From 3376a80f5e5fe03bdfa558a1e46993abd8f32ba5 Mon Sep 17 00:00:00 2001
From: mickey <mickey@openbsd.org>
Date: Mon, 6 Mar 2000 02:21:01 +0000
Subject: [PATCH] do not save/restore callee-save regs on trap, except for DDB

---
 sys/arch/hppa/hppa/locore.S | 230 ++++++++++++++++++------------------
 1 file changed, 115 insertions(+), 115 deletions(-)

diff --git a/sys/arch/hppa/hppa/locore.S b/sys/arch/hppa/hppa/locore.S
index b2332d4edd3..44018927875 100644
--- a/sys/arch/hppa/hppa/locore.S
+++ b/sys/arch/hppa/hppa/locore.S
@@ -1,4 +1,4 @@
-/*	$OpenBSD: locore.S,v 1.28 2000/02/10 15:16:59 mickey Exp $	*/
+/*	$OpenBSD: locore.S,v 1.29 2000/03/06 02:21:01 mickey Exp $	*/
 
 /*
  * Copyright (c) 1998-2000 Michael Shalayeff
@@ -319,7 +319,6 @@ ENTRY(pdc_call)
 
 	mfctl	eiem, t1
 	mtctl	r0, eiem
-	stw	t1, HPPA_FRAME_ARG(0)(sp)
 	stw	rp, HPPA_FRAME_CRP(sp)
 	copy	arg0, r31
 	copy	sp, ret1
@@ -331,64 +330,66 @@ ENTRY(pdc_call)
 	ldil	L%pdc_stack, ret1
 
 pdc_call_unmapped1
-	ldo	HPPA_FRAME_SIZE+24*4(ret1), ret1
+	copy	sp, r1
+	ldo	HPPA_FRAME_SIZE+24*4(ret1), sp
 
-	stw	sp, HPPA_FRAME_PSP(ret1)
+	stw	r1, HPPA_FRAME_PSP(sp)
+
+	/* save kernelmapped and eiem */
+	stw	ret0, HPPA_FRAME_ARG(21)(sp)
+	stw	t1, HPPA_FRAME_ARG(22)(sp)
+
+	/* copy arguments */
 	copy	arg2, arg0
 	copy	arg3, arg1
-	ldw	HPPA_FRAME_ARG(4)(sp), arg2
-	ldw	HPPA_FRAME_ARG(5)(sp), arg3
-	ldw	HPPA_FRAME_ARG(6)(sp), t1
-	ldw	HPPA_FRAME_ARG(7)(sp), t2
-	ldw	HPPA_FRAME_ARG(8)(sp), t3
-	ldw	HPPA_FRAME_ARG(9)(sp), t4
-	stw	t1, HPPA_FRAME_ARG(4)(ret1)
-	stw	t2, HPPA_FRAME_ARG(5)(ret1)
-	stw	t3, HPPA_FRAME_ARG(6)(ret1)
-	stw	t4, HPPA_FRAME_ARG(7)(ret1)
-	ldw	HPPA_FRAME_ARG(10)(sp), t1
-	ldw	HPPA_FRAME_ARG(11)(sp), t2
-	ldw	HPPA_FRAME_ARG(12)(sp), t3
-	ldw	HPPA_FRAME_ARG(13)(sp), t4
-	stw	t1, HPPA_FRAME_ARG(8)(ret1)
-	stw	t2, HPPA_FRAME_ARG(9)(ret1)
-	stw	t3, HPPA_FRAME_ARG(10)(ret1)
-	stw	t4, HPPA_FRAME_ARG(11)(ret1)
+	ldw	HPPA_FRAME_ARG(4)(r1), arg2
+	ldw	HPPA_FRAME_ARG(5)(r1), arg3
+	ldw	HPPA_FRAME_ARG(6)(r1), t1
+	ldw	HPPA_FRAME_ARG(7)(r1), t2
+	ldw	HPPA_FRAME_ARG(8)(r1), t3
+	ldw	HPPA_FRAME_ARG(9)(r1), t4
+	stw	t1, HPPA_FRAME_ARG(4)(sp)
+	stw	t2, HPPA_FRAME_ARG(5)(sp)
+	stw	t3, HPPA_FRAME_ARG(6)(sp)
+	stw	t4, HPPA_FRAME_ARG(7)(sp)
+	ldw	HPPA_FRAME_ARG(10)(r1), t1
+	ldw	HPPA_FRAME_ARG(11)(r1), t2
+	ldw	HPPA_FRAME_ARG(12)(r1), t3
+	ldw	HPPA_FRAME_ARG(13)(r1), t4
+	stw	t1, HPPA_FRAME_ARG(8)(sp)
+	stw	t2, HPPA_FRAME_ARG(9)(sp)
+	stw	t3, HPPA_FRAME_ARG(10)(sp)
+	stw	t4, HPPA_FRAME_ARG(11)(sp)
 
 	/* save temp control regs */
 	mfctl	cr24, t1
 	mfctl	cr25, t2
 	mfctl	cr26, t3
 	mfctl	cr27, t4	
-	stw	t1, HPPA_FRAME_ARG(12)(ret1)
-	stw	t2, HPPA_FRAME_ARG(13)(ret1)
-	stw	t3, HPPA_FRAME_ARG(14)(ret1)
-	stw	t4, HPPA_FRAME_ARG(15)(ret1)
+	stw	t1, HPPA_FRAME_ARG(12)(sp)
+	stw	t2, HPPA_FRAME_ARG(13)(sp)
+	stw	t3, HPPA_FRAME_ARG(14)(sp)
+	stw	t4, HPPA_FRAME_ARG(15)(sp)
 	mfctl	cr28, t1
 	mfctl	cr29, t2
 	mfctl	cr30, t3
 	mfctl	cr31, t4	
-	stw	t1, HPPA_FRAME_ARG(16)(ret1)
-	stw	t2, HPPA_FRAME_ARG(17)(ret1)
-	stw	t3, HPPA_FRAME_ARG(18)(ret1)
-	stw	t4, HPPA_FRAME_ARG(19)(ret1)
+	stw	t1, HPPA_FRAME_ARG(16)(sp)
+	stw	t2, HPPA_FRAME_ARG(17)(sp)
+	stw	t3, HPPA_FRAME_ARG(18)(sp)
+	stw	t4, HPPA_FRAME_ARG(19)(sp)
 
 	comb,=	r0, ret0, pdc_call_unmapped2
 	nop
 
 	copy	arg0, t4
-	mfctl	eiem, t1
-	mtctl	r0, eiem
-
 	ldi	PSW_Q, arg0 /* (!pdc_flag && args[0] == PDC_PIM)? PSW_M:0) */
 	break	HPPA_BREAK_KERNEL, HPPA_BREAK_SET_PSW
 	nop
-	stw	t1, HPPA_FRAME_ARG(22)(ret1)
-	stw	ret0, HPPA_FRAME_ARG(23)(ret1)
+	stw	ret0, HPPA_FRAME_ARG(23)(sp)
 	copy	t4, arg0
 
 pdc_call_unmapped2
-	copy	ret1, sp
 	.call
 	blr	r0, rp
 	bv,n    (r31)
@@ -412,27 +413,22 @@ pdc_call_unmapped2
 	mtctl	t3, cr30
 	mtctl	t4, cr31
 
-	copy	sp, t4	
-	ldw	HPPA_FRAME_PSP(t4), sp
-
-	ldil	L%kernelmapped, t1
-	ldw	R%kernelmapped(t1), t1
+	ldw	HPPA_FRAME_ARG(21)(sp), t1
+	ldw	HPPA_FRAME_ARG(22)(sp), t2
 	comb,=	r0, t1, pdc_call_unmapped3
 	nop
 
 	copy	ret0, t3
-	ldw	HPPA_FRAME_ARG(22)(t4), t1
-	ldw	HPPA_FRAME_ARG(23)(t4), arg0
+	ldw	HPPA_FRAME_ARG(23)(sp), arg0
 	break	HPPA_BREAK_KERNEL, HPPA_BREAK_SET_PSW
 	nop
 	copy	t3, ret0
-	mtctl	t1, eiem
 
 pdc_call_unmapped3
-	ldw	HPPA_FRAME_ARG(0)(sp), t1
+	ldw	HPPA_FRAME_PSP(sp), sp
 	ldw	HPPA_FRAME_CRP(sp), rp
 	bv	r0(rp)
-	mtctl	t1, eiem
+	mtctl	t2, eiem
 EXIT(pdc_call)
 
 /*
@@ -540,46 +536,45 @@ $syscall
 	stw	r1, TF_CR15-TF_SIZE(sr1, t3)	/* eiem */
 	mtctl	r1, eiem
 
-	stw	r2 , TF_R2 -TF_SIZE(sr1, t3)
+	/*
+	 * we beleive that any calee-save registers
+	 * will be saved accordingly in either syscall()
+	 * or deeper called functions
+	 */
 	stw	r3 , TF_R3 -TF_SIZE(sr1, t3)
-	stw	r4 , TF_R4 -TF_SIZE(sr1, t3)
-	stw	r5 , TF_R5 -TF_SIZE(sr1, t3)
-	stw	r6 , TF_R6 -TF_SIZE(sr1, t3)
-	stw	r7 , TF_R7 -TF_SIZE(sr1, t3)
-	stw	r8 , TF_R8 -TF_SIZE(sr1, t3)
-	stw	r9 , TF_R9 -TF_SIZE(sr1, t3)
-	stw	r10, TF_R10-TF_SIZE(sr1, t3)
-	stw	r11, TF_R11-TF_SIZE(sr1, t3)
-	stw	r12, TF_R12-TF_SIZE(sr1, t3)
-	stw	r13, TF_R13-TF_SIZE(sr1, t3)
-	stw	r14, TF_R14-TF_SIZE(sr1, t3)
-	stw	r15, TF_R15-TF_SIZE(sr1, t3)
-	stw	r16, TF_R16-TF_SIZE(sr1, t3)
-	stw	r17, TF_R17-TF_SIZE(sr1, t3)
-	stw	r18, TF_R18-TF_SIZE(sr1, t3)
 	stw	r27, TF_R27-TF_SIZE(sr1, t3)	/* dp */
+#ifdef DDB
+	stw	r0, 0(t3)	/* terminate frame */
+	copy	r0 , r3
+	stw	r0, HPPA_FRAME_PSP(sp)
+	stw	r0, HPPA_FRAME_CRP(sp)
+#endif
 
-	/* copy arguments */
-	stw	arg0,  0(sr1, t3)
-	stw	arg1,  4(sr1, t3)
-	stw	arg2,  8(sr1, t3)
-	stw	arg3, 12(sr1, t3)
+	/*
+	 * Copy Arguments
+	 * unfortunately mmap() under bsd requires 7 words;
+	 * linux is confined to 5, and hpux to 6.
+	 * assuming the `long' syscall it gives us the maximum
+	 * 9 words, which very much overkill for an average of 3.
+	 * we keep it at 10, since bundling will keep it
+	 * at the same speed as 9 anyway.
+	 */
+	stw	arg0, 1*4(sr1, t3)
+	stw	arg1, 2*4(sr1, t3)
+	stw	arg2, 3*4(sr1, t3)
+	stw	arg3, 4*4(sr1, t3)
 	ldw	HPPA_FRAME_ARG( 4)(t4), arg0
 	ldw	HPPA_FRAME_ARG( 5)(t4), arg1
 	ldw	HPPA_FRAME_ARG( 6)(t4), arg2
 	ldw	HPPA_FRAME_ARG( 7)(t4), arg3
-	stw	arg0, 16(sr1, t3)
-	stw	arg1, 20(sr1, t3)
-	stw	arg2, 24(sr1, t3)
-	stw	arg3, 28(sr1, t3)
+	stw	arg0, 5*4(sr1, t3)
+	stw	arg1, 6*4(sr1, t3)
+	stw	arg2, 7*4(sr1, t3)
+	stw	arg3, 8*4(sr1, t3)
 	ldw	HPPA_FRAME_ARG( 8)(t4), arg0
 	ldw	HPPA_FRAME_ARG( 9)(t4), arg1
-	ldw	HPPA_FRAME_ARG(10)(t4), arg2
-	ldw	HPPA_FRAME_ARG(11)(t4), arg3
-	stw	arg0, 32(sr1, t3)
-	stw	arg1, 36(sr1, t3)
-	stw	arg2, 40(sr1, t3)
-	stw	arg3, 44(sr1, t3)
+	stw	arg0, 9*4(sr1, t3)
+	stw	arg1,10*4(sr1, t3)
 
 	/*
 	 * Save the rest of the CPU context
@@ -655,12 +650,8 @@ $syscall
 	mtsp	r0, sr6
 	mtsp	r0, sr7
 
-	/* setup frame */
-	stw	r0, HPPA_FRAME_PSP(sp)
-	stw	r0, HPPA_FRAME_CRP(sp)
-
 	ldo	-TF_SIZE(t3), arg0
-	copy	t3, arg1
+	ldo	4(t3), arg1
 
 	ldil	L%$global$,dp
 	ldo	R%$global$(dp),dp
@@ -708,10 +699,11 @@ $syscall_return_copy_loop
 
 	/* 1b. restore most of the general registers */
 	ldw	TF_CR11(t3), t1
-	ldw	TF_R1(t3), r1
 	mtctl	t1, sar
+	ldw	TF_R1(t3), r1
 	ldw	TF_R2(t3), r2
 	ldw	TF_R3(t3), r3
+#ifdef DDB
 	ldw	TF_R4(t3), r4
 	ldw	TF_R5(t3), r5
 	ldw	TF_R6(t3), r6
@@ -727,10 +719,11 @@ $syscall_return_copy_loop
 	ldw	TF_R16(t3), r16
 	ldw	TF_R17(t3), r17
 	ldw	TF_R18(t3), r18
-	ldw	TF_R19(t3), r19
+#endif
+	ldw	TF_R19(t3), t4
 	/*	r20(t3) is used as a temporary and will be restored later */
-	/*	r21(t3) is used as a temporary and will be restored later */
-	/*	r22(t3) is used as a temporary and will be restored later */
+	/*	r21(t2) is used as a temporary and will be restored later */
+	/*	r22(t1) is used as a temporary and will be restored later */
 	ldw	TF_R23(t3), r23
 	ldw	TF_R24(t3), r24
 	ldw	TF_R25(t3), r25
@@ -1526,34 +1519,41 @@ $trapnowvirt
 	 * Save all general registers that we haven't saved already
 	 */
 
+#ifdef DDB
+	stw	rp, HPPA_FRAME_CRP(sp)
+	stw	r0, -HPPA_FRAME_SIZE(sp)
+#endif
+	stw	t3, -HPPA_FRAME_SIZE+4(sp)
+	/*
+	 * hmm, we don't need to save all the regs, only caller-save
+	 * (except when DDB)
+	 */
 	mfctl	sar, t1
 	stw	t1, TF_CR11(t3)
-	stw	r1,TF_R1(t3)
-	stw	r2,TF_R2(t3)
+	stw	r1, TF_R1(t3)
+	stw	r2, TF_R2(t3)
+	stw	r3, TF_R3(t3)
 #ifdef DDB
-	stw	rp, HPPA_FRAME_CRP(sp)
-	stw	r3, -HPPA_FRAME_SIZE(sp)
+	stw	r4, TF_R4(t3)
+	stw	r5, TF_R5(t3)
+	stw	r6, TF_R6(t3)
+	stw	r7, TF_R7(t3)
+	stw	r8, TF_R8(t3)
+	stw	r9, TF_R9(t3)
+	stw	t10, TF_R10(t3)
+	stw	t11, TF_R11(t3)
+	stw	t12, TF_R12(t3)
+	stw	t13, TF_R13(t3)
+	stw	t14, TF_R14(t3)
+	stw	t15, TF_R15(t3)
+	stw	t16, TF_R16(t3)
+	stw	t17, TF_R17(t3)
+	stw	t18, TF_R18(t3)
 #endif
-	stw	r3,TF_R3(t3)
-	stw	r4,TF_R4(t3)
-	stw	r5,TF_R5(t3)
-	stw	r6,TF_R6(t3)
-	stw	r7,TF_R7(t3)
-	stw	r8,TF_R8(t3)
-	stw	r9,TF_R9(t3)
-	stw	r10,TF_R10(t3)
-	stw	r11,TF_R11(t3)
-	stw	r12,TF_R12(t3)
-	stw	r13,TF_R13(t3)
-	stw	r14,TF_R14(t3)
-	stw	r15,TF_R15(t3)
-	stw	r16,TF_R16(t3)
-	stw	r17,TF_R17(t3)
-	stw	r18,TF_R18(t3)
-	stw	r19,TF_R19(t3)
-	/*	r20 already saved (t1)
+	stw	t4, TF_R19(t3)
+	/*	r20 already saved (t3)
 	 *	r21 already saved (t2)
-	 *	r22 already saved (t3) */
+	 *	r22 already saved (t1) */
 	stw	r23,TF_R23(t3)
 	stw	r24,TF_R24(t3)
 	stw	r25,TF_R25(t3)
@@ -1577,7 +1577,7 @@ $trap_copy_loop
 	stwm	t1, 4(arg1)
 
 	/*
-	 * Save the necessary control registers that were not already saved.
+	 * Save the necessary control registers that have not already saved.
 	 */
 
 	mfctl	rctr, t1
@@ -1622,22 +1622,22 @@ $trap_copy_loop
 	.import	trap, code
 	ldil	L%trap,t1
 	ldo	R%trap(t1),t1
-	copy	t3, r5
 	.call
 	blr	r0,rp
 	bv,n	r0(t1)
 	nop
 
+	ldw	-HPPA_FRAME_SIZE+4(sp), t3
 	/* see if curproc have changed */
-	ldw	TF_FLAGS(r5), arg0
+	ldw	TF_FLAGS(t3), arg0
 	bb,>=,n	arg0, TFF_LAST_POS, $syscall_return
-	copy	r5, t3
+	nop
 
 	/* see if curproc have really changed */
 	ldil	L%curproc, t1
 	ldw	R%curproc(t1), t2
 	comb,=,n r0, t2, $syscall_return
-	copy	r5, t3
+	ldw	-HPPA_FRAME_SIZE+4(sp), t3
 
 	/* means curproc have actually changed */
 	b	$syscall_return
@@ -2152,7 +2152,7 @@ ENTRY(switch_exit)
 	/* XXX we need to switch to some stupid stack here */
 
 	b	switch_search
-	nop
+	copy	r0, arg2	/* no old proc */
 EXIT(switch_exit)
 
 ENTRY(switch_trampoline)
-- 
2.20.1