Use double load/store instructions whenever possible.

2 files changed, 45 insertions, 72 deletions

diff --git a/lib/libpthread/arch/m88k/uthread_machdep.c b/lib/libpthread/arch/m88k/uthread_machdep.c
index 435e976d04e..c73fe479eb0 100644
--- a/lib/libpthread/arch/m88k/uthread_machdep.c
+++ b/lib/libpthread/arch/m88k/uthread_machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uthread_machdep.c,v 1.1 2004/02/27 22:19:46 deraadt Exp $	*/
+/*	$OpenBSD: uthread_machdep.c,v 1.2 2004/03/02 23:41:29 miod Exp $	*/
 
 /*
  * Copyright (c) 2004 Theo de Raadt
@@ -26,7 +26,7 @@
 #define ALIGNBYTES	7
 
 struct frame {
-	long	regs[30];
+	long	regs[28];	/* r4-r30, r1 */
 };
 
 /*
@@ -43,7 +43,7 @@ _thread_machdep_init(statep, base, len, entry)
 	struct frame *f;
 
 	f = (struct frame *)(((u_int32_t)base + len - sizeof *f) & ~ALIGNBYTES);
-	f->regs[29] = f->regs[27] = (u_int32_t)entry;
+	f->regs[27] = (u_int32_t)entry;		/* ``saved'' r1 */
 
 	statep->sp = (int)f;
 }
diff --git a/lib/libpthread/arch/m88k/uthread_machdep_asm.S b/lib/libpthread/arch/m88k/uthread_machdep_asm.S
index 9a74e5a7fae..dcf997650d8 100644
--- a/lib/libpthread/arch/m88k/uthread_machdep_asm.S
+++ b/lib/libpthread/arch/m88k/uthread_machdep_asm.S
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uthread_machdep_asm.S,v 1.1 2004/02/27 22:19:46 deraadt Exp $	*/
+/*	$OpenBSD: uthread_machdep_asm.S,v 1.2 2004/03/02 23:41:29 miod Exp $	*/
 
 /*
  * Copyright (c) 2004 Theo de Raadt
@@ -18,76 +18,49 @@
 
 #include <machine/asm.h>
 
-	global	__thread_machdep_switch
-	.type	__thread_machdep_switch, @function
-__thread_machdep_switch:
-	subu    r31,r31,4*30
+/*
+ * void _thread_machdep_switch(new, oldsave);
+ */
+ENTRY(_thread_machdep_switch)
+	subu	r31, r31, 4 * 28
 
-	st	r1,r31,4*29
-	st	r3,r31,4*0
-	st	r4,r31,4*1
-	st	r5,r31,4*2
-	st	r6,r31,4*3
-	st	r7,r31,4*4
-	st	r8,r31,4*5
-	st	r9,r31,4*6
-	st	r10,r31,4*7
-	st	r11,r31,4*8
-	st	r12,r31,4*9
-	st	r13,r31,4*10
-	st	r14,r31,4*11
-	st	r15,r31,4*12
-	st	r16,r31,4*13
-	st	r17,r31,4*14
-	st	r18,r31,4*15
-	st	r19,r31,4*16
-	st	r20,r31,4*17
-	st	r21,r31,4*18
-	st	r22,r31,4*19
-	st	r23,r31,4*20
-	st	r24,r31,4*21
-	st	r25,r31,4*22
-	st	r26,r31,4*23
-	st	r27,r31,4*24
-	st	r28,r31,4*25
-	st	r29,r31,4*26
-	st	r30,r31,4*27
+	/* save all registers but r0, r2, r3 and r31 on stack */
+	st.d	r4,  r31, 4 * 0
+	st.d	r6,  r31, 4 * 2
+	st.d	r8,  r31, 4 * 4
+	st.d	r10, r31, 4 * 6
+	st.d	r12, r31, 4 * 8
+	st.d	r14, r31, 4 * 10
+	st.d	r16, r31, 4 * 12
+	st.d	r18, r31, 4 * 14
+	st.d	r20, r31, 4 * 16
+	st.d	r22, r31, 4 * 18
+	st.d	r24, r31, 4 * 20
+	st.d	r26, r31, 4 * 22
+	st.d	r28, r31, 4 * 24
+	st	r30, r31, 4 * 26
+	st	r1,  r31, 4 * 27
 	
-	st	r31,r0,r3
-	ld	r31,r0,r2
+	/* exchange stacks */
+	st	r31, r3,  0	/* oldsave->frame = r31 */
+	ld	r31, r2,  0	/* r31 = oldsave->frame */
 
-	ld	r1,r31,4*29
-	ld	r3,r31,4*0
-	ld	r4,r31,4*1
-	ld	r5,r31,4*2
-	ld	r6,r31,4*3
-	ld	r7,r31,4*4
-	ld	r8,r31,4*5
-	ld	r9,r31,4*6
-	ld	r10,r31,4*7
-	ld	r11,r31,4*8
-	ld	r12,r31,4*9
-	ld	r13,r31,4*10
-	ld	r14,r31,4*11
-	ld	r15,r31,4*12
-	ld	r16,r31,4*13
-	ld	r17,r31,4*14
-	ld	r18,r31,4*15
-	ld	r19,r31,4*16
-	ld	r20,r31,4*17
-	ld	r21,r31,4*18
-	ld	r22,r31,4*19
-	ld	r23,r31,4*20
-	ld	r24,r31,4*21
-	ld	r25,r31,4*22
-	ld	r26,r31,4*23
-	ld	r27,r31,4*24
-	ld	r28,r31,4*25
-	ld	r29,r31,4*26
-	ld	r30,r31,4*27
+	/* restore registers */
+	ld.d	r4,  r31, 4 * 0
+	ld.d	r6,  r31, 4 * 2
+	ld.d	r8,  r31, 4 * 4
+	ld.d	r10, r31, 4 * 6
+	ld.d	r12, r31, 4 * 8
+	ld.d	r14, r31, 4 * 10
+	ld.d	r16, r31, 4 * 12
+	ld.d	r18, r31, 4 * 14
+	ld.d	r20, r31, 4 * 16
+	ld.d	r22, r31, 4 * 18
+	ld.d	r24, r31, 4 * 20
+	ld.d	r26, r31, 4 * 22
+	ld.d	r28, r31, 4 * 24
+	ld	r30, r31, 4 * 26
+	ld	r1,  r31, 4 * 27
 
 	jmp.n	r1
-	addu    r31,r31,4*30
-
-	.size	__thread_machdep_switch , . - __thread_machdep_switch
-	.align 8
+	 addu	r31, r31, 4 * 28