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>
This commit is contained in:
Paolo Bonzini 2022-10-19 13:22:06 +02:00
parent 314d3eff66
commit cf5ec6641e
9 changed files with 89 additions and 6 deletions

View File

@ -625,13 +625,12 @@ void x86_cpu_vendor_words2str(char *dst, uint32_t vendor1,
CPUID_EXT_SSE41 | CPUID_EXT_SSE42 | CPUID_EXT_POPCNT | \ CPUID_EXT_SSE41 | CPUID_EXT_SSE42 | CPUID_EXT_POPCNT | \
CPUID_EXT_XSAVE | /* CPUID_EXT_OSXSAVE is dynamic */ \ CPUID_EXT_XSAVE | /* CPUID_EXT_OSXSAVE is dynamic */ \
CPUID_EXT_MOVBE | CPUID_EXT_AES | CPUID_EXT_HYPERVISOR | \ CPUID_EXT_MOVBE | CPUID_EXT_AES | CPUID_EXT_HYPERVISOR | \
CPUID_EXT_RDRAND | CPUID_EXT_AVX) CPUID_EXT_RDRAND | CPUID_EXT_AVX | CPUID_EXT_F16C)
/* missing: /* missing:
CPUID_EXT_DTES64, CPUID_EXT_DSCPL, CPUID_EXT_VMX, CPUID_EXT_SMX, CPUID_EXT_DTES64, CPUID_EXT_DSCPL, CPUID_EXT_VMX, CPUID_EXT_SMX,
CPUID_EXT_EST, CPUID_EXT_TM2, CPUID_EXT_CID, CPUID_EXT_FMA, CPUID_EXT_EST, CPUID_EXT_TM2, CPUID_EXT_CID, CPUID_EXT_FMA,
CPUID_EXT_XTPR, CPUID_EXT_PDCM, CPUID_EXT_PCID, CPUID_EXT_DCA, CPUID_EXT_XTPR, CPUID_EXT_PDCM, CPUID_EXT_PCID, CPUID_EXT_DCA,
CPUID_EXT_X2APIC, CPUID_EXT_TSC_DEADLINE_TIMER, CPUID_EXT_X2APIC, CPUID_EXT_TSC_DEADLINE_TIMER */
CPUID_EXT_F16C */
#ifdef TARGET_X86_64 #ifdef TARGET_X86_64
#define TCG_EXT2_X86_64_FEATURES (CPUID_EXT2_SYSCALL | CPUID_EXT2_LM) #define TCG_EXT2_X86_64_FEATURES (CPUID_EXT2_SYSCALL | CPUID_EXT2_LM)

View File

@ -1258,6 +1258,7 @@ typedef union ZMMReg {
uint16_t _w_ZMMReg[512 / 16]; uint16_t _w_ZMMReg[512 / 16];
uint32_t _l_ZMMReg[512 / 32]; uint32_t _l_ZMMReg[512 / 32];
uint64_t _q_ZMMReg[512 / 64]; uint64_t _q_ZMMReg[512 / 64];
float16 _h_ZMMReg[512 / 16];
float32 _s_ZMMReg[512 / 32]; float32 _s_ZMMReg[512 / 32];
float64 _d_ZMMReg[512 / 64]; float64 _d_ZMMReg[512 / 64];
XMMReg _x_ZMMReg[512 / 128]; XMMReg _x_ZMMReg[512 / 128];
@ -1282,6 +1283,7 @@ typedef struct BNDCSReg {
#define ZMM_B(n) _b_ZMMReg[63 - (n)] #define ZMM_B(n) _b_ZMMReg[63 - (n)]
#define ZMM_W(n) _w_ZMMReg[31 - (n)] #define ZMM_W(n) _w_ZMMReg[31 - (n)]
#define ZMM_L(n) _l_ZMMReg[15 - (n)] #define ZMM_L(n) _l_ZMMReg[15 - (n)]
#define ZMM_H(n) _h_ZMMReg[31 - (n)]
#define ZMM_S(n) _s_ZMMReg[15 - (n)] #define ZMM_S(n) _s_ZMMReg[15 - (n)]
#define ZMM_Q(n) _q_ZMMReg[7 - (n)] #define ZMM_Q(n) _q_ZMMReg[7 - (n)]
#define ZMM_D(n) _d_ZMMReg[7 - (n)] #define ZMM_D(n) _d_ZMMReg[7 - (n)]
@ -1301,6 +1303,7 @@ typedef struct BNDCSReg {
#define ZMM_B(n) _b_ZMMReg[n] #define ZMM_B(n) _b_ZMMReg[n]
#define ZMM_W(n) _w_ZMMReg[n] #define ZMM_W(n) _w_ZMMReg[n]
#define ZMM_L(n) _l_ZMMReg[n] #define ZMM_L(n) _l_ZMMReg[n]
#define ZMM_H(n) _h_ZMMReg[n]
#define ZMM_S(n) _s_ZMMReg[n] #define ZMM_S(n) _s_ZMMReg[n]
#define ZMM_Q(n) _q_ZMMReg[n] #define ZMM_Q(n) _q_ZMMReg[n]
#define ZMM_D(n) _d_ZMMReg[n] #define ZMM_D(n) _d_ZMMReg[n]

View File

@ -586,6 +586,35 @@ void glue(helper_cvtpd2ps, SUFFIX)(CPUX86State *env, Reg *d, Reg *s)
} }
} }
#if SHIFT >= 1
void glue(helper_cvtph2ps, SUFFIX)(CPUX86State *env, Reg *d, Reg *s)
{
int i;
for (i = 2 << SHIFT; --i >= 0; ) {
d->ZMM_S(i) = float16_to_float32(s->ZMM_H(i), true, &env->sse_status);
}
}
void glue(helper_cvtps2ph, SUFFIX)(CPUX86State *env, Reg *d, Reg *s, int mode)
{
int i;
FloatRoundMode prev_rounding_mode = env->sse_status.float_rounding_mode;
if (!(mode & (1 << 2))) {
set_x86_rounding_mode(mode & 3, &env->sse_status);
}
for (i = 0; i < 2 << SHIFT; i++) {
d->ZMM_H(i) = float32_to_float16(s->ZMM_S(i), true, &env->sse_status);
}
for (i >>= 2; i < 1 << SHIFT; i++) {
d->Q(i) = 0;
}
env->sse_status.float_rounding_mode = prev_rounding_mode;
}
#endif
#if SHIFT == 1 #if SHIFT == 1
void helper_cvtss2sd(CPUX86State *env, Reg *d, Reg *v, Reg *s) void helper_cvtss2sd(CPUX86State *env, Reg *d, Reg *v, Reg *s)
{ {

View File

@ -353,6 +353,12 @@ DEF_HELPER_4(glue(aeskeygenassist, SUFFIX), void, env, Reg, Reg, i32)
DEF_HELPER_5(glue(pclmulqdq, SUFFIX), void, env, Reg, Reg, Reg, i32) DEF_HELPER_5(glue(pclmulqdq, SUFFIX), void, env, Reg, Reg, Reg, i32)
#endif #endif
/* F16C helpers */
#if SHIFT >= 1
DEF_HELPER_3(glue(cvtph2ps, SUFFIX), void, env, Reg, Reg)
DEF_HELPER_4(glue(cvtps2ph, SUFFIX), void, env, Reg, Reg, int)
#endif
/* AVX helpers */ /* AVX helpers */
#if SHIFT >= 1 #if SHIFT >= 1
DEF_HELPER_4(glue(vpermilpd, SUFFIX), void, env, Reg, Reg, Reg) DEF_HELPER_4(glue(vpermilpd, SUFFIX), void, env, Reg, Reg, Reg)

View File

@ -336,6 +336,7 @@ static const X86OpEntry opcodes_0F38_00toEF[240] = {
[0x07] = X86_OP_ENTRY3(PHSUBSW, V,x, H,x, W,x, vex4 cpuid(SSSE3) mmx avx2_256 p_00_66), [0x07] = X86_OP_ENTRY3(PHSUBSW, V,x, H,x, W,x, vex4 cpuid(SSSE3) mmx avx2_256 p_00_66),
[0x10] = X86_OP_ENTRY2(PBLENDVB, V,x, W,x, vex4 cpuid(SSE41) avx2_256 p_66), [0x10] = X86_OP_ENTRY2(PBLENDVB, V,x, W,x, vex4 cpuid(SSE41) avx2_256 p_66),
[0x13] = X86_OP_ENTRY2(VCVTPH2PS, V,x, W,ph, vex11 cpuid(F16C) p_66),
[0x14] = X86_OP_ENTRY2(BLENDVPS, V,x, W,x, vex4 cpuid(SSE41) p_66), [0x14] = X86_OP_ENTRY2(BLENDVPS, V,x, W,x, vex4 cpuid(SSE41) p_66),
[0x15] = X86_OP_ENTRY2(BLENDVPD, V,x, W,x, vex4 cpuid(SSE41) p_66), [0x15] = X86_OP_ENTRY2(BLENDVPD, V,x, W,x, vex4 cpuid(SSE41) p_66),
/* Listed incorrectly as type 4 */ /* Listed incorrectly as type 4 */
@ -525,6 +526,7 @@ static const X86OpEntry opcodes_0F3A[256] = {
[0x15] = X86_OP_ENTRY3(PEXTRW, E,w, V,dq, I,b, vex5 cpuid(SSE41) zext0 p_66), [0x15] = X86_OP_ENTRY3(PEXTRW, E,w, V,dq, I,b, vex5 cpuid(SSE41) zext0 p_66),
[0x16] = X86_OP_ENTRY3(PEXTR, E,y, V,dq, I,b, vex5 cpuid(SSE41) p_66), [0x16] = X86_OP_ENTRY3(PEXTR, E,y, V,dq, I,b, vex5 cpuid(SSE41) p_66),
[0x17] = X86_OP_ENTRY3(VEXTRACTPS, E,d, V,dq, I,b, vex5 cpuid(SSE41) p_66), [0x17] = X86_OP_ENTRY3(VEXTRACTPS, E,d, V,dq, I,b, vex5 cpuid(SSE41) p_66),
[0x1d] = X86_OP_ENTRY3(VCVTPS2PH, W,ph, V,x, I,b, vex11 cpuid(F16C) p_66),
[0x20] = X86_OP_ENTRY4(PINSRB, V,dq, H,dq, E,b, vex5 cpuid(SSE41) zext2 p_66), [0x20] = X86_OP_ENTRY4(PINSRB, V,dq, H,dq, E,b, vex5 cpuid(SSE41) zext2 p_66),
[0x21] = X86_OP_GROUP0(VINSERTPS), [0x21] = X86_OP_GROUP0(VINSERTPS),
@ -1051,6 +1053,10 @@ static bool decode_op_size(DisasContext *s, X86OpEntry *e, X86OpSize size, MemOp
*ot = s->vex_l ? MO_256 : MO_128; *ot = s->vex_l ? MO_256 : MO_128;
return true; return true;
case X86_SIZE_ph: /* SSE/AVX packed half precision */
*ot = s->vex_l ? MO_128 : MO_64;
return true;
case X86_SIZE_d64: /* Default to 64-bit in 64-bit mode */ case X86_SIZE_d64: /* Default to 64-bit in 64-bit mode */
*ot = CODE64(s) && s->dflag == MO_32 ? MO_64 : s->dflag; *ot = CODE64(s) && s->dflag == MO_32 ? MO_64 : s->dflag;
return true; return true;
@ -1342,6 +1348,8 @@ static bool has_cpuid_feature(DisasContext *s, X86CPUIDFeature cpuid)
switch (cpuid) { switch (cpuid) {
case X86_FEAT_None: case X86_FEAT_None:
return true; return true;
case X86_FEAT_F16C:
return (s->cpuid_ext_features & CPUID_EXT_F16C);
case X86_FEAT_MOVBE: case X86_FEAT_MOVBE:
return (s->cpuid_ext_features & CPUID_EXT_MOVBE); return (s->cpuid_ext_features & CPUID_EXT_MOVBE);
case X86_FEAT_PCLMULQDQ: case X86_FEAT_PCLMULQDQ:

View File

@ -92,6 +92,7 @@ typedef enum X86OpSize {
/* Custom */ /* Custom */
X86_SIZE_d64, X86_SIZE_d64,
X86_SIZE_f64, X86_SIZE_f64,
X86_SIZE_ph, /* SSE/AVX packed half precision */
} X86OpSize; } X86OpSize;
typedef enum X86CPUIDFeature { typedef enum X86CPUIDFeature {
@ -103,6 +104,7 @@ typedef enum X86CPUIDFeature {
X86_FEAT_AVX2, X86_FEAT_AVX2,
X86_FEAT_BMI1, X86_FEAT_BMI1,
X86_FEAT_BMI2, X86_FEAT_BMI2,
X86_FEAT_F16C,
X86_FEAT_MOVBE, X86_FEAT_MOVBE,
X86_FEAT_PCLMULQDQ, X86_FEAT_PCLMULQDQ,
X86_FEAT_SSE, X86_FEAT_SSE,

View File

@ -296,7 +296,7 @@ static void gen_writeback(DisasContext *s, X86DecodedInsn *decode, int opn, TCGv
case X86_OP_MMX: case X86_OP_MMX:
break; break;
case X86_OP_SSE: case X86_OP_SSE:
if (!op->has_ea && (s->prefix & PREFIX_VEX) && op->ot == MO_128) { if (!op->has_ea && (s->prefix & PREFIX_VEX) && op->ot <= MO_128) {
tcg_gen_gvec_dup_imm(MO_64, tcg_gen_gvec_dup_imm(MO_64,
offsetof(CPUX86State, xmm_regs[op->n].ZMM_X(1)), offsetof(CPUX86State, xmm_regs[op->n].ZMM_X(1)),
16, 16, 0); 16, 16, 0);
@ -852,6 +852,7 @@ UNARY_INT_SSE(VCVTTPD2DQ, cvttpd2dq)
UNARY_INT_SSE(VCVTDQ2PS, cvtdq2ps) UNARY_INT_SSE(VCVTDQ2PS, cvtdq2ps)
UNARY_INT_SSE(VCVTPS2DQ, cvtps2dq) UNARY_INT_SSE(VCVTPS2DQ, cvtps2dq)
UNARY_INT_SSE(VCVTTPS2DQ, cvttps2dq) UNARY_INT_SSE(VCVTTPS2DQ, cvttps2dq)
UNARY_INT_SSE(VCVTPH2PS, cvtph2ps)
static inline void gen_unary_imm_sse(DisasContext *s, CPUX86State *env, X86DecodedInsn *decode, static inline void gen_unary_imm_sse(DisasContext *s, CPUX86State *env, X86DecodedInsn *decode,
@ -1868,6 +1869,20 @@ static void gen_VCVTfp2fp(DisasContext *s, CPUX86State *env, X86DecodedInsn *dec
gen_helper_cvtsd2ss, gen_helper_cvtss2sd); gen_helper_cvtsd2ss, gen_helper_cvtss2sd);
} }
static void gen_VCVTPS2PH(DisasContext *s, CPUX86State *env, X86DecodedInsn *decode)
{
gen_unary_imm_fp_sse(s, env, decode,
gen_helper_cvtps2ph_xmm,
gen_helper_cvtps2ph_ymm);
/*
* VCVTPS2PH is the only instruction that performs an operation on a
* register source and then *stores* into memory.
*/
if (decode->op[0].has_ea) {
gen_store_sse(s, decode, decode->op[0].offset);
}
}
static void gen_VCVTSI2Sx(DisasContext *s, CPUX86State *env, X86DecodedInsn *decode) static void gen_VCVTSI2Sx(DisasContext *s, CPUX86State *env, X86DecodedInsn *decode)
{ {
int vec_len = vector_len(s, decode); int vec_len = vector_len(s, decode);

View File

@ -28,6 +28,7 @@ typedef struct {
} TestDef; } TestDef;
reg_state initI; reg_state initI;
reg_state initF16;
reg_state initF32; reg_state initF32;
reg_state initF64; reg_state initF64;
@ -221,6 +222,7 @@ static void run_all(void)
#define ARRAY_LEN(x) (sizeof(x) / sizeof(x[0])) #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}; 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}; double val_f64[] = {2.0, -1.0, 4.8, 0.8, 3, -42.0, 5e6, 7.5};
v4di val_i64[] = { v4di val_i64[] = {
@ -241,6 +243,12 @@ v4di indexd = {0x00000002000000efull, 0xfffffff500000010ull,
v4di gather_mem[0x20]; 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) void init_f32reg(v4di *r)
{ {
static int n; static int n;
@ -315,6 +323,15 @@ int main(int argc, char *argv[])
printf("Int:\n"); printf("Int:\n");
dump_regs(&initI); 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_all(&initF32);
init_f32reg(&initF32.ymm[10]); init_f32reg(&initF32.ymm[10]);
init_f32reg(&initF32.ymm[11]); init_f32reg(&initF32.ymm[11]);

View File

@ -9,6 +9,7 @@ from fnmatch import fnmatch
archs = [ archs = [
"SSE", "SSE2", "SSE3", "SSSE3", "SSE4_1", "SSE4_2", "SSE", "SSE2", "SSE3", "SSSE3", "SSE4_1", "SSE4_2",
"AES", "AVX", "AVX2", "AES+AVX", "VAES+AVX", "AES", "AVX", "AVX2", "AES+AVX", "VAES+AVX",
"F16C",
] ]
ignore = set(["FISTTP", ignore = set(["FISTTP",
@ -19,6 +20,7 @@ imask = {
'vBLENDPS': 0x0f, 'vBLENDPS': 0x0f,
'CMP[PS][SD]': 0x07, 'CMP[PS][SD]': 0x07,
'VCMP[PS][SD]': 0x1f, 'VCMP[PS][SD]': 0x1f,
'vCVTPS2PH': 0x7,
'vDPPD': 0x33, 'vDPPD': 0x33,
'vDPPS': 0xff, 'vDPPS': 0xff,
'vEXTRACTPS': 0x03, 'vEXTRACTPS': 0x03,
@ -221,8 +223,10 @@ def ArgGenerator(arg, op):
class InsnGenerator: class InsnGenerator:
def __init__(self, op, args): def __init__(self, op, args):
self.op = op self.op = op
if op[-2:] in ["PS", "PD", "SS", "SD"]: if op[-2:] in ["PH", "PS", "PD", "SS", "SD"]:
if op[-1] == 'S': if op[-1] == 'H':
self.optype = 'F16'
elif op[-1] == 'S':
self.optype = 'F32' self.optype = 'F32'
else: else:
self.optype = 'F64' self.optype = 'F64'