1 /* { dg-do compile } */
2 /* { dg-skip-if "Require optimsation to compile DCE tests" { *-*-* } { "-O0" } { "" } } */
3 /* { dg-require-effective-target arm_v8m_main_cde_ok } */
4 /* { dg-add-options arm_v8m_main_cde } */
5 /* { dg-final { check-function-bodies "**" "" } } */
6
7 /* These are the scalar intrinsics.
8 uint32_t __arm_cx1(int coproc, uint32_t imm);
9 uint32_t __arm_cx1a(int coproc, uint32_t acc, uint32_t imm);
10 uint32_t __arm_cx2(int coproc, uint32_t n, uint32_t imm);
11 uint32_t __arm_cx2a(int coproc, uint32_t acc, uint32_t n, uint32_t imm);
12 uint32_t __arm_cx3(int coproc, uint32_t n, uint32_t m, uint32_t imm);
13 uint32_t __arm_cx3a(int coproc, uint32_t acc, uint32_t n, uint32_t m, uint32_t imm);
14
15 uint64_t __arm_cx1d(int coproc, uint32_t imm);
16 uint64_t __arm_cx1da(int coproc, uint64_t acc, uint32_t imm);
17 uint64_t __arm_cx2d(int coproc, uint32_t n, uint32_t imm);
18 uint64_t __arm_cx2da(int coproc, uint64_t acc, uint32_t n, uint32_t imm);
19 uint64_t __arm_cx3d(int coproc, uint32_t n, uint32_t m, uint32_t imm);
20 uint64_t __arm_cx3da(int coproc, uint64_t acc, uint32_t n, uint32_t m, uint32_t imm); */
21
22 #include "arm_cde.h"
23
24 #define TEST_CDE_SCALAR_INTRINSIC(name, accum_type, arguments) \
25 accum_type test_cde_##name (__attribute__ ((unused)) uint32_t n, \
26 __attribute__ ((unused)) uint32_t m) \
27 { \
28 accum_type accum = 0; \
29 accum += __arm_##name arguments; \
30 return accum; \
31 }
32
33 /* Basic test that we produce the assembly as expected. */
34 /*
35 ** test_cde_cx1:
36 ** cx1 p0, r0, #33
37 ** bx lr
38 */
39 TEST_CDE_SCALAR_INTRINSIC (cx1, uint32_t, (0, 33))
40
41 /*
42 ** test_cde_cx1a:
43 ** movs r0, #0
44 ** cx1a p0, r0, #33
45 ** bx lr
46 */
47 TEST_CDE_SCALAR_INTRINSIC (cx1a, uint32_t, (0, accum, 33))
48
49 /*
50 ** test_cde_cx2:
51 ** cx2 p0, r0, r0, #33
52 ** bx lr
53 */
54 TEST_CDE_SCALAR_INTRINSIC (cx2, uint32_t, (0, n, 33))
55
56 /*
57 ** test_cde_cx2a:
58 ** movs (r[0-9]+), #0
59 ** cx2a p0, \1, r0, #33
60 ** mov r0, \1
61 ** bx lr
62 */
63 TEST_CDE_SCALAR_INTRINSIC (cx2a, uint32_t, (0, accum, n, 33))
64
65 /*
66 ** test_cde_cx3:
67 ** cx3 p0, r0, r0, r1, #33
68 ** bx lr
69 */
70 TEST_CDE_SCALAR_INTRINSIC (cx3, uint32_t, (0, n, m, 33))
71
72 /*
73 ** test_cde_cx3a:
74 ** movs (r[0-9]+), #0
75 ** cx3a p0, \1, r0, r1, #33
76 ** mov r0, \1
77 ** bx lr
78 */
79 TEST_CDE_SCALAR_INTRINSIC (cx3a, uint32_t, (0, accum, n, m, 33))
80
81 /*
82 ** test_cde_cx1d:
83 ** cx1d p0, r0, r1, #33
84 ** bx lr
85 */
86 TEST_CDE_SCALAR_INTRINSIC (cx1d, uint64_t, (0, 33))
87
88 /*
89 ** test_cde_cx1da:
90 ** movs r0, #0
91 ** movs r1, #0
92 ** cx1da p0, r0, r1, #33
93 ** bx lr
94 */
95 TEST_CDE_SCALAR_INTRINSIC (cx1da, uint64_t, (0, accum, 33))
96
97 /*
98 ** test_cde_cx2d:
99 ** cx2d p0, r0, r1, r0, #33
100 ** bx lr
101 */
102 TEST_CDE_SCALAR_INTRINSIC (cx2d, uint64_t, (0, n, 33))
103
104 /* This particular function gets optimised by the compiler in two different
105 ways depending on the optimisation level. So does test_cde_cx3da. That's
106 why we have two different regexes in each of these function body checks. */
107 /*
108 ** test_cde_cx2da:
109 ** (
110 ** mov (r[0-9]+), r0
111 ** movs r0, #0
112 ** movs r1, #0
113 ** cx2da p0, r0, r1, \1, #33
114 ** |
115 ** movs (r[0-9]+), #0
116 ** movs (r[0-9]+), #0
117 ** cx2da p0, \2, \3, r0, #33
118 ** mov r0, \2
119 ** mov r1, \3
120 ** )
121 ** bx lr
122 */
123 TEST_CDE_SCALAR_INTRINSIC (cx2da, uint64_t, (0, accum, n, 33))
124
125 /*
126 ** test_cde_cx3d:
127 ** cx3d p0, r0, r1, r0, r1, #33
128 ** bx lr
129 */
130 TEST_CDE_SCALAR_INTRINSIC (cx3d, uint64_t, (0, n, m, 33))
131
132 /*
133 ** test_cde_cx3da:
134 ** ...
135 ** (
136 ** movs (r[0-9]+), #0
137 ** movs (r[0-9]+), #0
138 ** cx3da p0, \1, \2, r0, r1, #33
139 ** mov r0, \1
140 ** mov r1, \2
141 ** |
142 ** movs r0, #0
143 ** movs r1, #0
144 ** cx3da p0, r0, r1, r[0-9]+, r[0-9]+, #33
145 ** )
146 ** ...
147 ** bx lr
148 */
149 TEST_CDE_SCALAR_INTRINSIC (cx3da, uint64_t, (0, accum, n, m, 33))
150
151
152
153 /* Ensure this function gets DCE'd out after optimisation.
154 Should be such since the ACLE specification mentions these functions are
155 stateless and pure. */
156 /*
157 ** test_cde_dce:
158 ** bx lr
159 */
160 void test_cde_dce (uint32_t n, uint32_t m)
161 {
162 uint64_t accum = 0;
163 __arm_cx1 (0, 33);
164 __arm_cx1a (0, accum, 33);
165 __arm_cx2 (0, n, 33);
166 __arm_cx2a (0, accum, n, 33);
167 __arm_cx3 (0, n, m, 33);
168 __arm_cx3a (0, accum, n, m, 33);
169 __arm_cx1d (0, 33);
170 __arm_cx1da (0, accum, 33);
171 __arm_cx2d (0, n, 33);
172 __arm_cx2da (0, accum, n, 33);
173 __arm_cx3d (0, n, m, 33);
174 __arm_cx3da (0, accum, n, m, 33);
175 }
176
177 /* Checking this function allows constants with symbolic names.
178 This test must be run under some level of optimisation.
179 The actual check we perform is that the function is provided something that,
180 at the point of expansion, is an immediate. That check is not as strict as
181 having something that is an immediate directly.
182
183 Since we've already checked these intrinsics generate code in the manner we
184 expect (above), here we just check that all the instructions we expect are
185 there. To ensure the instructions are from these functions we use different
186 constants and search for those specifically with `scan-assembler-times`. */
187
188 /* Checking this function allows constants with symbolic names. */
189 uint32_t test_cde2 (uint32_t n, uint32_t m)
190 {
191 int coproc = 6;
192 uint32_t imm = 30;
193 uint32_t accum = 0;
194 accum += __arm_cx1 (coproc, imm);
195 accum += __arm_cx1a (coproc, accum, imm);
196 accum += __arm_cx2 (coproc, n, imm);
197 accum += __arm_cx2a (coproc, accum, n, imm);
198 accum += __arm_cx3 (coproc, n, m, imm);
199 accum += __arm_cx3a (coproc, accum, n, m, imm);
200 return accum;
201 }
202
203 /* Checking this function allows constants with symbolic names. */
204 uint64_t test_cdedi2 (uint32_t n, uint32_t m)
205 {
206 int coproc = 6;
207 uint32_t imm = 30;
208 uint64_t accum = 0;
209 accum += __arm_cx1d (coproc, imm);
210 accum += __arm_cx1da (coproc, accum, imm);
211 accum += __arm_cx2d (coproc, n, imm);
212 accum += __arm_cx2da (coproc, accum, n, imm);
213 accum += __arm_cx3d (coproc, n, m, imm);
214 accum += __arm_cx3da (coproc, accum, n, m, imm);
215 return accum;
216 }
217
218 /* { dg-final { scan-assembler-times "cx1\\tp6" 1 } } */
219 /* { dg-final { scan-assembler-times "cx2\\tp6" 1 } } */
220 /* { dg-final { scan-assembler-times "cx3\\tp6" 1 } } */
221 /* { dg-final { scan-assembler-times "cx1a\\tp6" 1 } } */
222 /* { dg-final { scan-assembler-times "cx2a\\tp6" 1 } } */
223 /* { dg-final { scan-assembler-times "cx3a\\tp6" 1 } } */
224 /* { dg-final { scan-assembler-times "cx1d\\tp6" 1 } } */
225 /* { dg-final { scan-assembler-times "cx2d\\tp6" 1 } } */
226 /* { dg-final { scan-assembler-times "cx3d\\tp6" 1 } } */
227 /* { dg-final { scan-assembler-times "cx1da\\tp6" 1 } } */
228 /* { dg-final { scan-assembler-times "cx2da\\tp6" 1 } } */
229 /* { dg-final { scan-assembler-times "cx3da\\tp6" 1 } } */