tcg/i386: Detect AVX512

There are some operation sizes in some subsets of AVX512 that
are missing from previous iterations of AVX.  Detect them.

Tested-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This commit is contained in:
Richard Henderson 2021-12-15 18:16:36 -08:00
parent 21eab5bfae
commit ba597b66d9
3 changed files with 43 additions and 5 deletions

View File

@ -45,12 +45,26 @@
#ifndef bit_AVX2
#define bit_AVX2 (1 << 5)
#endif
#ifndef bit_AVX512F
#define bit_AVX512F (1 << 16)
#endif
#ifndef bit_BMI2
#define bit_BMI2 (1 << 8)
#endif
#ifndef bit_AVX512F
#define bit_AVX512F (1 << 16)
#endif
#ifndef bit_AVX512DQ
#define bit_AVX512DQ (1 << 17)
#endif
#ifndef bit_AVX512BW
#define bit_AVX512BW (1 << 30)
#endif
#ifndef bit_AVX512VL
#define bit_AVX512VL (1u << 31)
#endif
/* Leaf 7, %ecx */
#ifndef bit_AVX512VBMI2
#define bit_AVX512VBMI2 (1 << 6)
#endif
/* Leaf 0x80000001, %ecx */
#ifndef bit_LZCNT

View File

@ -171,6 +171,10 @@ bool have_bmi1;
bool have_popcnt;
bool have_avx1;
bool have_avx2;
bool have_avx512bw;
bool have_avx512dq;
bool have_avx512vbmi2;
bool have_avx512vl;
bool have_movbe;
#ifdef CONFIG_CPUID_H
@ -3839,12 +3843,12 @@ static void tcg_out_nop_fill(tcg_insn_unit *p, int count)
static void tcg_target_init(TCGContext *s)
{
#ifdef CONFIG_CPUID_H
unsigned a, b, c, d, b7 = 0;
unsigned a, b, c, d, b7 = 0, c7 = 0;
unsigned max = __get_cpuid_max(0, 0);
if (max >= 7) {
/* BMI1 is available on AMD Piledriver and Intel Haswell CPUs. */
__cpuid_count(7, 0, a, b7, c, d);
__cpuid_count(7, 0, a, b7, c7, d);
have_bmi1 = (b7 & bit_BMI) != 0;
have_bmi2 = (b7 & bit_BMI2) != 0;
}
@ -3874,6 +3878,22 @@ static void tcg_target_init(TCGContext *s)
if ((xcrl & 6) == 6) {
have_avx1 = (c & bit_AVX) != 0;
have_avx2 = (b7 & bit_AVX2) != 0;
/*
* There are interesting instructions in AVX512, so long
* as we have AVX512VL, which indicates support for EVEX
* on sizes smaller than 512 bits. We are required to
* check that OPMASK and all extended ZMM state are enabled
* even if we're not using them -- the insns will fault.
*/
if ((xcrl & 0xe0) == 0xe0
&& (b7 & bit_AVX512F)
&& (b7 & bit_AVX512VL)) {
have_avx512vl = true;
have_avx512bw = (b7 & bit_AVX512BW) != 0;
have_avx512dq = (b7 & bit_AVX512DQ) != 0;
have_avx512vbmi2 = (c7 & bit_AVX512VBMI2) != 0;
}
}
}
}

View File

@ -103,6 +103,10 @@ extern bool have_bmi1;
extern bool have_popcnt;
extern bool have_avx1;
extern bool have_avx2;
extern bool have_avx512bw;
extern bool have_avx512dq;
extern bool have_avx512vbmi2;
extern bool have_avx512vl;
extern bool have_movbe;
/* optional instructions */