// return zero if we can; otherwise fill with zero and return -1.
#if CPUFAM_X86
+# define OUT edi
pushreg ebx
pushreg edi
- mov edi, [esp + 12]
- mov eax, [esp + 16]
- mov ecx, [esp + 20]
-# define OUT edi
+ mov OUT, [SP + 12]
+ mov eax, [SP + 16]
+ mov ecx, [SP + 20]
#endif
#if CPUFAM_AMD64 && ABI_SYSV
+# define OUT rdi
pushreg rbx
mov eax, esi
mov ecx, edx
-# define OUT rdi
#endif
#if CPUFAM_AMD64 && ABI_WIN
+# define OUT r9
pushreg rbx
- mov r9, rcx
+ mov OUT, rcx
mov eax, edx
mov ecx, r8d
-# define OUT r9
#endif
endprologue
// First, check that this is even a thing, using the complicated
- // dance with the flags register.
+ // dance with the flags register. This is unnecessary on AMD64,
+ // which postdates the introduction of `cpuid'.
+#if CPUFAM_X86
pushf
- pop R_d(r) // current flags in d
+ pop DX // current flags in d
- or R_d(r), EFLAGS_ID // force the id bit on and check it
- push R_d(r)
+ or DX, EFLAGS_ID // force the id bit on and check it
+ push DX
popf
pushf
- pop R_d(r)
+ pop DX
test edx, EFLAGS_ID
jz 8f
- and R_d(r), ~EFLAGS_ID // force the id bit off and check it
- push R_d(r)
+ and DX, ~EFLAGS_ID // force the id bit off and check it
+ push DX
popf
pushf
- pop R_d(r)
+ pop DX
test edx, EFLAGS_ID
jnz 8f
+#endif
// OK, that seemed to work.
cpuid
ret
// Failed.
+#if CPUFAM_X86
8: xor eax, eax
mov [OUT + 0], eax
mov [OUT + 4], eax
mov [OUT + 12], eax
mov eax, -1
jmp 9b
+#endif
+
+#undef OUT
ENDFUNC
///--------------------------------------------------------------------------
// Enter with no arguments. Return nonzero if the XMM registers are
// usable.
- pushreg R_bp(r)
+ pushreg BP
setfp
stalloc 512
- and R_sp(r), ~15
+ and SP, ~15
endprologue
// Save the floating point and SIMD registers, and try to clobber
// xmm0.
- fxsave [R_sp(r)]
- mov eax, [R_sp(r) + 160]
- xor dword ptr [R_sp(r) + 160], 0xaaaa5555
- fxrstor [R_sp(r)]
+ lea DX, [SP + 160]
+ fxsave [SP]
+ mov eax, [DX]
+ xor dword ptr [DX], 0xaaaa5555
+ fxrstor [SP]
// Save them again, and read back the low word of xmm0. Undo the
// clobbering and restore.
- fxsave [R_sp(r)]
- mov ecx, [R_sp(r) + 160]
- mov [R_sp(r) + 160], eax
- fxrstor [R_sp(r)]
+ fxsave [SP]
+ mov ecx, [DX]
+ mov [DX], eax
+ fxrstor [SP]
// The register are live if we read different things.
xor eax, ecx
// Done.
dropfp
- popreg R_bp(r)
+ popreg BP
+ ret
+ENDFUNC
+
+///--------------------------------------------------------------------------
+/// Checking extended control registers.
+
+FUNC(dispatch_x86ish_xgetbv)
+ // Call with two arguments: a pointer Z_OUT to 8 bytes of output space, and
+ // a 32-bit integer C. Read the 64-bit value of XCR(C), and store it
+ // at Z_OUT.
+
+#if CPUFAM_X86
+# define Z_OUT edi
+ pushreg edi
+ mov Z_OUT, [esp + 8]
+ mov ecx, [esp + 12]
+#endif
+#if CPUFAM_AMD64 && ABI_SYSV
+# define Z_OUT rdi
+ mov ecx, esi
+#endif
+#if CPUFAM_AMD64 && ABI_WIN
+# define Z_OUT r8
+ mov r8, rcx
+ mov ecx, edx
+#endif
+ endprologue
+
+ xgetbv
+ mov [Z_OUT + 0], eax
+ mov [Z_OUT + 4], edx
+
+#if CPUFAM_X86
+ popreg edi
+#endif
ret
+
+#undef Z_OUT
+ENDFUNC
+
+///--------------------------------------------------------------------------
+/// Checking `rdrand'.
+
+FUNC(dispatch_x86ish_rdrand)
+ // Enter with two arguments: a code OP requesting either `rdrand' (0)
+ // or `rdseed' (1), and a pointer X_OUT to a 32-bit word. Try to
+ // generate a random word using the requested instruction'. If
+ // successful, set *X_OUT to the generated word, and return zero;
+ // otherwise, return -1.
+
+#if CPUFAM_X86
+# define OP eax
+# define X_OUT edx
+# define COUNT ecx
+ mov OP, [SP + 4]
+ mov X_OUT, [SP + 8]
+#endif
+#if CPUFAM_AMD64 && ABI_SYSV
+# define OP edi
+# define X_OUT rsi
+# define COUNT ecx
+#endif
+#if CPUFAM_AMD64 && ABI_WIN
+# define OP rcx
+# define X_OUT rdx
+# define COUNT r8d
+#endif
+ endprologue
+
+ cmp OP, 0
+ mov COUNT, 16 // fairly persistent
+ jne 1f
+
+0: rdrand eax
+ jc 9f
+ dec COUNT
+ jnz 0b
+ jmp 8f
+
+1: rdseed eax
+ jc 9f
+ dec COUNT
+ jnz 1b
+ jmp 8f
+
+ // Failed to come up with a random value.
+8: mov eax, -1
+ ret
+
+ // Success.
+9: mov [X_OUT], eax
+ xor eax, eax
+ ret
+
+#undef X_OUT
+#undef COUNT
+
ENDFUNC
///----- That's all, folks --------------------------------------------------
~mdw / catacomb / blobdiff