- Generally improve register usage in cpu_switch. Use the 'in' registers

  for temporaries relating to the state of the new process instead of the
  outs, so that functions can be called without fear of clobbering them.
- Use savefpctx instead of rolling our own.

1 files changed, 63 insertions, 71 deletions

diff --git a/sys/sparc64/sparc64/swtch.S b/sys/sparc64/sparc64/swtch.S
index 21f5a9b..7cea5ee 100644
--- a/sys/sparc64/sparc64/swtch.S
+++ b/sys/sparc64/sparc64/swtch.S
@@ -42,9 +42,8 @@
 ENTRY(cpu_throw)
 	save	%sp, -CCFSZ, %sp
 	flushw
-	mov	%i0, %l0
 	ba	%xcc, .Lsw1
-	 mov	%i1, %o0
+	 mov	%i1, %i0
 END(cpu_throw)
 
 /*
@@ -52,40 +51,34 @@ END(cpu_throw)
  */
 ENTRY(cpu_switch)
 	save	%sp, -CCFSZ, %sp
-	mov	%i0, %l0
-	mov	%i1, %o0
+	mov	%i1, %i0
 
 	/*
-	 * If the current thread was using floating point, save its context.
+	 * If the current thread was using floating point in userland, save
+	 * its context.
 	 */
-	ldx	[%l0 + TD_PCB], %l1
-	ldx	[%l0 + TD_FRAME], %l2
+1:	sub	PCB_REG, TF_SIZEOF, %l2
 	ldx	[%l2 + TF_FPRS], %l3
 	andcc	%l3, FPRS_FEF, %g0
-	bz,a,pt	%xcc, 1f
+	bz,a,pt	%xcc, 2f
 	 nop
-	wr	%g0, FPRS_FEF, %fprs
-	wr	%g0, ASI_BLK_S, %asi
-	stda	%f0, [%l1 + PCB_UFP + (0 * 64)] %asi
-	stda	%f16, [%l1 + PCB_UFP + (1 * 64)] %asi
-	stda	%f32, [%l1 + PCB_UFP + (2 * 64)] %asi
-	stda	%f48, [%l1 + PCB_UFP + (3 * 64)] %asi
-	membar	#Sync
-	wr	%g0, 0, %fprs
+	call	savefpctx
+	 add	PCB_REG, PCB_UFP, %o0
 	andn	%l3, FPRS_FEF, %l3
 	stx	%l3, [%l2 + TF_FPRS]
-	ldx	[%l1 + PCB_FLAGS], %l3
+
+	ldx	[PCB_REG + PCB_FLAGS], %l3
 	or	%l3, PCB_FEF, %l3
-	stx	%l3, [%l1 + PCB_FLAGS]
+	stx	%l3, [PCB_REG + PCB_FLAGS]
 
 	/*
 	 * Flush the windows out to the stack and save the current frame
 	 * pointer and program counter.
 	 */
-1:	flushw
+2:	flushw
 	wrpr	%g0, 0, %cleanwin
-	stx	%fp, [%l1 + PCB_SP]
-	stx	%i7, [%l1 + PCB_PC]
+	stx	%fp, [PCB_REG + PCB_SP]
+	stx	%i7, [PCB_REG + PCB_PC]
 
 	/*
 	 * Load the new thread's frame pointer and program counter, and set
@@ -95,42 +88,44 @@ ENTRY(cpu_switch)
 #if KTR_COMPILE & KTR_PROC
 	CATR(KTR_PROC, "cpu_switch: new td=%p pc=%#lx fp=%#lx"
 	    , %g1, %g2, %g3, 7, 8, 9)
-	stx	%o0, [%g1 + KTR_PARM1]
-	ldx	[%o0 + TD_PCB], %g2
+	stx	%i0, [%g1 + KTR_PARM1]
+	ldx	[%i0 + TD_PCB], %g2
 	ldx	[%g2 + PCB_PC], %g3
 	stx	%g3, [%g1 + KTR_PARM2]
 	ldx	[%g2 + PCB_SP], %g3
 	stx	%g3, [%g1 + KTR_PARM3]
 9:
 #endif
-	ldx	[%o0 + TD_PCB], %o1
-	ldx	[%o1 + PCB_SP], %fp
-	ldx	[%o1 + PCB_PC], %i7
-	sub	%fp, CCFSZ, %sp
-	stx	%o0, [PCPU(CURTHREAD)]
-	stx	%o1, [PCPU(CURPCB)]
+	ldx	[%i0 + TD_PCB], %i1
 
-	SET(sched_lock, %o3, %o2)
-	stx	%o0, [%o2 + MTX_LOCK]
+	stx	%i0, [PCPU(CURTHREAD)]
+	stx	%i1, [PCPU(CURPCB)]
 
 	wrpr	%g0, PSTATE_NORMAL, %pstate
-	mov	%o1, PCB_REG
+	mov	%i1, PCB_REG
 	wrpr	%g0, PSTATE_ALT, %pstate
-	mov	%o1, PCB_REG
+	mov	%i1, PCB_REG
 	wrpr	%g0, PSTATE_KERNEL, %pstate
 
+	ldx	[PCB_REG + PCB_SP], %fp
+	ldx	[PCB_REG + PCB_PC], %i7
+	sub	%fp, CCFSZ, %sp
+
+	SET(sched_lock, %i3, %i2)
+	stx	%i0, [%i2 + MTX_LOCK]
+
 	/*
 	 * Point to the vmspaces of the new process, and of the last non-kernel
 	 * process to run.
 	 */
-	ldx	[%o0 + TD_PROC], %o2
+	ldx	[%i0 + TD_PROC], %i2
 	ldx	[PCPU(VMSPACE)], %l2
-	ldx	[%o2 + P_VMSPACE], %o2
+	ldx	[%i2 + P_VMSPACE], %i2
 
 #if KTR_COMPILE & KTR_PROC
 	CATR(KTR_PROC, "cpu_switch: new vm=%p old vm=%p"
 	    , %g1, %g2, %g3, 7, 8, 9)
-	stx	%o2, [%g1 + KTR_PARM1]
+	stx	%i2, [%g1 + KTR_PARM1]
 	stx	%l2, [%g1 + KTR_PARM2]
 9:
 #endif
@@ -138,32 +133,32 @@ ENTRY(cpu_switch)
 	/*
 	 * If they are the same we are done.
 	 */
-	cmp	%l2, %o2
-	be,a,pn %xcc, 4f
+	cmp	%l2, %i2
+	be,a,pn %xcc, 5f
 	 nop
 
 	/*
 	 * If the new process has nucleus context we are done.
 	 */
-	lduw	[PCPU(CPUID)], %o3
-	sllx	%o3, INT_SHIFT, %o3
-	add	%o2, VM_PMAP + PM_CONTEXT, %o4
-	lduw	[%o3 + %o4], %o5
+	lduw	[PCPU(CPUID)], %i3
+	sllx	%i3, INT_SHIFT, %i3
+	add	%i2, VM_PMAP + PM_CONTEXT, %i4
+	lduw	[%i3 + %i4], %i5
 
 #if KTR_COMPILE & KTR_PROC
 	CATR(KTR_PROC, "cpu_switch: ctx=%#lx"
 	    , %g1, %g2, %g3, 7, 8, 9)
-	stx	%o5, [%g1 + KTR_PARM1]
+	stx	%i5, [%g1 + KTR_PARM1]
 9:
 #endif
 
-	brz,a,pn %o5, 4f
+	brz,a,pn %i5, 5f
 	 nop
 
 	/*
 	 * If there was no non-kernel vmspace, don't try to deactivate it.
 	 */
-	brz,a,pn %l2, 2f
+	brz,a,pn %l2, 3f
 	 nop
 
 	/*
@@ -188,31 +183,31 @@ ENTRY(cpu_switch)
 	 * Find the current free tlb context for this cpu and install it as
 	 * the new primary context.
 	 */
-2:	lduw	[PCPU(TLB_CTX)], %o5
-	stw	%o5, [%o3 + %o4]
-	mov	AA_DMMU_PCXR, %o4
-	stxa	%o5, [%o4] ASI_DMMU
+3:	lduw	[PCPU(TLB_CTX)], %i5
+	stw	%i5, [%i3 + %i4]
+	mov	AA_DMMU_PCXR, %i4
+	stxa	%i5, [%i4] ASI_DMMU
 	membar	#Sync
 
 	/*
 	 * See if we have run out of free contexts.
 	 */
-	lduw	[PCPU(TLB_CTX_MAX)], %o3
+	lduw	[PCPU(TLB_CTX_MAX)], %i3
 
 #if KTR_COMPILE & KTR_PROC
 	CATR(KTR_PROC, "cpu_switch: ctx=%#lx next=%#lx max=%#lx"
 	    , %g1, %g2, %g3, 7, 8, 9)
-	stx	%o5, [%g1 + KTR_PARM1]
-	add	%o5, 1, %g2
+	stx	%i5, [%g1 + KTR_PARM1]
+	add	%i5, 1, %g2
 	stx	%g2, [%g1 + KTR_PARM2]
-	stx	%o3, [%g1 + KTR_PARM3]
+	stx	%i3, [%g1 + KTR_PARM3]
 9:
 #endif
 
-	add	%o5, 1, %o5
-	cmp	%o3, %o5
-	bne,a,pt %xcc, 3f
-	 stw	%o5, [PCPU(TLB_CTX)]
+	add	%i5, 1, %i5
+	cmp	%i3, %i5
+	bne,a,pt %xcc, 4f
+	 stw	%i5, [PCPU(TLB_CTX)]
 
 #if KTR_COMPILE & KTR_PROC
 	CATR(KTR_PROC, "cpu_switch: context rollover"
@@ -226,34 +221,31 @@ ENTRY(cpu_switch)
 	 */
 	call	pmap_context_rollover
 	 nop
-	ldx	[PCPU(CURTHREAD)], %o0
-	ldx	[%o0 + TD_PROC], %o2
-	ldx	[%o2 + P_VMSPACE], %o2
 
 	/*
 	 * Mark the pmap as active on this cpu.
 	 */
-3:	lduw	[%o2 + VM_PMAP + PM_ACTIVE], %o3
-	lduw	[PCPU(CPUMASK)], %o4
-	or	%o3, %o4, %o3
-	stw	%o3, [%o2 + VM_PMAP + PM_ACTIVE]
+4:	lduw	[%i2 + VM_PMAP + PM_ACTIVE], %i3
+	lduw	[PCPU(CPUMASK)], %i4
+	or	%i3, %i4, %i3
+	stw	%i3, [%i2 + VM_PMAP + PM_ACTIVE]
 
 	/*
 	 * Make note of the change in vmspace.
 	 */
-	stx	%o2, [PCPU(VMSPACE)]
+	stx	%i2, [PCPU(VMSPACE)]
 
 	/*
 	 * Load the tsb registers.
 	 */
-	ldx	[%o2 + VM_PMAP + PM_TSB], %o3
-	mov	AA_DMMU_TSB, %o4
-	stxa	%o3, [%o4] ASI_DMMU
-	mov	AA_IMMU_TSB, %o4
-	stxa	%o3, [%o4] ASI_IMMU
+	ldx	[%i2 + VM_PMAP + PM_TSB], %i3
+	mov	AA_DMMU_TSB, %i4
+	stxa	%i3, [%i4] ASI_DMMU
+	mov	AA_IMMU_TSB, %i4
+	stxa	%i3, [%i4] ASI_IMMU
 	membar	#Sync
 
-4:
+5:
 #if KTR_COMPILE & KTR_PROC
 	CATR(KTR_PROC, "cpu_switch: return"
 	    , %g1, %g2, %g3, 7, 8, 9)