #12
152; CHECK: sub sp, sp, #16
153; CHECK: addvl sp, sp, #-2
154
155; Stack guard is placed below the SVE stack area (and above all fixed-width objects)
156; CHECK-DAG: add [[STACK_GUARD_SPILL_PART_LOC:x[0-9]+]], sp, #8, lsl #12
157; CHECK-DAG: add [[STACK_GUARD_SPILL_PART_LOC]], [[STACK_GUARD_SPILL_PART_LOC]], #16
158; CHECK-DAG: ldr [[STACK_GUARD:x[0-9]+]], [{{x[0-9]+}}, :lo12:__stack_chk_guard]
159; CHECK-DAG: str [[STACK_GUARD]], [[[STACK_GUARD_SPILL_PART_LOC]], #32760]
160
161; char_arr is below the stack guard
162; CHECK-DAG: add [[CHAR_ARR_LOC:x[0-9]+]], sp, #16, lsl #12
163; CHECK-DAG: strb wzr, [[[CHAR_ARR_LOC]]]
164
165; large1 is accessed via a virtual base register
166; CHECK-DAG: add [[LARGE1:x[0-9]+]], sp, #8, lsl #12
167; CHECK-DAG: stp x0, x0, [[[LARGE1]]]
168
169; large2 is at the bottom of the stack
170; CHECK-DAG: stp x0, x0, [sp]
171
172; vec1 and vec2 are in the SVE stack immediately below fp
173; CHECK-DAG: addvl x0, x29, #-1
174; CHECK-DAG: bl ptr_fn
175; CHECK-DAG: addvl x0, x29, #-2
176; CHECK-DAG: bl ptr_fn
177define void @local_stack_alloc(i64 %val) #0 {
178entry:
179  %char_arr = alloca [8 x i8], align 4
180  %gep0 = getelementptr [8 x i8], ptr %char_arr, i64 0, i64 0
181  store i8 0, ptr %gep0, align 8
182  %large1 = alloca [4096 x i64], align 8
183  %large2 = alloca [4096 x i64], align 8
184  %vec_1 = alloca <vscale x 4 x float>, align 16
185  %vec_2 = alloca <vscale x 4 x float>, align 16
186  %gep1 = getelementptr [4096 x i64], ptr %large1, i64 0, i64 0
187  %gep2 = getelementptr [4096 x i64], ptr %large1, i64 0, i64 1
188  store i64 %val, ptr %gep1, align 8
189  store i64 %val, ptr %gep2, align 8
190  %gep3 = getelementptr [4096 x i64], ptr %large2, i64 0, i64 0
191  %gep4 = getelementptr [4096 x i64], ptr %large2, i64 0, i64 1
192  store i64 %val, ptr %gep3, align 8
193  store i64 %val, ptr %gep4, align 8
194  call void @ptr_fn(ptr %vec_1)
195  call void @ptr_fn(ptr %vec_2)
196  ret void
197}
198
199; CHECK-LABEL: local_stack_alloc_strong:
200; CHECK: mov x29, sp
201; CHECK: sub sp, sp, #16, lsl #12
202; CHECK: sub sp, sp, #16
203; CHECK: addvl sp, sp, #-3
204
205; Stack guard is placed at the top of the SVE stack area
206; CHECK-DAG: ldr [[STACK_GUARD:x[0-9]+]], [{{x[0-9]+}}, :lo12:__stack_chk_guard]
207; CHECK-DAG: addvl [[STACK_GUARD_POS:x[0-9]+]], x29, #-1
208; CHECK-DAG: str [[STACK_GUARD]], [[[STACK_GUARD_POS]]]
209
210; char_arr is below the SVE stack area
211; CHECK-DAG: add [[CHAR_ARR:x[0-9]+]], sp, #15, lsl #12            // =61440
212; CHECK-DAG: add [[CHAR_ARR]], [[CHAR_ARR]], #9
213; CHECK-DAG: strb wzr, [[[CHAR_ARR]], #4095]
214
215; large1 is accessed via a virtual base register
216; CHECK-DAG: add [[LARGE1:x[0-9]+]], sp, #8, lsl #12
217; CHECK-DAG: stp x0, x0, [[[LARGE1]], #8]
218
219; large2 is at the bottom of the stack
220; CHECK-DAG: stp x0, x0, [sp, #8]
221
222; vec1 and vec2 are in the SVE stack area below the stack guard
223; CHECK-DAG: addvl x0, x29, #-2
224; CHECK-DAG: bl ptr_fn
225; CHECK-DAG: addvl x0, x29, #-3
226; CHECK-DAG: bl ptr_fn
227define void @local_stack_alloc_strong(i64 %val) #1 {
228entry:
229  %char_arr = alloca [8 x i8], align 4
230  %gep0 = getelementptr [8 x i8], ptr %char_arr, i64 0, i64 0
231  store i8 0, ptr %gep0, align 8
232  %large1 = alloca [4096 x i64], align 8
233  %large2 = alloca [4096 x i64], align 8
234  %vec_1 = alloca <vscale x 4 x float>, align 16
235  %vec_2 = alloca <vscale x 4 x float>, align 16
236  %gep1 = getelementptr [4096 x i64], ptr %large1, i64 0, i64 0
237  %gep2 = getelementptr [4096 x i64], ptr %large1, i64 0, i64 1
238  store i64 %val, ptr %gep1, align 8
239  store i64 %val, ptr %gep2, align 8
240  %gep3 = getelementptr [4096 x i64], ptr %large2, i64 0, i64 0
241  %gep4 = getelementptr [4096 x i64], ptr %large2, i64 0, i64 1
242  store i64 %val, ptr %gep3, align 8
243  store i64 %val, ptr %gep4, align 8
244  call void @ptr_fn(ptr %vec_1)
245  call void @ptr_fn(ptr %vec_2)
246  ret void
247}
248
249; A GEP addressing into a vector of <vscale x 4 x float> is in-bounds for
250; offsets up to 3, but out-of-bounds (and so triggers stack protection with
251; sspstrong) after that.
252
253; CHECK-LABEL: vector_gep_3:
254; CHECK-NOT: __stack_chk_guard
255define void @vector_gep_3() #0 {
256entry:
257  %vec = alloca <vscale x 4 x float>, align 16
258  %gep = getelementptr <vscale x 4 x float>, ptr %vec, i64 0, i64 3
259  store float 0.0, ptr %gep, align 4
260  ret void
261}
262
263; CHECK-LABEL: vector_gep_4:
264; CHECK-NOT: __stack_chk_guard
265define void @vector_gep_4() #0 {
266entry:
267  %vec = alloca <vscale x 4 x float>, align 16
268  %gep = getelementptr <vscale x 4 x float>, ptr %vec, i64 0, i64 4
269  store float 0.0, ptr %gep, align 4
270  ret void
271}
272
273; CHECK-LABEL: vector_gep_twice:
274; CHECK-NOT: __stack_chk_guard
275define void @vector_gep_twice() #0 {
276entry:
277  %vec = alloca <vscale x 4 x float>, align 16
278  %gep1 = getelementptr <vscale x 4 x float>, ptr %vec, i64 0, i64 3
279  store float 0.0, ptr %gep1, align 4
280  %gep2 = getelementptr float, ptr %gep1, i64 1
281  store float 0.0, ptr %gep2, align 4
282  ret void
283}
284
285; CHECK-LABEL: vector_gep_n:
286; CHECK-NOT: __stack_chk_guard
287define void @vector_gep_n(i64 %n) #0 {
288entry:
289  %vec = alloca <vscale x 4 x float>, align 16
290  %gep = getelementptr <vscale x 4 x float>, ptr %vec, i64 0, i64 %n
291  store float 0.0, ptr %gep, align 4
292  ret void
293}
294
295; CHECK-LABEL: vector_gep_3_strong:
296; CHECK-NOT: __stack_chk_guard
297define void @vector_gep_3_strong() #1 {
298entry:
299  %vec = alloca <vscale x 4 x float>, align 16
300  %gep = getelementptr <vscale x 4 x float>, ptr %vec, i64 0, i64 3
301  store float 0.0, ptr %gep, align 4
302  ret void
303}
304
305; CHECK-LABEL: vector_gep_4_strong:
306; CHECK: __stack_chk_guard
307define void @vector_gep_4_strong(i64 %val) #1 {
308entry:
309  %vec = alloca <vscale x 4 x float>, align 16
310  %gep = getelementptr <vscale x 4 x float>, ptr %vec, i64 0, i64 4
311  store float 0.0, ptr %gep, align 4
312  ret void
313}
314
315
316; CHECK-LABEL: vector_gep_twice_strong:
317; CHECK: __stack_chk_guard
318define void @vector_gep_twice_strong() #1 {
319entry:
320  %vec = alloca <vscale x 4 x float>, align 16
321  %gep1 = getelementptr <vscale x 4 x float>, ptr %vec, i64 0, i64 3
322  store float 0.0, ptr %gep1, align 4
323  %gep2 = getelementptr float, ptr %gep1, i64 1
324  store float 0.0, ptr %gep2, align 4
325  ret void
326}
327
328; CHECK-LABEL: vector_gep_n_strong:
329; CHECK: __stack_chk_guard
330define void @vector_gep_n_strong(i64 %n) #1 {
331entry:
332  %vec = alloca <vscale x 4 x float>, align 16
333  %gep = getelementptr <vscale x 4 x float>, ptr %vec, i64 0, i64 %n
334  store float 0.0, ptr %gep, align 4
335  ret void
336}
337
338attributes #0 = { ssp "frame-pointer"="non-leaf" }
339attributes #1 = { sspstrong "frame-pointer"="non-leaf" }
340
341!llvm.module.flags = !{!0}
342!0 = !{i32 7, !"direct-access-external-data", i32 1}
343