Implement mcount trampoline (untested).

1 files changed, 64 insertions, 114 deletions

diff --git a/sys/ia64/include/profile.h b/sys/ia64/include/profile.h
index 70a9cf7..038c805 100644
--- a/sys/ia64/include/profile.h
+++ b/sys/ia64/include/profile.h
@@ -34,126 +34,76 @@
 
 typedef u_long	fptrdiff_t;
 
-#if 0
-/*
- * XXX The definition of MCOUNT below is really the following code, run
- * XXX through cpp, since the inline assembly isn't preprocessed.
- */
-#define	OFFSET_AT	0
-#define OFFSET_V0	8
-#define OFFSET_T0	16
-#define OFFSET_T1	24
-#define OFFSET_T2	32
-#define OFFSET_T3	40
-#define OFFSET_T4	48
-#define OFFSET_T5	56
-#define OFFSET_T6	64
-#define OFFSET_T7	72
-#define OFFSET_S6	80
-#define OFFSET_A0	88
-#define OFFSET_A1	96
-#define OFFSET_A2	104
-#define OFFSET_A3	112
-#define OFFSET_A4	120
-#define OFFSET_A5	128
-#define OFFSET_T8	136
-#define OFFSET_T9	144
-#define OFFSET_T10	152
-#define OFFSET_T11	160
-#define OFFSET_RA	168
-#define OFFSET_T12	176
-#define OFFSET_GP	184
-#define	FRAME_SIZE	192
-
-LEAF(_mcount,0)			/* XXX */
-	.set noat
-	.set noreorder
-
-	lda	sp, -FRAME_SIZE(sp)
-
-	stq	at_reg, OFFSET_AT(sp)
-	stq	v0, OFFSET_V0(sp)
-	stq	t0, OFFSET_T0(sp)
-	stq	t1, OFFSET_T1(sp)
-	stq	t2, OFFSET_T2(sp)
-	stq	t3, OFFSET_T3(sp)
-	stq	t4, OFFSET_T4(sp)
-	stq	t5, OFFSET_T5(sp)
-	stq	t6, OFFSET_T6(sp)
-	stq	t7, OFFSET_T7(sp)
-	stq	s6, OFFSET_S6(sp)	/* XXX because run _after_ prologue. */
-	stq	a0, OFFSET_A0(sp)
-	stq	a1, OFFSET_A1(sp)
-	stq	a2, OFFSET_A2(sp)
-	stq	a3, OFFSET_A3(sp)
-	stq	a4, OFFSET_A4(sp)
-	stq	a5, OFFSET_A5(sp)
-	stq	t8, OFFSET_T8(sp)
-	stq	t9, OFFSET_T9(sp)
-	stq	t10, OFFSET_T10(sp)
-	stq	t11, OFFSET_T11(sp)
-	stq	ra, OFFSET_RA(sp)
-	stq	t12, OFFSET_T12(sp)
-	stq	gp, OFFSET_GP(sp)
-
-	br	pv, LX99
-LX99:	SETGP(pv)
-	mov	ra, a0
-	mov	at_reg, a1
-	CALL(mcount)
-
-	ldq	v0, OFFSET_V0(sp)
-	ldq	t0, OFFSET_T0(sp)
-	ldq	t1, OFFSET_T1(sp)
-	ldq	t2, OFFSET_T2(sp)
-	ldq	t3, OFFSET_T3(sp)
-	ldq	t4, OFFSET_T4(sp)
-	ldq	t5, OFFSET_T5(sp)
-	ldq	t6, OFFSET_T6(sp)
-	ldq	t7, OFFSET_T7(sp)
-	ldq	s6, OFFSET_S6(sp)	/* XXX because run _after_ prologue. */
-	ldq	a0, OFFSET_A0(sp)
-	ldq	a1, OFFSET_A1(sp)
-	ldq	a2, OFFSET_A2(sp)
-	ldq	a3, OFFSET_A3(sp)
-	ldq	a4, OFFSET_A4(sp)
-	ldq	a5, OFFSET_A5(sp)
-	ldq	t8, OFFSET_T8(sp)
-	ldq	t9, OFFSET_T9(sp)
-	ldq	t10, OFFSET_T10(sp)
-	ldq	t11, OFFSET_T11(sp)
-	ldq	ra, OFFSET_RA(sp)
-	stq	t12, OFFSET_T12(sp)
-	ldq	gp, OFFSET_GP(sp)
-
-	ldq	at_reg, OFFSET_AT(sp)
-
-	lda	sp, FRAME_SIZE(sp)
-	ret	zero, (at_reg), 1
-
-	END(_mcount)
-#endif /* 0 */
-
-#define MCOUNT __asm ("		\
-	.globl	_mcount;	\
-	.proc	_mcount;	\
-_mcount:;			\
-				\
+#define MCOUNT __asm ("							\n\
+	.globl	_mcount							\n\
+	.proc	_mcount							\n\
+_mcount:								\n\
+	alloc	loc0=ar.pfs,8,7,2,0	// space to save r8-r11,rp,b7	\n\
+	add	sp=-8*16,sp		// space to save f8-f15		\n\
+	mov	loc1=rp			// caller's return address	\n\
+	mov	loc2=b7			// our return back to caller	\n\
+	;;								\n\
+	add	r17=16,sp		// leave 16 bytes for mcount	\n\
+	add	r18=32,sp						\n\
+	;;								\n\
+	mov	loc3=r8			// structure return address	\n\
+	mov	loc4=r9			// language specific		\n\
+	mov	loc5=r10		// language specific		\n\
+	mov	loc6=r11		// language specific		\n\
+	;;								\n\
+	stf.spill [r17]=f8,32		// save float arguments		\n\
+	stf.spill [r18]=f9,32						\n\
+	mov	out0=rp			// frompc			\n\
+	;;								\n\
+	stf.spill [r17]=f10,32						\n\
+	stf.spill [r18]=f11,32						\n\
+	mov	out1=b7			// selfpc			\n\
+	;;								\n\
+	stf.spill [r17]=f12,32						\n\
+	stf.spill [r18]=f13,32						\n\
+	;;								\n\
+	stf.spill [r17]=f14,32						\n\
+	stf.spill [r18]=f15,32						\n\
+	;;								\n\
+	br.call.sptk.many rp=mcount					\n\
+	;;								\n\
+	add	r17=16,sp						\n\
+	add	r18=32,sp						\n\
+	;;								\n\
+	ldf.fill f8=[r17],32						\n\
+	ldf.fill f9=[r18],32						\n\
+	mov	r8=loc3			// restore structure pointer	\n\
+	;;								\n\
+	ldf.fill f10=[r17],32		// restore float arguments	\n\
+	ldf.fill f11=[r18],32						\n\
+	mov	r9=loc4							\n\
+	;;								\n\
+	ldf.fill f12=[r17],32		// etc.				\n\
+	ldf.fill f13=[r18],32						\n\
+	mov	r10=loc5						\n\
+	;;								\n\
+	ldf.fill f14=[r17],32						\n\
+	ldf.fill f15=[r18],32						\n\
+	mov	r11=loc6						\n\
+	;;								\n\
+	mov	b7=loc2			// clean up			\n\
+	mov	rp=loc1							\n\
+	mov	ar.pfs=loc0						\n\
+	;;								\n\
+	alloc	r14=ar.pfs,0,0,8,0	// drop our register frame	\n\
+	br.sptk.many b7			// back to caller		\n\
+									\n\
 	.end	_mcount");
 
 #ifdef _KERNEL
 /*
  * The following two macros do splhigh and splx respectively.
- * _alpha_pal_swpipl is a special version of alpha_pal_swpipl which
- * doesn't include profiling support.
- *
- * XXX These macros should probably use inline assembly.
  */
-#define MCOUNT_ENTER(s) \
-	s = _alpha_pal_swpipl(ALPHA_PSL_IPL_HIGH)
-#define MCOUNT_EXIT(s) \
-	(void)_alpha_pal_swpipl(s);
-#define	MCOUNT_DECL(s)	u_long s;
+#define MCOUNT_ENTER(s) \n\
+	_c = critical_enter()
+#define MCOUNT_EXIT(s) \n\
+	(void)critical_exit(_c)
+#define	MCOUNT_DECL(s)	critical_t c;
 #ifdef GUPROF
 struct gmonparam;