sse kernel support (sources)

sys/src/9/pc/dat.h

@@ -2,7 +2,9 @@ typedef struct BIOS32si	BIOS32si;
 typedef struct BIOS32ci	BIOS32ci;
 typedef struct Conf	Conf;
 typedef struct Confmem	Confmem;
-typedef struct FPsave	FPsave;
+typedef union FPsave	FPsave;
+typedef struct FPssestate FPssestate;
+typedef struct FPstate	FPstate;
 typedef struct ISAConf	ISAConf;
 typedef struct Label	Label;
 typedef struct Lock	Lock;
@@ -64,7 +66,7 @@ enum
 	FPillegal=	0x100,
 };
 
-struct	FPsave
+struct	FPstate
 {
 	ushort	control;
 	ushort	r1;
@@ -81,6 +83,33 @@ struct	FPsave
 	uchar	regs[80];	/* floating point registers */
 };
 
+struct	FPssestate		/* SSE fp state */
+{
+	ushort	fcw;		/* control */
+	ushort	fsw;		/* status */
+	ushort	ftw;		/* tag */
+	ushort	fop;		/* opcode */
+	ulong	fpuip;		/* pc */
+	ushort	cs;		/* pc segment */
+	ushort	r1;		/* reserved */
+	ulong	fpudp;		/* data pointer */
+	ushort	ds;		/* data pointer segment */
+	ushort	r2;
+	ulong	mxcsr;		/* MXCSR register state */
+	ulong	mxcsr_mask;	/* MXCSR mask register */
+	uchar	xregs[480];	/* extended registers */
+	uchar	alignpad[FPalign];
+};
+
+/*
+ * the FP regs must be stored here, not somewhere pointed to from here.
+ * port code assumes this.
+ */
+union FPsave {
+	FPstate;
+	FPssestate;
+};
+
 struct Confmem
 {
 	ulong	base;
@@ -227,6 +256,7 @@ struct Mach
 	uvlong	tscticks;
 	int	pdballoc;
 	int	pdbfree;
+	FPsave	*fpsavalign;
 
 	vlong	mtrrcap;
 	vlong	mtrrdef;
@@ -297,6 +327,7 @@ enum {
 	Clflush = 1<<19,
 	Acpif	= 1<<22,	/* therm control msr */
 	Mmx	= 1<<23,
+	Fxsr	= 1<<24,	/* have SSE FXSAVE/FXRSTOR */
 	Sse	= 1<<25,	/* thus sfence instr. */
 	Sse2	= 1<<26,	/* thus mfence & lfence instr.s */
 	Rdrnd	= 1<<30,	/* RDRAND support bit */

sys/src/9/pc/devarch.c

@@ -38,6 +38,11 @@ enum {
 
 	Qmax = 16,
 };
+
+enum {
+	CR4Osfxsr = 1 << 9,
+};
+
 enum {				/* cpuid standard function codes */
 	Highstdfunc = 0,	/* also returns vendor string */
 	Procsig,
@@ -850,6 +855,15 @@ cpuidentify(void)
 			rdmsr(0x01, &mct);
 	}
 
+	if(m->cpuiddx & Fxsr){			/* have sse fp? */
+		fpsave = fpssesave;
+		fprestore = fpsserestore;
+		putcr4(getcr4() | CR4Osfxsr);
+	} else {
+		fpsave = fpx87save;
+		fprestore = fpx87restore;
+	}
+
 	cputype = t;
 	return t->family;
 }

sys/src/9/pc/fns.h

@@ -33,9 +33,15 @@ void	fpclear(void);
 void	fpenv(FPsave*);
 void	fpinit(void);
 void	fpoff(void);
-void	fprestore(FPsave*);
-void	fpsave(FPsave*);
+void	(*fprestore)(FPsave*);
+void	(*fpsave)(FPsave*);
+void	fpsserestore(FPsave*);
+void	fpsserestore0(FPsave*);
+void	fpssesave(FPsave*);
+void	fpssesave0(FPsave*);
 ulong	fpstatus(void);
+void	fpx87restore(FPsave*);
+void	fpx87save(FPsave*);
 ulong	getcr0(void);
 ulong	getcr2(void);
 ulong	getcr3(void);

sys/src/9/pc/l.s

@@ -657,13 +657,13 @@ TEXT fpinit(SB), $0				/* enable and init */
 	WAIT
 	RET
 
-TEXT fpsave(SB), $0				/* save state and disable */
+TEXT fpx87save(SB), $0				/* save state and disable */
 	MOVL	p+0(FP), AX
 	FSAVE	0(AX)				/* no WAIT */
 	FPOFF(l2)
 	RET
 
-TEXT fprestore(SB), $0				/* enable and restore state */
+TEXT fpx87restore(SB), $0				/* enable and restore state */
 	FPON
 	MOVL	p+0(FP), AX
 	FRSTOR	0(AX)
@@ -685,6 +685,19 @@ TEXT fpclear(SB), $0				/* clear pending exceptions */
 	FPOFF(l3)
 	RET
 
+TEXT fpssesave0(SB), $0				/* save state and disable */
+	MOVL	p+0(FP), AX
+	FXSAVE	0(AX)				/* no WAIT */
+	FPOFF(l4)
+	RET
+
+TEXT fpsserestore0(SB), $0			/* enable and restore state */
+	FPON
+	MOVL	p+0(FP), AX
+	FXRSTOR	0(AX)
+	WAIT
+	RET
+
 /*
  */
 TEXT splhi(SB), $0

sys/src/9/pc/main.c

@@ -479,14 +479,37 @@ static char* mathmsg[] =
 	"precision loss",
 };
 
+static void
+mathstate(ulong *stsp, ulong *pcp, ulong *ctlp)
+{
+	ulong sts, fpc, ctl;
+	FPsave *f = &up->fpsave;
+
+	if(fpsave == fpx87save){
+		sts = f->status;
+		fpc = f->pc;
+		ctl = f->control;
+	} else {
+		sts = f->fsw;
+		fpc = f->fpuip;
+		ctl = f->fcw;
+	}
+	if(stsp)
+		*stsp = sts;
+	if(pcp)
+		*pcp = fpc;
+	if(ctlp)
+		*ctlp = ctl;
+}
+
 static void
 mathnote(void)
 {
 	int i;
-	ulong status;
+	ulong status, pc;
 	char *msg, note[ERRMAX];
 
-	status = up->fpsave.status;
+	mathstate(&status, &pc, nil);
 
 	/*
 	 * Some attention should probably be paid here to the
@@ -513,12 +536,50 @@ mathnote(void)
 	postnote(up, 1, note, NDebug);
 }
 
+/*
+ * sse fp save and restore buffers have to be 16-byte (FPalign) aligned,
+ * so we shuffle the data up and down as needed or make copies.
+ */
+
+void
+fpssesave(FPsave *fps)
+{
+	FPsave *afps;
+
+	afps = (FPsave *)ROUND(((uintptr)fps), FPalign);
+	fpssesave0(afps);
+	if (fps != afps)  /* not aligned? shuffle down from aligned buffer */
+		memmove(fps, afps, sizeof(FPssestate) - FPalign);
+}
+
+void
+fpsserestore(FPsave *fps)
+{
+	FPsave *afps;
+
+	afps = (FPsave *)ROUND(((uintptr)fps), FPalign);
+	if (fps != afps) {
+		if (m->fpsavalign == nil)
+			m->fpsavalign = mallocalign(sizeof(FPssestate),
+				FPalign, 0, 0);
+		if (m->fpsavalign)
+			afps = m->fpsavalign;
+		/* copy or shuffle up to make aligned */
+		memmove(afps, fps, sizeof(FPssestate) - FPalign);
+	}
+	fpsserestore0(afps);
+	/* if we couldn't make a copy, shuffle regs back down */
+	if (fps != afps && afps != m->fpsavalign)
+		memmove(fps, afps, sizeof(FPssestate) - FPalign);
+}
+
 /*
  *  math coprocessor error
  */
 static void
 matherror(Ureg *ur, void*)
 {
+	ulong status, pc;
 	/*
 	 *  a write cycle to port 0xF0 clears the interrupt latch attached
 	 *  to the error# line from the 387
@@ -532,9 +593,11 @@ matherror(Ureg *ur, void*)
 	fpenv(&up->fpsave);
 	mathnote();
 
-	if((ur->pc & 0xf0000000) == KZERO)
+	if((ur->pc & 0xf0000000) == KZERO){
+		mathstate(&status, &pc, nil);
 		panic("fp: status %ux fppc=0x%lux pc=0x%lux",
 			up->fpsave.status, up->fpsave.pc, ur->pc);
+	}
 }
 
 /*
@@ -543,6 +606,8 @@ matherror(Ureg *ur, void*)
 static void
 mathemu(Ureg *ureg, void*)
 {
+	ulong status, control;
+
 	if(up->fpstate & FPillegal){
 		/* someone did floating point in a note handler */
 		postnote(up, 1, "sys: floating point in note handler", NDebug);
@@ -561,7 +626,8 @@ mathemu(Ureg *ureg, void*)
 		 * More attention should probably be paid here to the
 		 * exception masks and error summary.
 		 */
-		if((up->fpsave.status & ~up->fpsave.control) & 0x07F){
+		mathstate(&status, nil, &control);
+		if((status & ~control) & 0x07F){
 			mathnote();
 			break;
 		}

sys/src/9/pc/mem.h

@@ -19,6 +19,7 @@
 #define	ROUND(s, sz)	(((s)+((sz)-1))&~((sz)-1))
 #define	PGROUND(s)	ROUND(s, BY2PG)
 #define	BLOCKALIGN	8
+#define	FPalign		16
 
 /*
  * In 32-bit mode, the MAXMACH limit is 32 without

sys/src/cmd/8l/optab.c

@@ -434,12 +434,6 @@ uchar	ymskb[] =
 	Ymr,	Yrl,	Zm_r_xm,	1,
 	0
 };
-uchar	yxaes[] =
-{
-	Yxm,	Yxr,	Zm_r_xm,	2,
-	Yxm,	Yxr,	Zm_r_i_xm,	2,
-	0
-};
 
 Optab optab[] =
 /*	as, ytab, andproto, opcode */