Rewrite cpu_switch(). The most notable change is the fact that we now

have f16-f31 as part of the context. The PCB has been reorganized to
better match how we save and restore the (preserved) registers. This
commit also moves the context restoriation to its own function (named
pcb_restore), as we did with pcb_save.

Only minimal effort has been put in writing optimal assembly. The
expectation is that there will be more rounds of changes.

7 files changed, 364 insertions, 303 deletions

diff --git a/sys/ia64/ia64/genassym.c b/sys/ia64/ia64/genassym.c
index 8dbc394..4a04842 100644
--- a/sys/ia64/ia64/genassym.c
+++ b/sys/ia64/ia64/genassym.c
@@ -112,36 +112,10 @@ ASSYM(TF_R_R14,		offsetof(struct trapframe, tf_r[FRAME_R14]));
 ASSYM(TF_R_R15,		offsetof(struct trapframe, tf_r[FRAME_R15]));
 ASSYM(TF_F,		offsetof(struct trapframe, tf_f));
 
-ASSYM(PCB_R4,		offsetof(struct pcb, pcb_r4));
-ASSYM(PCB_R5,		offsetof(struct pcb, pcb_r5));
-ASSYM(PCB_R6,		offsetof(struct pcb, pcb_r6));
-ASSYM(PCB_R7,		offsetof(struct pcb, pcb_r7));
-
-ASSYM(PCB_F2,		offsetof(struct pcb, pcb_f2));
-ASSYM(PCB_F3,		offsetof(struct pcb, pcb_f3));
-ASSYM(PCB_F4,		offsetof(struct pcb, pcb_f4));
-ASSYM(PCB_F5,		offsetof(struct pcb, pcb_f5));
-
-ASSYM(PCB_B0,		offsetof(struct pcb, pcb_b0));
-ASSYM(PCB_B1,		offsetof(struct pcb, pcb_b1));
-ASSYM(PCB_B2,		offsetof(struct pcb, pcb_b2));
-ASSYM(PCB_B3,		offsetof(struct pcb, pcb_b3));
-ASSYM(PCB_B4,		offsetof(struct pcb, pcb_b4));
-ASSYM(PCB_B5,		offsetof(struct pcb, pcb_b5));
-
-ASSYM(PCB_OLD_UNAT,	offsetof(struct pcb, pcb_old_unat));
-ASSYM(PCB_SP,		offsetof(struct pcb, pcb_sp));
-ASSYM(PCB_PFS,		offsetof(struct pcb, pcb_pfs));
-ASSYM(PCB_BSPSTORE,	offsetof(struct pcb, pcb_bspstore));
-
-ASSYM(PCB_UNAT,		offsetof(struct pcb, pcb_unat));
-ASSYM(PCB_RNAT,		offsetof(struct pcb, pcb_rnat));
-ASSYM(PCB_PR,		offsetof(struct pcb, pcb_pr));
-ASSYM(PCB_PMAP,		offsetof(struct pcb, pcb_pmap));
-
+ASSYM(PCB_CURRENT_PMAP,	offsetof(struct pcb, pcb_current_pmap));
 ASSYM(PCB_ONFAULT,	offsetof(struct pcb, pcb_onfault));
-
-ASSYM(PCB_HIGHFP,	offsetof(struct pcb, pcb_highfp));
+ASSYM(PCB_RP,		offsetof(struct pcb, pcb_rp));
+ASSYM(PCB_UNAT47,	offsetof(struct pcb, pcb_unat47));
 
 ASSYM(UC_MCONTEXT_MC_AR_BSP,  offsetof(ucontext_t, uc_mcontext.mc_ar_bsp));
 ASSYM(UC_MCONTEXT_MC_AR_RNAT, offsetof(ucontext_t, uc_mcontext.mc_ar_rnat));
diff --git a/sys/ia64/ia64/locore.S b/sys/ia64/ia64/locore.S
index c7dc7a7..726586d 100644
--- a/sys/ia64/ia64/locore.S
+++ b/sys/ia64/ia64/locore.S
@@ -120,7 +120,7 @@ ENTRY(__start, 1)
 	;;
 	ld8	out0=[out0]
 	;; 
-	add	r16=PCB_B0,out0		// return to mi_startup_trampoline
+	add	r16=PCB_RP,out0		// return to mi_startup_trampoline
 	movl	r17=mi_startup_trampoline
 	;;
 	st8	[r16]=r17
diff --git a/sys/ia64/ia64/locore.s b/sys/ia64/ia64/locore.s
index c7dc7a7..726586d 100644
--- a/sys/ia64/ia64/locore.s
+++ b/sys/ia64/ia64/locore.s
@@ -120,7 +120,7 @@ ENTRY(__start, 1)
 	;;
 	ld8	out0=[out0]
 	;; 
-	add	r16=PCB_B0,out0		// return to mi_startup_trampoline
+	add	r16=PCB_RP,out0		// return to mi_startup_trampoline
 	movl	r17=mi_startup_trampoline
 	;;
 	st8	[r16]=r17
diff --git a/sys/ia64/ia64/machdep.c b/sys/ia64/ia64/machdep.c
index 4f739d2..030f8629 100644
--- a/sys/ia64/ia64/machdep.c
+++ b/sys/ia64/ia64/machdep.c
@@ -704,7 +704,7 @@ ia64_init(u_int64_t arg1, u_int64_t arg2)
 	 */
 	thread0.td_frame = (struct trapframe *)thread0.td_pcb - 1;
 	thread0.td_pcb->pcb_sp = (u_int64_t)thread0.td_frame - 16;
-	thread0.td_pcb->pcb_bspstore = (u_int64_t)proc0kstack;
+	thread0.td_pcb->pcb_ar_bsp = (u_int64_t)proc0kstack;
 
 	mutex_init();
 
@@ -1261,18 +1261,26 @@ fill_fpregs(td, fpregs)
 	struct thread *td;
 	struct fpreg *fpregs;
 {
-	fpregs->fpr_regs[2] = td->td_pcb->pcb_f2;
-	fpregs->fpr_regs[3] = td->td_pcb->pcb_f3;
-	fpregs->fpr_regs[4] = td->td_pcb->pcb_f4;
-	fpregs->fpr_regs[5] = td->td_pcb->pcb_f5;
+	struct pcb *pcb = td->td_pcb;
 
-	bcopy(td->td_frame->tf_f, fpregs->fpr_regs+6,
+	/*
+	 * XXX - The PCB pointer should not point to the actual PCB,
+	 * because it will not contain the preserved registers of
+	 * the program being debugged. Instead, it should point to
+	 * a PCB constructed by unwinding all the way up to the
+	 * IVT handler.
+	 */
+	bcopy(pcb->pcb_f + PCB_F2, fpregs->fpr_regs + 2,
+	    sizeof(pcb->pcb_f[0]) * 4);
+
+	bcopy(td->td_frame->tf_f, fpregs->fpr_regs + 6,
 	    sizeof(td->td_frame->tf_f));
 
-	/* XXX f16-f31 */
+	bcopy(pcb->pcb_f + PCB_F16, fpregs->fpr_regs + 16,
+	    sizeof(pcb->pcb_f[0]) * 16);
 
 	ia64_fpstate_save(td, 0);
-	bcopy(td->td_pcb->pcb_highfp, fpregs->fpr_regs+32,
+	bcopy(pcb->pcb_highfp, fpregs->fpr_regs + 32,
 	    sizeof(td->td_pcb->pcb_highfp));
 
 	return (0);
@@ -1283,18 +1291,30 @@ set_fpregs(td, fpregs)
 	struct thread *td;
 	struct fpreg *fpregs;
 {
-	td->td_pcb->pcb_f2 = fpregs->fpr_regs[2];
-	td->td_pcb->pcb_f3 = fpregs->fpr_regs[3];
-	td->td_pcb->pcb_f4 = fpregs->fpr_regs[4];
-	td->td_pcb->pcb_f5 = fpregs->fpr_regs[5];
+	struct pcb *pcb = td->td_pcb;
+
+	/*
+	 * XXX - The PCB pointer should not point to the actual PCB,
+	 * because it will not contain the preserved registers of
+	 * the program being debugged. Instead, it should point to
+	 * a PCB constructed by unwinding all the way up to the
+	 * IVT handler.
+	 * XXX - An additional complication here is that we need to
+	 * have the actual location of where the values should be
+	 * stored as well. Some values may still reside in registers,
+	 * while other may have been saved somewhere.
+	 */
+	bcopy(fpregs->fpr_regs + 2, pcb->pcb_f + PCB_F2,
+	    sizeof(pcb->pcb_f[0]) * 4);
 
-	bcopy(fpregs->fpr_regs+6, td->td_frame->tf_f,
+	bcopy(fpregs->fpr_regs + 6, td->td_frame->tf_f,
 	    sizeof(td->td_frame->tf_f));
 
-	/* XXX f16-f31 */
+	bcopy(fpregs->fpr_regs + 16, pcb->pcb_f + PCB_F16,
+	    sizeof(pcb->pcb_f[0]) * 16);
 
 	ia64_fpstate_drop(td);
-	bcopy(fpregs->fpr_regs+32, td->td_pcb->pcb_highfp,
+	bcopy(fpregs->fpr_regs + 32, pcb->pcb_highfp,
 	    sizeof(td->td_pcb->pcb_highfp));
 
 	return (0);
diff --git a/sys/ia64/ia64/swtch.s b/sys/ia64/ia64/swtch.s
index da1e270..b17c92a 100644
--- a/sys/ia64/ia64/swtch.s
+++ b/sys/ia64/ia64/swtch.s
@@ -31,175 +31,304 @@
 #include "assym.s"
 
 /*
- * pcb_save: save process context, i.e. callee-saved registers. This
+ * pcb_save - save process context, i.e. callee-saved registers. This
  * function is called by savectx() and cpu_switch().
  *
+ * The callee-saved (ie preserved) registers are:
+ *	r4-r7
+ *	f2-f5, f16-f31
+ *	p1-p5, p16-p63
+ *	b1-b5
+ *	ar.unat, ar.lc
+ *
+ * Notes:
+ * o  We don't need to save and restore gp because its value is constant.
+ * o  We don't need to save and restore ar.rsc because its value is known.
+ * o  We don't need to save and restore r13 because it always points to
+ *    the pcpu structure.
+ *
  * Arguments:
- *	in0	'struct pcb *' of the process that needs its context saved.
- *	in1	return address.
+ *	r8	'struct pcb *' of the process that needs its context saved.
+ *	b6	return address.
  */
-
-ENTRY(pcb_save,2)
-{	.mmi
-	flushrs				// push out caller's dirty regs
-	mov	r22=ar.unat		// caller's value for ar.unat
-	mov	r16=in0
+ENTRY(pcb_save,0)
+{	.mii
+	mov	r16=ar.unat
+	add	r9=8,r8
+	mov	r17=b0
 	;;
 }
 {	.mmi
-	mov	ar.rsc=0		// stop the RSE after the flush
-	;;
-	mov	r22=ar.rnat		// read RSE's NaT collection
-	add	r17=8,r16
+	st8	[r8]=r12,16		// sp
+	st8	[r9]=r16,16		// ar.unat
+	mov	r18=pr
 	;;
 }
 {	.mmi
-	.mem.offset	0,0
-	st8.spill [r16]=r4,16		// save r4, ofs=&r6
-	.mem.offset	8,0
-	st8.spill [r17]=r5,16		// save r5, ofs=&r7
-	mov	r18=b0
+	st8	[r8]=r17,16		// rp
+	st8	[r9]=r18,24		// pr
+	add	r31=PC_CURRENT_PMAP,r13
 	;;
 }
-{	.mmi
-	.mem.offset	16,0
-	st8.spill [r16]=r6,16		// save r6, ofs=&f2	
-	.mem.offset	24,0
-	st8.spill [r17]=r7,24		// save r7, ofs=&f3
-	mov	r19=b1
+	stf.spill [r8]=f2,32
+	stf.spill [r9]=f3,32
 	;;
-}
-{	.mmi
-	stf.spill [r16]=f2,32		// save f2, ofs=&f4
-	stf.spill [r17]=f3,32		// save f3, ofs=&f5
-	mov	r20=b2
+	stf.spill [r8]=f4,32
+	stf.spill [r9]=f5,32
 	;;
-}
-{	.mmi
-	stf.spill [r16]=f4,32		// save f4, ofs=&b0
-	stf.spill [r17]=f5,24		// save f5, ofs=&b1
-	mov	r21=b3
+	stf.spill [r8]=f16,32
+	stf.spill [r9]=f17,32
 	;;
-}
-{	.mmi
-	st8	[r16]=r18,16		// save b0, ofs=&b2
-	st8	[r17]=r19,16		// save b1, ofs=&b3
-	mov	r18=b4
+	stf.spill [r8]=f18,32
+	stf.spill [r9]=f19,32
+	;;
+	stf.spill [r8]=f20,32
+	stf.spill [r9]=f21,32
+	;;
+	stf.spill [r8]=f22,32
+	stf.spill [r9]=f23,32
+	;;
+	stf.spill [r8]=f24,32
+	stf.spill [r9]=f25,32
+	;;
+	stf.spill [r8]=f26,32
+	stf.spill [r9]=f27,32
+	;;
+	stf.spill [r8]=f28,32
+	stf.spill [r9]=f29,32
+	;;
+	stf.spill [r8]=f30,32
+	stf.spill [r9]=f31,24		// BEWARE
 	;;
-}
 {	.mmi
-	st8	[r16]=r20,16		// save b2, ofs=&b4
-	st8	[r17]=r21,16		// save b3, ofs=&b5
-	mov	r19=b5
+	.mem.offset 0,0
+	st8.spill [r8]=r4,16
+	.mem.offset 8,0
+	st8.spill [r9]=r5,16
+	mov	r17=b1
 	;;
 }
 {	.mmi
-	st8	[r16]=r18,16		// save b4, ofs=&old_unat
-	st8	[r17]=r19,16		// save b5, ofs=&sp
-	mov	r18=ar.pfs
+	.mem.offset 16,0
+	st8.spill [r8]=r6,16
+	.mem.offset 24,0
+	st8.spill [r9]=r7,16
+	mov	r18=b2
 	;;
 }
 {	.mmi
-	st8	[r16]=r22,16		// save old_unat, ofs=&pfs
-	st8	[r17]=sp,16		// save sp, ofs=&bspstore
-	mov	r19=ar.lc
+	flushrs
+	mov	r16=ar.unat
+	mov	r19=b3
 	;;
 }
 {	.mmi
-	mov	r20=ar.bspstore
-	mov	r21=ar.unat
-	add	r23=PC_CURRENT_PMAP,r13
+	st8	[r8]=r16,16		// ar.unat
+	st8	[r9]=r17,16		// b1
+	mov	r20=b4
 	;;
 }
 {	.mmi
-	st8	[r16]=r18,16		// save pfs, ofs=&lc
-	st8	[r17]=r20,16		// save bspstore, ofs=&unat
-	mov	r18=pr
+	st8	[r8]=r18,16		// b2
+	st8	[r9]=r19,16		// b3
+	mov	r21=b5
 	;;
 }
+	st8	[r8]=r20,16		// b4
+	st8	[r9]=r21,16		// b5
+	;;
 {	.mmi
-	st8	[r16]=r19,16		// save lc, ofs=&rnat
-	st8	[r17]=r21,16		// save unat, ofs=&pr
-	mov	b6=in1
+	mov	ar.rsc=0
+	mov	r16=ar.bsp
+	mov	r17=ar.pfs
 	;;
 }
 {	.mmi
-	st8	[r16]=r22,16		// save rnat, ofs=&pmap
-	st8	[r17]=r18,16		// save pr, ofs=&ar.fsr
-	nop	1
+	st8	[r8]=r16,16		// ar.bsp
+	st8	[r9]=r17,16		// ar.pfs
+	mov	r18=ar.lc
 	;;
 }
-{	.mmb
-	ld8	r19=[r23]
-	mov	r18=ar.fsr
-	nop	2
+	mov	r17=ar.rnat
+	ld8	r19=[r31]
 	;;
-}
-{	.mmi
-	st8	[r16]=r19,16		// save pmap, ofs=&ar.fcr
-	st8	[r17]=r18,16		// save ar.fsr, ofs=&ar.fir
-	nop	3
+	st8	[r8]=r17,16		// ar.rnat
+	st8	[r9]=r18,16		// ar.lc
 	;;
-}
-{	.mmb
-	mov	r19=ar.fcr
-	mov	r18=ar.fir
-	nop	4
+	st8	[r8]=r19,16		// pc_current_pmap
+	mov	r16=ar.fcr
 	;;
-}
-{	.mmi
-	st8	[r16]=r19,16		// save ar.fcr, ofs=&ar.fdr
-	st8	[r17]=r18,16		// save ar.fir, ofs=&ar.eflag
-	nop	5
+	st8	[r9]=r16,16		// ar.fcr
+	mov	r17=ar.eflag
+	;;
+	st8	[r8]=r17,16		// ar.eflag
+	mov	r16=ar.csd
+	;;
+	st8	[r9]=r16,16		// ar.csd
+	mov	r17=ar.ssd
+	;;
+	st8	[r8]=r17,16		// ar.ssd
+	mov	r16=ar.fsr
+	;;
+	st8	[r9]=r16,16		// ar.fsr
+	mov	r17=ar.fir
+	;;
+	st8	[r8]=r17,16		// ar.fir
+	mov	r16=ar.fdr
 	;;
-}
 {	.mmb
-	mov	r19=ar.fdr
-	mov	r18=ar.eflag
-	nop	6
+	st8	[r9]=r16,16		// ar.fdr
+	mov	ar.rsc=3
+	br.sptk.many b6
 	;;
 }
+END(pcb_save)
+
+/*
+ * pcb_restore - restore a process context. This does not return.
+ *
+ * Arguments:
+ *	r8	'struct pcb *' of the process that needs its context saved.
+ */
+ENTRY(pcb_restore,0)
+	mov	ar.rsc=0
+	add	r9=8,r8
+	add	r31=PCB_UNAT47,r8
+	;;
+	ld8	r12=[r8],16		// sp
+	ld8	r30=[r9],16		// ar.unat
+	;;
+	ld8	r16=[r8],16		// rp
+	ld8	r17=[r9],24		// pr
+	;;
+	ldf.fill f2=[r8],32
+	ldf.fill f3=[r9],32
+	mov	b0=r16
+	;;
+	ldf.fill f4=[r8],32
+	ldf.fill f5=[r9],32
+	mov	pr=r17,0x1fffe
+	;;
+	ldf.fill f16=[r8],32
+	ldf.fill f17=[r9],32
+	;;
+	ldf.fill f18=[r8],32
+	ldf.fill f19=[r9],32
+	;;
+	ldf.fill f20=[r8],32
+	ldf.fill f21=[r9],32
+	;;
+	ldf.fill f22=[r8],32
+	ldf.fill f23=[r9],32
+	;;
+	ldf.fill f24=[r8],32
+	ldf.fill f25=[r9],32
+	;;
+	ldf.fill f26=[r8],32
+	ldf.fill f27=[r9],32
+	;;
+	ldf.fill f28=[r8],32
+	ldf.fill f29=[r9],32
+	;;
+	ldf.fill f30=[r8],32
+	ldf.fill f31=[r9],24		// BEWARE
+	;;
+	ld8	r16=[r31]
+	;;
+	mov	ar.unat=r16
+	;;
+	.mem.offset 0,0
+	ld8.fill r4=[r8],16
+	.mem.offset 8,0
+	ld8.fill r5=[r9],16
+	;;
+	.mem.offset 16,0
+	ld8.fill r6=[r8],24
+	.mem.offset 24,0
+	ld8.fill r7=[r9],24
+	;;
+	loadrs
+	mov	ar.unat=r30
+	;;
+	ld8	r16=[r8],16		// b1
+	ld8	r17=[r9],16		// b2
+	;;
+	ld8	r18=[r8],16		// b3
+	ld8	r19=[r9],16		// b4
+	mov	b1=r16
+	;;
+	ld8	r20=[r8],16		// b5
+	ld8	r21=[r9],16		// ar.bsp
+	mov	b2=r17
+	;;
+	ld8	r22=[r8],16		// ar.pfs
+	ld8	r23=[r9],24		// ar.rnat
+	mov	b3=r18
+	;;
+	ld8	r24=[r8],24		// ar.lc
+	mov	b4=r19
+	mov	b5=r20
+	;;
 {	.mmi
-	st8	[r16]=r19,16		// save ar.fdr, ofs=&ar.csd
-	st8	[r17]=r18,16		// save ar.eflag, ofs=&ssd
-	nop	7
+	mov	ar.bspstore=r21
 	;;
-}
-{	.mmb
-	mov	r19=ar.csd
-	mov	r18=ar.ssd
-	nop	8
+	mov	ar.rnat=r23
+	mov	ar.pfs=r22
 	;;
 }
 {	.mmi
-	st8	[r16]=r19		// save ar.csd
-	st8	[r17]=r18		// save ar.ssd
-	nop	9
+	invala
+	;;
+	ld8	r16=[r9],16		// ar.fcr
+	mov	r18=ar.lc
+	;;
 }
-{	.mib
-	mov	ar.rsc=3		// turn RSE back on
-	mov	ret0=r0
-	br.sptk.few	b6
+	ld8	r17=[r8],16		// ar.eflag
+	mov	ar.fcr=r16
+	;;
+	ld8	r16=[r9],16		// ar.csd
+	mov	ar.eflag=r17
+	;;
+	ld8	r17=[r8],16		// ar.ssd
+	mov	ar.csd=r16
+	;;
+	ld8	r16=[r9],16		// ar.fsr
+	mov	ar.ssd=r17
+	;;
+	ld8	r17=[r8],16		// ar.fir
+	mov	ar.fsr=r16
+	;;
+	ld8	r16=[r9],16		// ar.fdr
+	mov	ar.fir=r17
+	;;
+{	.mmb
+	mov	ar.fdr=r16
+	mov	ar.rsc=3
+	br.ret.sptk b0
 	;;
 }
-END(pcb_save)
+END(pcb_restore)
 
+/*
+ * savectx - Save current context.
+ */
 ENTRY(savectx,1)
 {	.mii
-1:	alloc	r16=ar.pfs,0,0,2,0
+1:	nop	1
 	mov	r17=ip
 	;;
-	add	r33=2f-1b,r17
+	add	r16=2f-1b,r17
+	;;
 }
-{	.mfb
-	nop	1
-	nop	2
-	br.sptk.few	pcb_save
+{	.mib
+	mov	r8=r32
+	mov	b6=r16
+	br.sptk.many pcb_save
 }
 {	.mfb
-2:	nop	3
-	nop	4
-	br.ret.sptk	rp
+2:	nop	2
+	nop	3
+	br.ret.sptk rp
 }
 END(savectx)
 
@@ -213,18 +342,15 @@ END(savectx)
  *	Does not return. We arrange things so that savectx appears to
  *	return a second time with a non-zero return value.
  */
-
 ENTRY(restorectx, 1)
-
-	mov	r15=in0
-	br.cond.sptk.many 4f
-	
+	mov	r8=in0
+	br.cond.sptk.many pcb_restore
+	;;
 END(restorectx)	
 
 ENTRY(cpu_throw, 0)
-	
-	br.sptk.few 2f
-
+	br.sptk.many 2f
+	;;
 END(cpu_throw)
 
 ENTRY(cpu_switch, 0)
@@ -233,16 +359,17 @@ ENTRY(cpu_switch, 0)
 	mov	r17=ip
 	;;
 }
-	add	r33=1f-1b,r17
+	add	r16=1f-1b,r17
 	add	r17=PC_CURTHREAD,r13
 	;;
 	ld8	r17=[r17]
 	;;
 	add	r17=TD_PCB,r17
+	mov	b6=r16
+	;;
+	ld8	r8=[r17]
+	br.sptk.many pcb_save
 	;;
-	ld8	r32=[r17]
-	br.sptk.few	pcb_save
-
 1:
 #ifdef SMP
 	add	r17 = PC_CPUID, r13
@@ -263,10 +390,8 @@ ENTRY(cpu_switch, 0)
 	br.sptk	3f
 #endif
 
-2:	srlz.i
-	mf
-	;; 
-	br.call.sptk.few rp=choosethread
+2:	br.call.sptk.few rp=choosethread
+	;;
 	mov	r4=ret0			// save from call
 
 3:
@@ -276,6 +401,7 @@ ENTRY(cpu_switch, 0)
 	ld8	out0=[r14]
 	mov	out1=1
 	br.call.sptk.few rp=ia64_fpstate_save // clear fpcurthread
+	;;
 #endif
 
 	add	r14=PC_CURTHREAD,r13
@@ -287,96 +413,20 @@ ENTRY(cpu_switch, 0)
 	;;
 	ld8	r15=[r15]
 	;;
-	add	r15=PCB_PMAP,r15	// &pcb_pmap
+	add	r15=PCB_CURRENT_PMAP,r15 // &pcb_pmap
 	;;
 	ld8	out0=[r15]
 	br.call.sptk.few rp=pmap_install // install RIDs etc.
-
+	;;
 	add	r15=TD_PCB,r4
 	add	r16=TD_KSTACK,r4
 	;;
-	ld8	r15=[r15]
+	ld8	r8=[r15]
 	ld8	r16=[r16]
 	;;
 	mov	ar.k5=r16
-
-	// pcb_restore
-4:	
-	add	r3=PCB_UNAT,r15		// point at NaT for r4..r7
-	mov	ar.rsc=0 ;;		// switch off the RSE
-	ld8	r16=[r3]		// load NaT for r4..r7
-	;;
-	mov	ar.unat=r16
-	;;
-	ld8.fill r4=[r15],8 ;;		// restore r4
-	ld8.fill r5=[r15],8 ;;		// restore r5
-	ld8.fill r6=[r15],8 ;;		// restore r6
-	ld8.fill r7=[r15],8 ;;		// restore r7
-
-	ldf.fill f2=[r15],16 ;;		// restore f2
-	ldf.fill f3=[r15],16 ;;		// restore f3
-	ldf.fill f4=[r15],16 ;;		// restore f4
-	ldf.fill f5=[r15],16 ;;		// restore f5
-
-	ld8	r16=[r15],8 ;;		// restore b0
-	ld8	r17=[r15],8 ;;		// restore b1
-	ld8	r18=[r15],8 ;;		// restore b2
-	ld8	r19=[r15],8 ;;		// restore b3
-	ld8	r20=[r15],8 ;;		// restore b4
-	ld8	r21=[r15],8 ;;		// restore b5
-
-	mov	b0=r16
-	mov	b1=r17
-	mov	b2=r18
-	mov	b3=r19
-	mov	b4=r20
-	mov	b5=r21
-
-	ld8	r16=[r15],8 ;;		// caller's ar.unat
-	ld8	sp=[r15],8 ;;		// stack pointer
-	ld8	r17=[r15],8 ;;		// ar.pfs
-	ld8	r18=[r15],8 ;;		// ar.bspstore
-	ld8	r21=[r15],16 ;;		// ar.lc, skip ar.unat
-	ld8	r19=[r15],8 ;;		// ar.rnat
-	ld8	r20=[r15],16 ;;		// pr, skip pmap
-
-	loadrs				// invalidate register stack
-	;;
-	mov	ar.unat=r16
-	mov	ar.pfs=r17
-	mov	ar.lc=r21
-	mov	ar.bspstore=r18 ;;
-	mov	ar.rnat=r19
-	mov	pr=r20,0x1ffff
-	;;
-	ld8	r16=[r15],8		// ar.fsr
-	;;
-	ld8	r17=[r15],8		// ar.fcr
-	mov	ar.fsr=r16
-	;; 
-	ld8	r16=[r15],8		// ar.fir
-	mov	ar.fcr=r17
-	;; 
-	ld8	r17=[r15],8		// ar.fdr
-	mov	ar.fir=r16
-	;; 
-	ld8	r16=[r15],8		// ar.eflag
-	mov	ar.fdr=r17
-	;; 
-	ld8	r17=[r15],8		// ar.csd
-	mov	ar.eflag=r16
-	;; 
-	ld8	r16=[r15],8		// ar.ssd
-	mov	ar.csd=r17
-	;; 
-	mov	ar.ssd=r16
-	
-	mov	ar.rsc=3		// restart RSE
-	invala
+	br.cond.sptk pcb_restore
 	;;
-9:
-	br.ret.sptk.few rp
-
 END(cpu_switch)
 
 /*
@@ -397,13 +447,13 @@ ENTRY(savehighfp, 1)
 	;; 
 	stf.spill [in0]=f36,32
 	stf.spill [r14]=f37,32
-	;; 
+	;;
 	stf.spill [in0]=f38,32
 	stf.spill [r14]=f39,32
-	;; 
+	;;
 	stf.spill [in0]=f40,32
 	stf.spill [r14]=f41,32
-	;; 
+	;;
 	stf.spill [in0]=f42,32
 	stf.spill [r14]=f43,32
 	;; 
@@ -721,5 +771,5 @@ ENTRY(fork_trampoline, 0)
 	br.call.sptk.few rp=fork_exit
 	;; 
 	br.cond.sptk.many exception_restore
-
-	END(fork_trampoline)
+	;;
+END(fork_trampoline)
diff --git a/sys/ia64/ia64/vm_machdep.c b/sys/ia64/ia64/vm_machdep.c
index f0cdab0..6a63482 100644
--- a/sys/ia64/ia64/vm_machdep.c
+++ b/sys/ia64/ia64/vm_machdep.c
@@ -215,7 +215,7 @@ cpu_fork(td1, p2, td2, flags)
 	if (p2tf->tf_cr_ipsr & IA64_PSR_IS) {
 		p2tf->tf_r[FRAME_R8] = 0; /* child returns zero (eax) */
 		p2tf->tf_r[FRAME_R10] = 1; /* is child (edx) */
-		td2->td_pcb->pcb_eflag &= ~PSL_C; /* no error */
+		td2->td_pcb->pcb_ar_eflag &= ~PSL_C; /* no error */
 	} else {
 		p2tf->tf_r[FRAME_R8] = 0; /* child's pid (linux) 	*/
 		p2tf->tf_r[FRAME_R9] = 1; /* is child (FreeBSD) 	*/
@@ -282,10 +282,10 @@ cpu_fork(td1, p2, td2, flags)
 	 * straight into exception_restore. Also initialise its
 	 * pmap to the containing proc's vmspace.
 	 */
-	td2->td_pcb->pcb_bspstore = (u_int64_t)p2bs + td1->td_frame->tf_ndirty;
-	td2->td_pcb->pcb_rnat = rnat;
-	td2->td_pcb->pcb_pfs = 0;
-	td2->td_pcb->pcb_pmap = (u_int64_t)
+	td2->td_pcb->pcb_ar_bsp = (u_int64_t)p2bs + td1->td_frame->tf_ndirty;
+	td2->td_pcb->pcb_ar_rnat = rnat;
+	td2->td_pcb->pcb_ar_pfs = 0;
+	td2->td_pcb->pcb_current_pmap = (u_int64_t)
 		vmspace_pmap(td2->td_proc->p_vmspace);
 
 	/*
@@ -298,10 +298,10 @@ cpu_fork(td1, p2, td2, flags)
 	 * available as scratch space.
 	 */
 	td2->td_pcb->pcb_sp = (u_int64_t)p2tf - 16;	
-	td2->td_pcb->pcb_r4 = (u_int64_t)fork_return;
-	td2->td_pcb->pcb_r5 = FDESC_FUNC(exception_restore);
-	td2->td_pcb->pcb_r6 = (u_int64_t)td2;
-	td2->td_pcb->pcb_b0 = FDESC_FUNC(fork_trampoline);
+	td2->td_pcb->pcb_r[PCB_R4] = (u_int64_t)fork_return;
+	td2->td_pcb->pcb_r[PCB_R5] = FDESC_FUNC(exception_restore);
+	td2->td_pcb->pcb_r[PCB_R6] = (u_int64_t)td2;
+	td2->td_pcb->pcb_rp = FDESC_FUNC(fork_trampoline);
 }
 
 /*
@@ -316,8 +316,8 @@ cpu_set_fork_handler(td, func, arg)
 	void (*func)(void *);
 	void *arg;
 {
-	td->td_pcb->pcb_r4 = (u_int64_t) func;
-	td->td_pcb->pcb_r6 = (u_int64_t) arg;
+	td->td_pcb->pcb_r[PCB_R4] = (u_int64_t) func;
+	td->td_pcb->pcb_r[PCB_R6] = (u_int64_t) arg;
 }
 
 /*
diff --git a/sys/ia64/include/pcb.h b/sys/ia64/include/pcb.h
index beba354..c5ee145 100644
--- a/sys/ia64/include/pcb.h
+++ b/sys/ia64/include/pcb.h
@@ -33,46 +33,63 @@
  * PCB: process control block
  */
 struct pcb {
-	u_int64_t		pcb_r4;
-	u_int64_t		pcb_r5;
-	u_int64_t		pcb_r6;
-	u_int64_t		pcb_r7;
+	uint64_t		pcb_sp;
+	uint64_t		pcb_ar_unat;
+	uint64_t		pcb_rp;
+	uint64_t		pcb_pr;
+	struct ia64_fpreg	pcb_f[20];
+#define	PCB_F2		0
+#define	PCB_F3		1
+#define	PCB_F4		2
+#define	PCB_F5		3
+#define	PCB_F16		4
+#define	PCB_F17		5
+#define	PCB_F18		6
+#define	PCB_F19		7
+#define	PCB_F20		8
+#define	PCB_F21		9
+#define	PCB_F22		10
+#define	PCB_F23		11
+#define	PCB_F24		12
+#define	PCB_F25		13
+#define	PCB_F26		14
+#define	PCB_F27		15
+#define	PCB_F28		16
+#define	PCB_F29		17
+#define	PCB_F30		18
+#define	PCB_F31		19
+	uint64_t		pcb_r[4];
+#define	PCB_R4		0
+#define	PCB_R5		1
+#define	PCB_R6		2
+#define	PCB_R7		3
+	uint64_t		pcb_unat47;
+	uint64_t		pcb_b[5];
+#define	PCB_B1		0
+#define	PCB_B2		1
+#define	PCB_B3		2
+#define	PCB_B4		3
+#define	PCB_B5		4
+	uint64_t		pcb_ar_bsp;
+	uint64_t		pcb_ar_pfs;
+	uint64_t		pcb_ar_rnat;
+	uint64_t		pcb_ar_lc;
 
-	struct ia64_fpreg	pcb_f2;
-	struct ia64_fpreg	pcb_f3;
-	struct ia64_fpreg	pcb_f4;
-	struct ia64_fpreg	pcb_f5;
+	uint64_t		pcb_current_pmap;
 
-	u_int64_t		pcb_b0;		/* really restart address */
-	u_int64_t		pcb_b1;
-	u_int64_t		pcb_b2;
-	u_int64_t		pcb_b3;
-	u_int64_t		pcb_b4;
-	u_int64_t		pcb_b5;
-
-	u_int64_t		pcb_old_unat;	/* caller's ar.unat */
-	u_int64_t		pcb_sp;
-	u_int64_t		pcb_pfs;
-	u_int64_t		pcb_bspstore;
-	u_int64_t		pcb_lc;
-
-	u_int64_t		pcb_unat;	/* ar.unat for r4..r7 */
-	u_int64_t		pcb_rnat;
-	u_int64_t		pcb_pr;		/* predicates */
-	u_int64_t		pcb_pmap;	/* current pmap */
-
-	u_int64_t		pcb_fsr;
-	u_int64_t		pcb_fcr;
-	u_int64_t		pcb_fir;
-	u_int64_t		pcb_fdr;
-	u_int64_t		pcb_eflag;
-	u_int64_t		pcb_csd;
-	u_int64_t		pcb_ssd;
-
-	u_int64_t		pcb_onfault;	/* for copy faults */
-	u_int64_t		pcb_accessaddr;	/* for [fs]uswintr */
+	uint64_t		pcb_ar_fcr;
+	uint64_t		pcb_ar_eflag;
+	uint64_t		pcb_ar_csd;
+	uint64_t		pcb_ar_ssd;
+	uint64_t		pcb_ar_fsr;
+	uint64_t		pcb_ar_fir;
+	uint64_t		pcb_ar_fdr;
 
+	/* Aligned! */
 	struct ia64_fpreg	pcb_highfp[96];	/* f32-f127 */
+
+	uint64_t		pcb_onfault;	/* for copy faults */
+	uint64_t		pcb_accessaddr;	/* for [fs]uswintr */
 };
 
 #ifdef _KERNEL