qemu/tests/tcg/i386/test-avx.c
Paolo Bonzini cf5ec6641e target/i386: implement F16C instructions
F16C only consists of two instructions, which are a bit peculiar
nevertheless.

First, they access only the low half of an YMM or XMM register for the
packed-half operand; the exact size still depends on the VEX.L flag.
This is similar to the existing avx_movx flag, but not exactly because
avx_movx is hardcoded to affect operand 2.  To this end I added a "ph"
format name; it's possible to reuse this approach for the VPMOVSX and
VPMOVZX instructions, though that would also require adding two more
formats for the low-quarter and low-eighth of an operand.

Second, VCVTPS2PH is somewhat weird because it *stores* the result of
the instruction into memory rather than loading it.

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
2022-10-20 15:16:18 +02:00

365 lines
8.7 KiB
C

#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
typedef void (*testfn)(void);
typedef struct {
uint64_t q0, q1, q2, q3;
} __attribute__((aligned(32))) v4di;
typedef struct {
uint64_t mm[8];
v4di ymm[16];
uint64_t r[16];
uint64_t flags;
uint32_t ff;
uint64_t pad;
v4di mem[4];
v4di mem0[4];
} reg_state;
typedef struct {
int n;
testfn fn;
const char *s;
reg_state *init;
} TestDef;
reg_state initI;
reg_state initF16;
reg_state initF32;
reg_state initF64;
static void dump_ymm(const char *name, int n, const v4di *r, int ff)
{
printf("%s%d = %016lx %016lx %016lx %016lx\n",
name, n, r->q3, r->q2, r->q1, r->q0);
if (ff == 64) {
double v[4];
memcpy(v, r, sizeof(v));
printf(" %16g %16g %16g %16g\n",
v[3], v[2], v[1], v[0]);
} else if (ff == 32) {
float v[8];
memcpy(v, r, sizeof(v));
printf(" %8g %8g %8g %8g %8g %8g %8g %8g\n",
v[7], v[6], v[5], v[4], v[3], v[2], v[1], v[0]);
}
}
static void dump_regs(reg_state *s)
{
int i;
for (i = 0; i < 16; i++) {
dump_ymm("ymm", i, &s->ymm[i], 0);
}
for (i = 0; i < 4; i++) {
dump_ymm("mem", i, &s->mem0[i], 0);
}
}
static void compare_state(const reg_state *a, const reg_state *b)
{
int i;
for (i = 0; i < 8; i++) {
if (a->mm[i] != b->mm[i]) {
printf("MM%d = %016lx\n", i, b->mm[i]);
}
}
for (i = 0; i < 16; i++) {
if (a->r[i] != b->r[i]) {
printf("r%d = %016lx\n", i, b->r[i]);
}
}
for (i = 0; i < 16; i++) {
if (memcmp(&a->ymm[i], &b->ymm[i], 32)) {
dump_ymm("ymm", i, &b->ymm[i], a->ff);
}
}
for (i = 0; i < 4; i++) {
if (memcmp(&a->mem0[i], &a->mem[i], 32)) {
dump_ymm("mem", i, &a->mem[i], a->ff);
}
}
if (a->flags != b->flags) {
printf("FLAGS = %016lx\n", b->flags);
}
}
#define LOADMM(r, o) "movq " #r ", " #o "[%0]\n\t"
#define LOADYMM(r, o) "vmovdqa " #r ", " #o "[%0]\n\t"
#define STOREMM(r, o) "movq " #o "[%1], " #r "\n\t"
#define STOREYMM(r, o) "vmovdqa " #o "[%1], " #r "\n\t"
#define MMREG(F) \
F(mm0, 0x00) \
F(mm1, 0x08) \
F(mm2, 0x10) \
F(mm3, 0x18) \
F(mm4, 0x20) \
F(mm5, 0x28) \
F(mm6, 0x30) \
F(mm7, 0x38)
#define YMMREG(F) \
F(ymm0, 0x040) \
F(ymm1, 0x060) \
F(ymm2, 0x080) \
F(ymm3, 0x0a0) \
F(ymm4, 0x0c0) \
F(ymm5, 0x0e0) \
F(ymm6, 0x100) \
F(ymm7, 0x120) \
F(ymm8, 0x140) \
F(ymm9, 0x160) \
F(ymm10, 0x180) \
F(ymm11, 0x1a0) \
F(ymm12, 0x1c0) \
F(ymm13, 0x1e0) \
F(ymm14, 0x200) \
F(ymm15, 0x220)
#define LOADREG(r, o) "mov " #r ", " #o "[rax]\n\t"
#define STOREREG(r, o) "mov " #o "[rax], " #r "\n\t"
#define REG(F) \
F(rbx, 0x248) \
F(rcx, 0x250) \
F(rdx, 0x258) \
F(rsi, 0x260) \
F(rdi, 0x268) \
F(r8, 0x280) \
F(r9, 0x288) \
F(r10, 0x290) \
F(r11, 0x298) \
F(r12, 0x2a0) \
F(r13, 0x2a8) \
F(r14, 0x2b0) \
F(r15, 0x2b8) \
static void run_test(const TestDef *t)
{
reg_state result;
reg_state *init = t->init;
memcpy(init->mem, init->mem0, sizeof(init->mem));
printf("%5d %s\n", t->n, t->s);
asm volatile(
MMREG(LOADMM)
YMMREG(LOADYMM)
"sub rsp, 128\n\t"
"push rax\n\t"
"push rbx\n\t"
"push rcx\n\t"
"push rdx\n\t"
"push %1\n\t"
"push %2\n\t"
"mov rax, %0\n\t"
"pushf\n\t"
"pop rbx\n\t"
"shr rbx, 8\n\t"
"shl rbx, 8\n\t"
"mov rcx, 0x2c0[rax]\n\t"
"and rcx, 0xff\n\t"
"or rbx, rcx\n\t"
"push rbx\n\t"
"popf\n\t"
REG(LOADREG)
"mov rax, 0x240[rax]\n\t"
"call [rsp]\n\t"
"mov [rsp], rax\n\t"
"mov rax, 8[rsp]\n\t"
REG(STOREREG)
"mov rbx, [rsp]\n\t"
"mov 0x240[rax], rbx\n\t"
"mov rbx, 0\n\t"
"mov 0x270[rax], rbx\n\t"
"mov 0x278[rax], rbx\n\t"
"pushf\n\t"
"pop rbx\n\t"
"and rbx, 0xff\n\t"
"mov 0x2c0[rax], rbx\n\t"
"add rsp, 16\n\t"
"pop rdx\n\t"
"pop rcx\n\t"
"pop rbx\n\t"
"pop rax\n\t"
"add rsp, 128\n\t"
MMREG(STOREMM)
YMMREG(STOREYMM)
: : "r"(init), "r"(&result), "r"(t->fn)
: "memory", "cc",
"rsi", "rdi",
"r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
"mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7",
"ymm0", "ymm1", "ymm2", "ymm3", "ymm4", "ymm5",
"ymm6", "ymm7", "ymm8", "ymm9", "ymm10", "ymm11",
"ymm12", "ymm13", "ymm14", "ymm15"
);
compare_state(init, &result);
}
#define TEST(n, cmd, type) \
static void __attribute__((naked)) test_##n(void) \
{ \
asm volatile(cmd); \
asm volatile("ret"); \
}
#include "test-avx.h"
static const TestDef test_table[] = {
#define TEST(n, cmd, type) {n, test_##n, cmd, &init##type},
#include "test-avx.h"
{-1, NULL, "", NULL}
};
static void run_all(void)
{
const TestDef *t;
for (t = test_table; t->fn; t++) {
run_test(t);
}
}
#define ARRAY_LEN(x) (sizeof(x) / sizeof(x[0]))
uint16_t val_f16[] = { 0x4000, 0xbc00, 0x44cd, 0x3a66, 0x4200, 0x7a1a, 0x4780, 0x4826 };
float val_f32[] = {2.0, -1.0, 4.8, 0.8, 3, -42.0, 5e6, 7.5, 8.3};
double val_f64[] = {2.0, -1.0, 4.8, 0.8, 3, -42.0, 5e6, 7.5};
v4di val_i64[] = {
{0x3d6b3b6a9e4118f2lu, 0x355ae76d2774d78clu,
0xac3ff76c4daa4b28lu, 0xe7fabd204cb54083lu},
{0xd851c54a56bf1f29lu, 0x4a84d1d50bf4c4fflu,
0x56621e553d52b56clu, 0xd0069553da8f584alu},
{0x5826475e2c5fd799lu, 0xfd32edc01243f5e9lu,
0x738ba2c66d3fe126lu, 0x5707219c6e6c26b4lu},
};
v4di deadbeef = {0xa5a5a5a5deadbeefull, 0xa5a5a5a5deadbeefull,
0xa5a5a5a5deadbeefull, 0xa5a5a5a5deadbeefull};
v4di indexq = {0x000000000000001full, 0x000000000000008full,
0xffffffffffffffffull, 0xffffffffffffff5full};
v4di indexd = {0x00000002000000efull, 0xfffffff500000010ull,
0x0000000afffffff0ull, 0x000000000000000eull};
v4di gather_mem[0x20];
void init_f16reg(v4di *r)
{
memset(r, 0, sizeof(*r));
memcpy(r, val_f16, sizeof(val_f16));
}
void init_f32reg(v4di *r)
{
static int n;
float v[8];
int i;
for (i = 0; i < 8; i++) {
v[i] = val_f32[n++];
if (n == ARRAY_LEN(val_f32)) {
n = 0;
}
}
memcpy(r, v, sizeof(*r));
}
void init_f64reg(v4di *r)
{
static int n;
double v[4];
int i;
for (i = 0; i < 4; i++) {
v[i] = val_f64[n++];
if (n == ARRAY_LEN(val_f64)) {
n = 0;
}
}
memcpy(r, v, sizeof(*r));
}
void init_intreg(v4di *r)
{
static uint64_t mask;
static int n;
r->q0 = val_i64[n].q0 ^ mask;
r->q1 = val_i64[n].q1 ^ mask;
r->q2 = val_i64[n].q2 ^ mask;
r->q3 = val_i64[n].q3 ^ mask;
n++;
if (n == ARRAY_LEN(val_i64)) {
n = 0;
mask *= 0x104C11DB7;
}
}
static void init_all(reg_state *s)
{
int i;
s->r[3] = (uint64_t)&s->mem[0]; /* rdx */
s->r[4] = (uint64_t)&gather_mem[ARRAY_LEN(gather_mem) / 2]; /* rsi */
s->r[5] = (uint64_t)&s->mem[2]; /* rdi */
s->flags = 2;
for (i = 0; i < 16; i++) {
s->ymm[i] = deadbeef;
}
s->ymm[13] = indexd;
s->ymm[14] = indexq;
for (i = 0; i < 4; i++) {
s->mem0[i] = deadbeef;
}
}
int main(int argc, char *argv[])
{
int i;
init_all(&initI);
init_intreg(&initI.ymm[10]);
init_intreg(&initI.ymm[11]);
init_intreg(&initI.ymm[12]);
init_intreg(&initI.mem0[1]);
printf("Int:\n");
dump_regs(&initI);
init_all(&initF16);
init_f16reg(&initF16.ymm[10]);
init_f16reg(&initF16.ymm[11]);
init_f16reg(&initF16.ymm[12]);
init_f16reg(&initF16.mem0[1]);
initF16.ff = 16;
printf("F16:\n");
dump_regs(&initF16);
init_all(&initF32);
init_f32reg(&initF32.ymm[10]);
init_f32reg(&initF32.ymm[11]);
init_f32reg(&initF32.ymm[12]);
init_f32reg(&initF32.mem0[1]);
initF32.ff = 32;
printf("F32:\n");
dump_regs(&initF32);
init_all(&initF64);
init_f64reg(&initF64.ymm[10]);
init_f64reg(&initF64.ymm[11]);
init_f64reg(&initF64.ymm[12]);
init_f64reg(&initF64.mem0[1]);
initF64.ff = 64;
printf("F64:\n");
dump_regs(&initF64);
for (i = 0; i < ARRAY_LEN(gather_mem); i++) {
init_intreg(&gather_mem[i]);
}
if (argc > 1) {
int n = atoi(argv[1]);
run_test(&test_table[n]);
} else {
run_all();
}
return 0;
}