xfnw/plan9fox
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

amd64: FP: always use enough to fit AVX state and align to 64 bytes

Browse source
This commit is contained in:
Sigrid 2020-12-06 19:31:56 +01:00
parent 66b6185845
commit 334c5e1134
4 changed files with 4 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

1
sys/src/9/pc64/dat.h
View file

@ -242,7 +242,6 @@ struct Mach
u64int dr7; /* shadow copy of dr7 */ u64int dr7; /* shadow copy of dr7 */
u64int xcr0; u64int xcr0;
u32int fpsavesz; u32int fpsavesz;
u32int fpalign;
void* vmx; void* vmx;

4
sys/src/9/pc64/fpu.c
View file

@ -16,8 +16,7 @@ fpuinit(void)
uintptr cr4; uintptr cr4;
ulong regs[4]; ulong regs[4];
m->fpsavesz = sizeof(FPssestate); m->fpsavesz = sizeof(FPsave); /* always enough to fit sse+avx */
m->fpalign = 16;
if((m->cpuiddx & (Sse|Fxsr)) == (Sse|Fxsr)){ /* have sse fp? */ if((m->cpuiddx & (Sse|Fxsr)) == (Sse|Fxsr)){ /* have sse fp? */
cr4 = getcr4() | CR4Osfxsr|CR4Oxmmex; cr4 = getcr4() | CR4Osfxsr|CR4Oxmmex;
putcr4(cr4); putcr4(cr4);
@ -34,7 +33,6 @@ fpuinit(void)
cpuid(0xd, 0, regs); cpuid(0xd, 0, regs);
m->fpsavesz = regs[1]; m->fpsavesz = regs[1];
m->fpalign = 64;
cpuid(0xd, 1, regs); cpuid(0xd, 1, regs);
if(regs[0] & Xsaveopt) if(regs[0] & Xsaveopt)

4
sys/src/9/pc64/main.c
View file

@ -488,7 +488,7 @@ mathemu(Ureg *ureg, void*)
up->fpstate |= FPkernel; up->fpstate |= FPkernel;
} }
while(up->fpslot[index] == nil) while(up->fpslot[index] == nil)
up->fpslot[index] = mallocalign(m->fpsavesz, m->fpalign, 0, 0); up->fpslot[index] = mallocalign(m->fpsavesz, FPalign, 0, 0);
up->fpsave = up->fpslot[index]; up->fpsave = up->fpslot[index];
up->fpstate = FPactive | (up->fpstate & (FPnouser|FPkernel|FPindexm)); up->fpstate = FPactive | (up->fpstate & (FPnouser|FPkernel|FPindexm));
break; break;
@ -574,7 +574,7 @@ procfork(Proc *p)
case FPinactive | FPpush: case FPinactive | FPpush:
case FPinactive: case FPinactive:
while(p->fpslot[0] == nil) while(p->fpslot[0] == nil)
p->fpslot[0] = mallocalign(m->fpsavesz, m->fpalign, 0, 0); p->fpslot[0] = mallocalign(m->fpsavesz, FPalign, 0, 0);
memmove(p->fpsave = p->fpslot[0], up->fpslot[0], m->fpsavesz); memmove(p->fpsave = p->fpslot[0], up->fpslot[0], m->fpsavesz);
p->fpstate = FPinactive; p->fpstate = FPinactive;
} }

1
sys/src/9/pc64/mem.h
View file

@ -26,6 +26,7 @@
#define ROUND(s, sz) (((s)+((sz)-1))&~((sz)-1)) #define ROUND(s, sz) (((s)+((sz)-1))&~((sz)-1))
#define PGROUND(s) ROUND(s, BY2PG) #define PGROUND(s) ROUND(s, BY2PG)
#define BLOCKALIGN 8 #define BLOCKALIGN 8
#define FPalign 64
#define MAXMACH 128 /* max # cpus system can run */ #define MAXMACH 128 /* max # cpus system can run */