s/fpu_curproc/fpu_curpcb/ which also points into physical.

move it's definition into locore (where it's used the most).
adjust fpu context switch and save on swapout accordingly.
fix context save (in switch).
add accounting for fpu context switches (since it differes
from proc switching, we might get provided w/ some usefull
numbers as to how effective this lazy switch technique is).
enable handling fpu/sfu exceptions.

3 files changed, 73 insertions, 34 deletions

diff --git a/sys/arch/hppa/hppa/locore.S b/sys/arch/hppa/hppa/locore.S
index 9d05a718c9b..c71f4dcf2f5 100644
--- a/sys/arch/hppa/hppa/locore.S
+++ b/sys/arch/hppa/hppa/locore.S
@@ -1,4 +1,4 @@
-/*	$OpenBSD: locore.S,v 1.23 1999/12/31 09:07:10 mickey Exp $	*/
+/*	$OpenBSD: locore.S,v 1.24 2000/01/11 20:27:57 mickey Exp $	*/
 
 /*
  * Copyright (c) 1998,1999 Michael Shalayeff
@@ -86,7 +86,6 @@
 	.import	bootdev, data
 	.import	esym, data
 	.import	curproc, data
-	.import	fpu_curproc, data
 	.import	want_resched, data
 	.import virtual_avail, data
 	.import	proc0, data
@@ -101,6 +100,12 @@ pdc_stack
 	.block	3*NBPG
 kernelmapped			/* set when kernel is mapped */
 	.block	4
+	.export fpu_csw, data
+fpu_csw
+	.block	4
+	.export	fpu_curpcb, data
+fpu_curpcb
+	.block	4
 
 /*
  * This is the starting location for the kernel
@@ -700,9 +705,9 @@ $syscall_return
 $syscall_return_copy_loop
 	ldw	0(arg0), t1
 	ldw	4(arg0), t2
-	ldo	8(arg0), arg0
 	stw	t1, 0(arg1)
 	stw	t2, 4(arg1)
+	ldo	8(arg0), arg0
 	comb,<,n arg0, arg2, $syscall_return_copy_loop
 	ldo	8(arg1), arg1
 
@@ -871,7 +876,7 @@ $ivaaddr
 	ATRAP(privr,T_PRIV_REG)		/* 11. privileged register trap */
 	ATRAP(ovrfl,T_OVERFLOW)		/* 12. overflow trap */
 	ATRAP(cond,T_CONDITION)		/* 13. conditional trap */
-#ifdef FPEMUL_notyet
+#ifdef FPEMUL
 	CTRAP(excpt,T_EXCEPTION,)	/* 14. assist exception trap */
 #else
 	ATRAP(excpt,T_EXCEPTION)
@@ -952,14 +957,12 @@ TLABEL(emu)
 	mtctl	t2, tr3
 	mtctl	t3, tr5
 
-	ldil	L%fpu_curproc, t1
-	ldw	R%fpu_curproc(t1), t1
+	ldil	L%fpu_curpcb, t1
+	ldw	R%fpu_curpcb(t1), t1
 
 	mfctl	ccr, t3
-
-	ldil	L%curproc, t2
-	ldw	R%curproc(t2), t2
-
+	mfctl	cr30, t2
+	
 	/* enable coprocessor */
 	depi	3, 25, 2, t3
 	mtctl	t3, ccr
@@ -967,8 +970,7 @@ TLABEL(emu)
 	comb,=,n t1, t2, $fpusw_done
 	comb,=,n r0, t1, $fpusw_nosave
 
-	ldw	p_addr(t1), t3
-	ldo	pcb_fpregs+u_pcb(t3), t3
+	/* ldo	pcb_fpregs+u_pcb(t1), t1 */
 
 	fstds,ma fr0 , 8(t3)	/* fr0 must be saved first */
 	fstds,ma fr1 , 8(t3)
@@ -1001,12 +1003,17 @@ TLABEL(emu)
 	fstds,ma fr28, 8(t3)
 	fstds,ma fr29, 8(t3)
 	fstds,ma fr30, 8(t3)
-	fstds    fr31, 8(t3)
+	fstds    fr31, 0(t3)
 
 $fpusw_nosave
 
-	ldw	p_addr(t2), t3
-	ldo	31*8+pcb_fpregs+u_pcb(t3), t3
+	/* count switches */
+	ldil	L%fpu_csw, t1
+	ldw	R%fpu_csw(t1), t3
+	ldo	1(t3), t3
+	stw	t3, R%fpu_csw(t1)
+
+	ldo	31*8+pcb_fpregs+u_pcb(t2), t3
 
 	fldds,ma -8(t3), fr31
 	fldds,ma -8(t3), fr30
@@ -1041,8 +1048,8 @@ $fpusw_nosave
 	fldds,ma -8(t3), fr1
 	fldds     0(t3), fr0	/* fr0 must be restored last */
 
-	ldil	L%fpu_curproc, t1
-	stw	t2, R%fpu_curproc(t1)
+	ldil	L%fpu_curpcb, t1
+	stw	t2, R%fpu_curpcb(t1)
 
 $fpusw_done
 	mfctl	tr5, t3
@@ -1051,7 +1058,7 @@ $fpusw_done
 	rfi
 	nop
 
-#ifdef notyet
+#ifdef FPEMUL
 	.export TLABEL(excpt), entry
 	/*
 	 * Emulate FPU/SFU if none/disabled
@@ -1062,7 +1069,7 @@ TLABEL(excpt)
 	mtctl	sp, tr3
 	mtctl	r31, tr2
 
-	.import	$fpemu_stack
+	.import	$fpemu_stack, data
 	ldil	L%$fpemu_stack, r31
 	ldo	R%$fpemu_stack(r31), r31
 	ldo	R%TF_SIZE+HPPA_FRAME_SIZE(r31), sp
@@ -1093,6 +1100,7 @@ TLABEL(excpt)
 	ldil	L%fpu_emulate,t1
 	ldo	R%fpu_emulate(t1),t1
 	mfctl	iir, arg0
+	/* arg3 -- regs */
 	.call
 	blr	r0,rp
 	bv,n	0(t1)
@@ -1125,13 +1133,13 @@ TLABEL(excpt)
 	mfctl	tr7, r1
 	rfi
 	nop
-#endif
 	
 	.export	$sfu_emu, entry
 $sfu_emu
+	bv	r0(rp)
 	ldo	1(r0), ret0	/* none supported by now */
-	rfi
-	nop
+
+#endif /* FPEMUL */
 
 	/* Compute the hpt entry ptr */
 #define	HPTENT \
diff --git a/sys/arch/hppa/hppa/machdep.c b/sys/arch/hppa/hppa/machdep.c
index dee4a372ddb..f6e35cf96bc 100644
--- a/sys/arch/hppa/hppa/machdep.c
+++ b/sys/arch/hppa/hppa/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.17 1999/12/31 09:07:13 mickey Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.18 2000/01/11 20:27:56 mickey Exp $	*/
 
 /*
  * Copyright (c) 1999 Michael Shalayeff
@@ -147,7 +147,6 @@ int	totalphysmem, resvmem, physmem, esym;
  * Things for MI glue to stick on.
  */
 struct user *proc0paddr;
-struct proc *fpu_curproc;
 long mem_ex_storage[EXTENT_FIXED_STORAGE_SIZE(8) / sizeof(long)];
 struct extent *hppa_ex;
 
diff --git a/sys/arch/hppa/hppa/vm_machdep.c b/sys/arch/hppa/hppa/vm_machdep.c
index 37bc642176e..a6fa790437d 100644
--- a/sys/arch/hppa/hppa/vm_machdep.c
+++ b/sys/arch/hppa/hppa/vm_machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: vm_machdep.c,v 1.11 1999/12/31 09:07:13 mickey Exp $	*/
+/*	$OpenBSD: vm_machdep.c,v 1.12 2000/01/11 20:27:57 mickey Exp $	*/
 
 /*
  * Copyright (c) 1999 Michael Shalayeff
@@ -135,17 +135,49 @@ void
 cpu_swapout(p)
 	struct proc *p;
 {
-	extern struct proc *fpu_curproc;
-	struct proc *q = fpu_curproc;
+	extern paddr_t fpu_curpcb;
+	paddr_t q = fpu_curpcb;
 
-	fpu_curproc = NULL;
+	fpu_curpcb = 0;
 
 	/*
-	 * TODO: explicit FPU save state if we own the FPU,
-	 * since user area might get swapped out as well,
-	 * and we won't be able to save it no more
+	 * TODO: determine if we have an fpu
 	 */
-	if (p == q) {
+	if (kvtop((caddr_t)&p->p_addr->u_pcb) == q) {
+		__asm __volatile(
+		    "fstds,ma %%fr0 , 8(%0)\n\t"
+		    "fstds,ma %%fr1 , 8(%0)\n\t"
+		    "fstds,ma %%fr2 , 8(%0)\n\t"
+		    "fstds,ma %%fr3 , 8(%0)\n\t"
+		    "fstds,ma %%fr4 , 8(%0)\n\t"
+		    "fstds,ma %%fr5 , 8(%0)\n\t"
+		    "fstds,ma %%fr6 , 8(%0)\n\t"
+		    "fstds,ma %%fr7 , 8(%0)\n\t"
+		    "fstds,ma %%fr8 , 8(%0)\n\t"
+		    "fstds,ma %%fr9 , 8(%0)\n\t"
+		    "fstds,ma %%fr10, 8(%0)\n\t"
+		    "fstds,ma %%fr11, 8(%0)\n\t"
+		    "fstds,ma %%fr12, 8(%0)\n\t"
+		    "fstds,ma %%fr13, 8(%0)\n\t"
+		    "fstds,ma %%fr14, 8(%0)\n\t"
+		    "fstds,ma %%fr15, 8(%0)\n\t"
+		    "fstds,ma %%fr16, 8(%0)\n\t"
+		    "fstds,ma %%fr17, 8(%0)\n\t"
+		    "fstds,ma %%fr18, 8(%0)\n\t"
+		    "fstds,ma %%fr19, 8(%0)\n\t"
+		    "fstds,ma %%fr20, 8(%0)\n\t"
+		    "fstds,ma %%fr21, 8(%0)\n\t"
+		    "fstds,ma %%fr22, 8(%0)\n\t"
+		    "fstds,ma %%fr23, 8(%0)\n\t"
+		    "fstds,ma %%fr24, 8(%0)\n\t"
+		    "fstds,ma %%fr25, 8(%0)\n\t"
+		    "fstds,ma %%fr26, 8(%0)\n\t"
+		    "fstds,ma %%fr27, 8(%0)\n\t"
+		    "fstds,ma %%fr28, 8(%0)\n\t"
+		    "fstds,ma %%fr29, 8(%0)\n\t"
+		    "fstds,ma %%fr30, 8(%0)\n\t"
+		    "fstds    %%fr31, 0(%0)\n\t"
+		    : "+r" (q) :: "memory");
 	}
 }
 
@@ -226,13 +258,13 @@ void
 cpu_exit(p)
 	struct proc *p;
 {
-	extern struct proc *fpu_curproc;	/* from machdep.c */
+	extern paddr_t fpu_curpcb;	/* from locore.S */
 
 	uvmexp.swtch++;
 
 	splhigh();
 	curproc = NULL;
-	fpu_curproc = NULL;
+	fpu_curpcb = 0;
 	uvmspace_free(p->p_vmspace);
 
 	/* XXX should be in the locore? */