24     int n;  member
35 static void dump_ymm(const char *name, int n, const v4di *r, int ff)  in dump_ymm()  argument
37     printf("%s%d = %016lx %016lx %016lx %016lx\n",  in dump_ymm()
38            name, n, r->q3, r->q2, r->q1, r->q0);  in dump_ymm()
42         printf("        %16g %16g %16g %16g\n",  in dump_ymm()
47         printf(" %8g %8g %8g %8g %8g %8g %8g %8g\n",  in dump_ymm()
57         dump_ymm("ymm", i, &s->ymm[i], 0);  in dump_regs()
60         dump_ymm("mem", i, &s->mem0[i], 0);  in dump_regs()
68         if (a->mm[i] != b->mm[i]) {  in compare_state()
69             printf("MM%d = %016lx\n", i, b->mm[i]);  in compare_state()
73         if (a->r[i] != b->r[i]) {  in compare_state()
74             printf("r%d = %016lx\n", i, b->r[i]);  in compare_state()
78         if (memcmp(&a->ymm[i], &b->ymm[i], 32)) {  in compare_state()
79             dump_ymm("ymm", i, &b->ymm[i], a->ff);  in compare_state()
83         if (memcmp(&a->mem0[i], &a->mem[i], 32)) {  in compare_state()
84             dump_ymm("mem", i, &a->mem[i], a->ff);  in compare_state()
87     if (a->flags != b->flags) {  in compare_state()
88         printf("FLAGS = %016lx\n", b->flags);  in compare_state()
92 #define LOADMM(r, o) "movq " #r ", " #o "[%0]\n\t"
93 #define LOADYMM(r, o) "vmovdqa " #r ", " #o "[%0]\n\t"
94 #define STOREMM(r, o) "movq " #o "[%1], " #r "\n\t"
95 #define STOREYMM(r, o) "vmovdqa " #o "[%1], " #r "\n\t"
122 #define LOADREG(r, o) "mov " #r ", " #o "[rax]\n\t"
123 #define STOREREG(r, o) "mov " #o "[rax], " #r "\n\t"
142     reg_state *init = t->init;  in run_test()
143     memcpy(init->mem, init->mem0, sizeof(init->mem));  in run_test()
144     printf("%5d %s\n", t->n, t->s);  in run_test()
148             "sub rsp, 128\n\t"  in run_test()
149             "push rax\n\t"  in run_test()
150             "push rbx\n\t"  in run_test()
151             "push rcx\n\t"  in run_test()
152             "push rdx\n\t"  in run_test()
153             "push %1\n\t"  in run_test()
154             "push %2\n\t"  in run_test()
155             "mov rax, %0\n\t"  in run_test()
156             "pushf\n\t"  in run_test()
157             "pop rbx\n\t"  in run_test()
158             "shr rbx, 8\n\t"  in run_test()
159             "shl rbx, 8\n\t"  in run_test()
160             "mov rcx, 0x2c0[rax]\n\t"  in run_test()
161             "and rcx, 0xff\n\t"  in run_test()
162             "or rbx, rcx\n\t"  in run_test()
163             "push rbx\n\t"  in run_test()
164             "popf\n\t"  in run_test()
166             "mov rax, 0x240[rax]\n\t"  in run_test()
167             "call [rsp]\n\t"  in run_test()
168             "mov [rsp], rax\n\t"  in run_test()
169             "mov rax, 8[rsp]\n\t"  in run_test()
171             "mov rbx, [rsp]\n\t"  in run_test()
172             "mov 0x240[rax], rbx\n\t"  in run_test()
173             "mov rbx, 0\n\t"  in run_test()
174             "mov 0x270[rax], rbx\n\t"  in run_test()
175             "mov 0x278[rax], rbx\n\t"  in run_test()
176             "pushf\n\t"  in run_test()
177             "pop rbx\n\t"  in run_test()
178             "and rbx, 0xff\n\t"  in run_test()
179             "mov 0x2c0[rax], rbx\n\t"  in run_test()
180             "add rsp, 16\n\t"  in run_test()
181             "pop rdx\n\t"  in run_test()
182             "pop rcx\n\t"  in run_test()
183             "pop rbx\n\t"  in run_test()
184             "pop rax\n\t"  in run_test()
185             "add rsp, 128\n\t"  in run_test()
188             : : "r"(init), "r"(&result), "r"(t->fn)  in run_test()
200 #define TEST(n, cmd, type) \  argument
201 static void __attribute__((naked)) test_##n(void) \
206 #include "test-avx.h"
210 #define TEST(n, cmd, type) {n, test_##n, cmd, &init##type},  argument
211 #include "test-avx.h"
212     {-1, NULL, "", NULL}
218     for (t = test_table; t->fn; t++) {  in run_all()
226 float val_f32[] = {2.0, -1.0, 4.8, 0.8, 3, -42.0, 5e6, 7.5, 8.3};
227 double val_f64[] = {2.0, -1.0, 4.8, 0.8, 3, -42.0, 5e6, 7.5};
239 /* &gather_mem[0x10] is 512 bytes from the base; indices must be >=-64, <64
257     static int n;  in init_f32reg()  local
261         v[i] = val_f32[n++];  in init_f32reg()
262         if (n == ARRAY_LEN(val_f32)) {  in init_f32reg()
263             n = 0;  in init_f32reg()
271     static int n;  in init_f64reg()  local
275         v[i] = val_f64[n++];  in init_f64reg()
276         if (n == ARRAY_LEN(val_f64)) {  in init_f64reg()
277             n = 0;  in init_f64reg()
286     static int n;  in init_intreg()  local
288     r->q0 = val_i64[n].q0 ^ mask;  in init_intreg()
289     r->q1 = val_i64[n].q1 ^ mask;  in init_intreg()
290     r->q2 = val_i64[n].q2 ^ mask;  in init_intreg()
291     r->q3 = val_i64[n].q3 ^ mask;  in init_intreg()
292     n++;  in init_intreg()
293     if (n == ARRAY_LEN(val_i64)) {  in init_intreg()
294         n = 0;  in init_intreg()
303     s->r[3] = (uint64_t)&s->mem[0]; /* rdx */  in init_all()
304     s->r[4] = (uint64_t)&gather_mem[ARRAY_LEN(gather_mem) / 2]; /* rsi */  in init_all()
305     s->r[5] = (uint64_t)&s->mem[2]; /* rdi */  in init_all()
306     s->flags = 2;  in init_all()
308         s->ymm[i] = deadbeef;  in init_all()
310     s->ymm[13] = indexd;  in init_all()
311     s->ymm[14] = indexq;  in init_all()
313         s->mem0[i] = deadbeef;  in init_all()
328     printf("Int:\n");  in main()
339     printf("F16:\n");  in main()
350     printf("F32:\n");  in main()
361     printf("F64:\n");  in main()
369         int n = atoi(argv[1]);  in main()  local
370         run_test(&test_table[n]);  in main()