Intermediate, fully functional, rework for variadic functions

* check/cva_list.c: New file implementing a test to ensure
	the value returned by jit_va_start is a valid C va_list.

	* check/va_list.ok: New simple helper file, as now the
	va_list.tst test is enabled.

	* check/va_list.tst: Rewritten for an extensive variadic
	jit functions test.

	* check/Makefile.am: Update for the new tests.

	* lib/jit_arm-cpu.c, lib/jit_arm-swf.c, lib/jit_arm-vfp.c,
	lib/jit_arm.c: Correct broken software float in a previous
	commit. Note that the hard float abi implementation is known
	broken at this time, for special cases involving variadic
	functions, and should be corrected next.

	lib/jit_x86-cpu.c, lib/jit_x86-sz.c, lib/jit_x86.c: Correct
	the jit_va_list_t semantics to match C va_list.

ChangeLog

@@ -1,3 +1,25 @@
+2015-06-25 Paulo Andrade <pcpa@gnu.org>
+
+	* check/cva_list.c: New file implementing a test to ensure
+	the value returned by jit_va_start is a valid C va_list.
+
+	* check/va_list.ok: New simple helper file, as now the
+	va_list.tst test is enabled.
+
+	* check/va_list.tst: Rewritten for an extensive variadic
+	jit functions test.
+
+	* check/Makefile.am: Update for the new tests.
+
+	* lib/jit_arm-cpu.c, lib/jit_arm-swf.c, lib/jit_arm-vfp.c,
+	lib/jit_arm.c: Correct broken software float in a previous
+	commit. Note that the hard float abi implementation is known
+	broken at this time, for special cases involving variadic
+	functions, and should be corrected next.
+
+	lib/jit_x86-cpu.c, lib/jit_x86-sz.c, lib/jit_x86.c: Correct
+	the jit_va_list_t semantics to match C va_list.
+
 2015-06-24 Paulo Andrade <pcpa@gnu.org>
 
 	* lib/Makefile.am: Bump library major. This is a preparation

check/Makefile.am

@@ -16,7 +16,7 @@
 
 AM_CFLAGS = -I$(top_srcdir)/include -D_GNU_SOURCE
 
-check_PROGRAMS = lightning ccall self setcode nodata ctramp carg
+check_PROGRAMS = lightning ccall self setcode nodata ctramp carg cva_list
 
 lightning_LDADD = $(top_builddir)/lib/liblightning.la -lm $(SHLIB)
 lightning_SOURCES = lightning.c
@@ -39,6 +39,9 @@ ctramp_SOURCES = ctramp.c
 carg_LDADD = $(top_builddir)/lib/liblightning.la -lm $(SHLIB)
 carg_SOURCES = carg.c
 
+cva_list_LDADD = $(top_builddir)/lib/liblightning.la -lm $(SHLIB)
+cva_list_SOURCES = cva_list.c
+
 $(top_builddir)/lib/liblightning.la:
 	cd $(top_builddir)/lib; $(MAKE) $(AM_MAKEFLAGS) liblightning.la
 
@@ -96,6 +99,7 @@ EXTRA_DIST =				\
 	ranger.tst	ranger.ok	\
 	ret.tst		ret.ok		\
 	tramp.tst	tramp.ok	\
+	va_list.tst	va_list.ok	\
 	check.sh			\
 	check.x87.sh			\
 	check.arm.sh	check.swf.sh	\
@@ -123,7 +127,8 @@ base_TESTS =				\
 	clobber carry call		\
 	float jmpr put			\
 	qalu_mul qalu_div		\
-	range ranger ret tramp
+	range ranger ret tramp		\
+	va_list
 
 $(base_TESTS):	check.sh
 	$(LN_S) $(srcdir)/check.sh $@
@@ -148,7 +153,8 @@ x87_TESTS =					\
 	fop_abs.x87 fop_sqrt.x87		\
 	varargs.x87 stack.x87			\
 	clobber.x87 carry.x87 call.x87		\
-	float.x87 jmpr.x87 put.x87
+	float.x87 jmpr.x87 put.x87		\
+	va_list.x87
 $(x87_TESTS):	check.x87.sh
 	$(LN_S) $(srcdir)/check.x87.sh $@
 TESTS += $(x87_TESTS)
@@ -170,7 +176,8 @@ x87_nodata_TESTS =							\
 	fop_abs.x87.nodata fop_sqrt.x87.nodata				\
 	varargs.x87.nodata stack.x87.nodata				\
 	clobber.x87.nodata carry.x87.nodata call.x87.nodata		\
-	float.x87.nodata jmpr.x87.nodata put.x87.nodata
+	float.x87.nodata jmpr.x87.nodata put.x87.nodata			\
+	va_list.x87.nodata
 $(x87_nodata_TESTS):	check.x87.nodata.sh
 	$(LN_S) $(srcdir)/check.x87.nodata.sh $@
 TESTS += $(x87_nodata_TESTS)
@@ -195,7 +202,7 @@ arm_TESTS =					\
 	varargs.arm stack.arm			\
 	clobber.arm carry.arm call.arm		\
 	float.arm jmpr.arm tramp.arm range.arm	\
-	ranger.arm put.arm
+	ranger.arm put.arm va_list.arm
 $(arm_TESTS):	check.arm.sh
 	$(LN_S) $(srcdir)/check.arm.sh $@
 TESTS += $(arm_TESTS)
@@ -220,7 +227,7 @@ swf_TESTS =					\
 	varargs.swf stack.swf			\
 	clobber.swf carry.swf call.swf		\
 	float.swf jmpr.swf tramp.swf range.swf	\
-	ranger.swf put.swf
+	ranger.swf put.swf va_list.swf
 $(swf_TESTS):	check.swf.sh
 	$(LN_S) $(srcdir)/check.swf.sh $@
 TESTS += $(swf_TESTS)
@@ -243,7 +250,7 @@ arm_swf_TESTS =							\
 	varargs.arm.swf stack.arm.swf				\
 	clobber.arm.swf carry.arm.swf call.arm.swf		\
 	float.arm.swf jmpr.arm.swf tramp.arm.swf range.arm.swf	\
-	ranger.arm.swf put.arm.swf
+	ranger.arm.swf put.arm.swf va_list.arm.swf
 $(arm_swf_TESTS):	check.arm.swf.sh
 	$(LN_S) $(srcdir)/check.arm.swf.sh $@
 TESTS += $(arm_swf_TESTS)
@@ -267,7 +274,8 @@ arm4_swf_TESTS =						\
 	varargs.arm4.swf stack.arm4.swf				\
 	clobber.arm4.swf carry.arm4.swf call.arm4.swf		\
 	float.arm4.swf jmpr.arm4.swf tramp.arm4.swf		\
-	range.arm4.swf ranger.arm4.swf put.arm4.swf
+	range.arm4.swf ranger.arm4.swf put.arm4.swf		\
+	va_list.arm4.swf
 $(arm4_swf_TESTS):	check.arm4.swf.sh
 	$(LN_S) $(srcdir)/check.arm4.swf.sh $@
 TESTS += $(arm4_swf_TESTS)
@@ -293,13 +301,14 @@ nodata_TESTS =						\
 	varargs.nodata stack.nodata			\
 	clobber.nodata carry.nodata call.nodata		\
 	float.nodata jmpr.nodata tramp.nodata		\
-	range.nodata ranger.nodata put.nodata
+	range.nodata ranger.nodata put.nodata		\
+	va_list.nodata
 $(nodata_TESTS):	check.nodata.sh
 	$(LN_S) $(srcdir)/check.nodata.sh $@
 TESTS += $(nodata_TESTS)
 endif
 
-TESTS += ccall self setcode nodata ctramp carg
+TESTS += ccall self setcode nodata ctramp carg cva_list
 CLEANFILES = $(TESTS)
 
 #TESTS_ENVIRONMENT=$(srcdir)/run-test;

check/va_list.ok

@@ -0,0 +1 @@
+ok

check/va_list.tst

@@ -1,30 +1,743 @@
-.data	16
+.data	8
-fmt:
+ok:
-.c	"%d %f\n"
+.c	"ok\n"
 .code
 	jmpi main
-varargs:
+
+#define BEGIN(L)						\
+L:								\
 	prolog
-	ellipsis
+#define VA_START()						\
+	ellipsis						\
 	va_start %v0
-	va_arg %r0 %v0
+#define VARG(L,N)						\
-	va_arg_d %f0 %v0
+	va_arg %r0 %v0						\
-	va_end %v0
+	beqi L##N %r0 N						\
-	prepare
+	calli @abort						\
-		pushargi fmt
+L##N:
-		ellipsis
+#define VARGD(L,N)						\
-		pushargr %r0
+	va_arg_d %f0 %v0					\
-		pushargr_d %f0
+	beqi_d L##N %f0 N					\
-	finishi @printf
+	calli @abort						\
-	ret
+L##N:
+#define VA_END()						\
+	va_end %v0						\
+	ret							\
 	epilog
+#define ARG(N)			arg $arg##N
+#define ARGD(N)			arg_d $arg##N
+#define GET(L,N)						\
+	getarg %r0 $arg##N					\
+	beqi L##N %r0 N						\
+	calli @abort						\
+L##N:
+#define GETD(L,N)						\
+	getarg_d %f0 $arg##N					\
+	beqi_d L##N %f0 N					\
+	calli @abort						\
+L##N:
+
+#define ARG1()				ARG(1)
+#define ARG2()		ARG1()		ARG(2)
+#define ARG3()		ARG2()		ARG(3)
+#define ARG4()		ARG3()		ARG(4)
+#define ARG5()		ARG4()		ARG(5)
+#define ARG6()		ARG5()		ARG(6)
+#define ARG7()		ARG6()		ARG(7)
+#define ARG8()		ARG7()		ARG(8)
+#define ARG9()		ARG8()		ARG(9)
+#define GET1(L)				GET(L,1)
+#define GET2(L)		GET1(L)		GET(L,2)
+#define GET3(L)		GET2(L)		GET(L,3)
+#define GET4(L)		GET3(L)		GET(L,4)
+#define GET5(L)		GET4(L)		GET(L,5)
+#define GET6(L)		GET5(L)		GET(L,6)
+#define GET7(L)		GET6(L)		GET(L,7)
+#define GET8(L)		GET7(L)		GET(L,8)
+#define GET9(L)		GET8(L)		GET(L,9)
+#define ARGD1()				ARGD(1)
+#define ARGD2()		ARGD1()		ARGD(2)
+#define ARGD3()		ARGD2()		ARGD(3)
+#define ARGD4()		ARGD3()		ARGD(4)
+#define ARGD5()		ARGD4()		ARGD(5)
+#define ARGD6()		ARGD5()		ARGD(6)
+#define ARGD7()		ARGD6()		ARGD(7)
+#define ARGD8()		ARGD7()		ARGD(8)
+#define ARGD9()		ARGD8()		ARGD(9)
+#define GETD1(L)			GETD(L,1)
+#define GETD2(L)	GETD1(L)	GETD(L,2)
+#define GETD3(L)	GETD2(L)	GETD(L,3)
+#define GETD4(L)	GETD3(L)	GETD(L,4)
+#define GETD5(L)	GETD4(L)	GETD(L,5)
+#define GETD6(L)	GETD5(L)	GETD(L,6)
+#define GETD7(L)	GETD6(L)	GETD(L,7)
+#define GETD8(L)	GETD7(L)	GETD(L,8)
+#define GETD9(L)	GETD8(L)	GETD(L,9)
+#define IDARG1()			ARG(1)
+#define IDARG2()	IDARG1()	ARGD(2)
+#define IDARG3()	IDARG2()	ARG(3)
+#define IDARG4()	IDARG3()	ARGD(4)
+#define IDARG5()	IDARG4()	ARG(5)
+#define IDARG6()	IDARG5()	ARGD(6)
+#define IDARG7()	IDARG6()	ARG(7)
+#define IDARG8()	IDARG7()	ARGD(8)
+#define IDARG9()	IDARG8()	ARG(9)
+#define IDGET1(L)			GET(L,1)
+#define IDGET2(L)	IDGET1(L)	GETD(L,2)
+#define IDGET3(L)	IDGET2(L)	GET(L,3)
+#define IDGET4(L)	IDGET3(L)	GETD(L,4)
+#define IDGET5(L)	IDGET4(L)	GET(L,5)
+#define IDGET6(L)	IDGET5(L)	GETD(L,6)
+#define IDGET7(L)	IDGET6(L)	GET(L,7)
+#define IDGET8(L)	IDGET7(L)	GETD(L,8)
+#define IDGET9(L)	IDGET8(L)	GET(L,9)
+#define DIARG1()			ARGD(1)
+#define DIARG2()	DIARG1()	ARG(2)
+#define DIARG3()	DIARG2()	ARGD(3)
+#define DIARG4()	DIARG3()	ARG(4)
+#define DIARG5()	DIARG4()	ARGD(5)
+#define DIARG6()	DIARG5()	ARG(6)
+#define DIARG7()	DIARG6()	ARGD(7)
+#define DIARG8()	DIARG7()	ARG(8)
+#define DIARG9()	DIARG8()	ARGD(9)
+#define DIGET1(L)			GETD(L,1)
+#define DIGET2(L)	DIGET1(L)	GET(L,2)
+#define DIGET3(L)	DIGET2(L)	GETD(L,3)
+#define DIGET4(L)	DIGET3(L)	GET(L,4)
+#define DIGET5(L)	DIGET4(L)	GETD(L,5)
+#define DIGET6(L)	DIGET5(L)	GET(L,6)
+#define DIGET7(L)	DIGET6(L)	GETD(L,7)
+#define DIGET8(L)	DIGET7(L)	GET(L,8)
+#define DIGET9(L)	DIGET8(L)	GETD(L,9)
+
+#define VARG1(L)						\
+	VARG(L, 10)
+#define VARG2(L)						\
+	VARG(L, 9)						\
+	VARG1(L)
+#define VARG3(L)						\
+	VARG(L, 8)						\
+	VARG2(L)
+#define VARG4(L)						\
+	VARG(L, 7)						\
+	VARG3(L)
+#define VARG5(L)						\
+	VARG(L, 6)						\
+	VARG4(L)
+#define VARG6(L)						\
+	VARG(L, 5)						\
+	VARG5(L)
+#define VARG7(L)						\
+	VARG(L, 4)						\
+	VARG6(L)
+#define VARG8(L)						\
+	VARG(L, 3)						\
+	VARG7(L)
+#define VARG9(L)						\
+	VARG(L, 2)						\
+	VARG8(L)
+#define VARG10(L)						\
+	VARG(L, 1)						\
+	VARG9(L)
+#define VARGD1(L)						\
+	VARGD(L, 10)
+#define VARGD2(L)						\
+	VARGD(L, 9)						\
+	VARGD1(L)
+#define VARGD3(L)						\
+	VARGD(L, 8)						\
+	VARGD2(L)
+#define VARGD4(L)						\
+	VARGD(L, 7)						\
+	VARGD3(L)
+#define VARGD5(L)						\
+	VARGD(L, 6)						\
+	VARGD4(L)
+#define VARGD6(L)						\
+	VARGD(L, 5)						\
+	VARGD5(L)
+#define VARGD7(L)						\
+	VARGD(L, 4)						\
+	VARGD6(L)
+#define VARGD8(L)						\
+	VARGD(L, 3)						\
+	VARGD7(L)
+#define VARGD9(L)						\
+	VARGD(L, 2)						\
+	VARGD8(L)
+#define VARGD10(L)						\
+	VARGD(L, 1)						\
+	VARGD9(L)
+#define IDVARG1(L)						\
+	VARGD(L, 10)
+#define IDVARG2(L)						\
+	VARG(L, 9)						\
+	IDVARG1(L)
+#define IDVARG3(L)						\
+	VARGD(L, 8)						\
+	IDVARG2(L)
+#define IDVARG4(L)						\
+	VARG(L, 7)						\
+	IDVARG3(L)
+#define IDVARG5(L)						\
+	VARGD(L, 6)						\
+	IDVARG4(L)
+#define IDVARG6(L)						\
+	VARG(L, 5)						\
+	IDVARG5(L)
+#define IDVARG7(L)						\
+	VARGD(L, 4)						\
+	IDVARG6(L)
+#define IDVARG8(L)						\
+	VARG(L, 3)						\
+	IDVARG7(L)
+#define IDVARG9(L)						\
+	VARGD(L, 2)						\
+	IDVARG8(L)
+#define IDVARG10(L)						\
+	VARG(L, 1)						\
+	IDVARG9(L)
+#define DIVARG1(L)						\
+	VARG(L, 10)
+#define DIVARG2(L)						\
+	VARGD(L, 9)						\
+	DIVARG1(L)
+#define DIVARG3(L)						\
+	VARG(L, 8)						\
+	DIVARG2(L)
+#define DIVARG4(L)						\
+	VARGD(L, 7)						\
+	DIVARG3(L)
+#define DIVARG5(L)						\
+	VARG(L, 6)						\
+	DIVARG4(L)
+#define DIVARG6(L)						\
+	VARGD(L, 5)						\
+	DIVARG5(L)
+#define DIVARG7(L)						\
+	VARG(L, 4)						\
+	DIVARG6(L)
+#define DIVARG8(L)						\
+	VARGD(L, 3)						\
+	DIVARG7(L)
+#define DIVARG9(L)						\
+	VARG(L, 2)						\
+	DIVARG8(L)
+#define DIVARG10(L)						\
+	VARGD(L, 1)						\
+	DIVARG9(L)
+
+BEGIN(_iiiiiiiiii)
+	VA_START()
+	VARG10(_iiiiiiiiii)
+	VA_END()
+BEGIN(i_iiiiiiiii)
+	ARG1()
+	GET1(i_iiiiiiiii)
+	VA_START()
+	VARG9(i_iiiiiiiii)
+	VA_END()
+BEGIN(ii_iiiiiiii)
+	ARG2()
+	GET2(ii_iiiiiiii)
+	VA_START()
+	VARG8(ii_iiiiiiii)
+	VA_END()
+BEGIN(iii_iiiiiii)
+	ARG3()
+	GET3(iii_iiiiiii)
+	VA_START()
+	VARG7(iii_iiiiiii)
+	VA_END()
+BEGIN(iiii_iiiiii)
+	ARG4()
+	GET4(iiii_iiiiii)
+	VA_START()
+	VARG6(iiii_iiiiii)
+	VA_END()
+BEGIN(iiiii_iiiii)
+	ARG5()
+	GET5(iiiii_iiiii)
+	VA_START()
+	VARG5(iiiii_iiiii)
+	VA_END()
+BEGIN(iiiiii_iiii)
+	ARG6()
+	GET6(iiiiii_iiii)
+	VA_START()
+	VARG4(iiiiii_iiii)
+	VA_END()
+BEGIN(iiiiiii_iii)
+	ARG7()
+	GET7(iiiiiii_iii)
+	VA_START()
+	VARG3(iiiiiii_iii)
+	VA_END()
+BEGIN(iiiiiiii_ii)
+	ARG8()
+	GET8(iiiiiiii_ii)
+	VA_START()
+	VARG2(iiiiiiii_ii)
+	VA_END()
+BEGIN(iiiiiiiii_i)
+	ARG9()
+	GET9(iiiiiiiii_i)
+	VA_START()
+	VARG1(iiiiiiiii_i)
+	VA_END()
+BEGIN(_dddddddddd)
+	VA_START()
+	VARGD10(_dddddddddd)
+	VA_END()
+BEGIN(d_ddddddddd)
+	ARGD1()
+	GETD1(d_ddddddddd)
+	VA_START()
+	VARGD9(d_ddddddddd)
+	VA_END()
+BEGIN(dd_dddddddd)
+	ARGD2()
+	GETD2(dd_dddddddd)
+	VA_START()
+	VARGD8(dd_dddddddd)
+	VA_END()
+BEGIN(ddd_ddddddd)
+	ARGD3()
+	GETD3(ddd_ddddddd)
+	VA_START()
+	VARGD7(ddd_ddddddd)
+	VA_END()
+BEGIN(dddd_dddddd)
+	ARGD4()
+	GETD4(dddd_dddddd)
+	VA_START()
+	VARGD6(dddd_dddddd)
+	VA_END()
+BEGIN(ddddd_ddddd)
+	ARGD5()
+	GETD5(ddddd_ddddd)
+	VA_START()
+	VARGD5(ddddd_ddddd)
+	VA_END()
+BEGIN(dddddd_dddd)
+	ARGD6()
+	GETD6(dddddd_dddd)
+	VA_START()
+	VARGD4(dddddd_dddd)
+	VA_END()
+BEGIN(ddddddd_ddd)
+	ARGD7()
+	GETD7(ddddddd_ddd)
+	VA_START()
+	VARGD3(ddddddd_ddd)
+	VA_END()
+BEGIN(dddddddd_dd)
+	ARGD8()
+	GETD8(dddddddd_dd)
+	VA_START()
+	VARGD2(dddddddd_dd)
+	VA_END()
+BEGIN(ddddddddd_d)
+	ARGD9()
+	GETD9(ddddddddd_d)
+	VA_START()
+	VARGD1(ddddddddd_d)
+	VA_END()
+BEGIN(_ididididid)
+	VA_START()
+	IDVARG10(_ididididid)
+	VA_END()
+BEGIN(i_didididid)
+	IDARG1()
+	IDGET1(i_didididid)
+	VA_START()
+	IDVARG9(i_didididid)
+	VA_END()
+BEGIN(id_idididid)
+	IDARG2()
+	IDGET2(id_idididid)
+	VA_START()
+	IDVARG8(id_idididid)
+	VA_END()
+BEGIN(idi_dididid)
+	IDARG3()
+	IDGET3(idi_dididid)
+	VA_START()
+	IDVARG7(idi_dididid)
+	VA_END()
+BEGIN(idid_ididid)
+	IDARG4()
+	IDGET4(idid_ididid)
+	VA_START()
+	IDVARG6(idid_ididid)
+	VA_END()
+BEGIN(ididi_didid)
+	IDARG5()
+	IDGET5(ididi_didid)
+	VA_START()
+	IDVARG5(ididi_didid)
+	VA_END()
+BEGIN(ididid_idid)
+	IDARG6()
+	IDGET6(ididid_idid)
+	VA_START()
+	IDVARG4(ididid_idid)
+	VA_END()
+BEGIN(idididi_did)
+	IDARG7()
+	IDGET7(idididi_did)
+	VA_START()
+	IDVARG3(idididi_did)
+	VA_END()
+BEGIN(idididid_id)
+	IDARG8()
+	IDGET8(idididid_id)
+	VA_START()
+	IDVARG2(idididid_id)
+	VA_END()
+BEGIN(ididididi_d)
+	IDARG9()
+	IDGET9(ididididi_d)
+	VA_START()
+	IDVARG1(ididididi_d)
+	VA_END()
+BEGIN(_dididididi)
+	VA_START()
+	DIVARG10(_dididididi)
+	VA_END()
+BEGIN(d_ididididi)
+	DIARG1()
+	DIGET1(d_ididididi)
+	VA_START()
+	DIVARG9(d_ididididi)
+	VA_END()
+BEGIN(di_didididi)
+	DIARG2()
+	DIGET2(di_didididi)
+	VA_START()
+	DIVARG8(di_didididi)
+	VA_END()
+BEGIN(did_idididi)
+	DIARG3()
+	DIGET3(did_idididi)
+	VA_START()
+	DIVARG7(did_idididi)
+	VA_END()
+BEGIN(didi_dididi)
+	DIARG4()
+	DIGET4(didi_dididi)
+	VA_START()
+	DIVARG6(didi_dididi)
+	VA_END()
+BEGIN(didid_ididi)
+	DIARG5()
+	DIGET5(didid_ididi)
+	VA_START()
+	DIVARG5(didid_ididi)
+	VA_END()
+BEGIN(dididi_didi)
+	DIARG6()
+	DIGET6(dididi_didi)
+	VA_START()
+	DIVARG4(dididi_didi)
+	VA_END()
+BEGIN(dididid_idi)
+	DIARG7()
+	DIGET7(dididid_idi)
+	VA_START()
+	DIVARG3(dididid_idi)
+	VA_END()
+BEGIN(didididi_di)
+	DIARG8()
+	DIGET8(didididi_di)
+	VA_START()
+	DIVARG2(didididi_di)
+	VA_END()
+BEGIN(didididid_i)
+	DIARG9()
+	DIGET9(didididid_i)
+	VA_START()
+	DIVARG1(didididid_i)
+	VA_END()
+
+#define PUSH1()		pushargi 1
+#define PUSH2()		PUSH1()		pushargi 2
+#define PUSH3()		PUSH2()		pushargi 3
+#define PUSH4()		PUSH3()		pushargi 4
+#define PUSH5()		PUSH4()		pushargi 5
+#define PUSH6()		PUSH5()		pushargi 6
+#define PUSH7()		PUSH6()		pushargi 7
+#define PUSH8()		PUSH7()		pushargi 8
+#define PUSH9()		PUSH8()		pushargi 9
+#define VPUSH1()	pushargi 1	VPUSH2()
+#define VPUSH2()	pushargi 2	VPUSH3()
+#define VPUSH3()	pushargi 3	VPUSH4()
+#define VPUSH4()	pushargi 4	VPUSH5()
+#define VPUSH5()	pushargi 5	VPUSH6()
+#define VPUSH6()	pushargi 6	VPUSH7()
+#define VPUSH7()	pushargi 7	VPUSH8()
+#define VPUSH8()	pushargi 8	VPUSH9()
+#define VPUSH9()	pushargi 9	VPUSH10()
+#define VPUSH10()	pushargi 10
+#define PUSHD1()	pushargi_d 1
+#define PUSHD2()	PUSHD1()	pushargi_d 2
+#define PUSHD3()	PUSHD2()	pushargi_d 3
+#define PUSHD4()	PUSHD3()	pushargi_d 4
+#define PUSHD5()	PUSHD4()	pushargi_d 5
+#define PUSHD6()	PUSHD5()	pushargi_d 6
+#define PUSHD7()	PUSHD6()	pushargi_d 7
+#define PUSHD8()	PUSHD7()	pushargi_d 8
+#define PUSHD9()	PUSHD8()	pushargi_d 9
+#define VPUSHD1()	pushargi_d 1	VPUSHD2()
+#define VPUSHD2()	pushargi_d 2	VPUSHD3()
+#define VPUSHD3()	pushargi_d 3	VPUSHD4()
+#define VPUSHD4()	pushargi_d 4	VPUSHD5()
+#define VPUSHD5()	pushargi_d 5	VPUSHD6()
+#define VPUSHD6()	pushargi_d 6	VPUSHD7()
+#define VPUSHD7()	pushargi_d 7	VPUSHD8()
+#define VPUSHD8()	pushargi_d 8	VPUSHD9()
+#define VPUSHD9()	pushargi_d 9	VPUSHD10()
+#define VPUSHD10()	pushargi_d 10
+#define IDPUSH1()	pushargi 1
+#define IDPUSH2()	IDPUSH1()	pushargi_d 2
+#define IDPUSH3()	IDPUSH2()	pushargi 3
+#define IDPUSH4()	IDPUSH3()	pushargi_d 4
+#define IDPUSH5()	IDPUSH4()	pushargi 5
+#define IDPUSH6()	IDPUSH5()	pushargi_d 6
+#define IDPUSH7()	IDPUSH6()	pushargi 7
+#define IDPUSH8()	IDPUSH7()	pushargi_d 8
+#define IDPUSH9()	IDPUSH8()	pushargi 9
+#define IDVPUSH1()	pushargi 1	IDVPUSH2()
+#define IDVPUSH2()	pushargi_d 2	IDVPUSH3()
+#define IDVPUSH3()	pushargi 3	IDVPUSH4()
+#define IDVPUSH4()	pushargi_d 4	IDVPUSH5()
+#define IDVPUSH5()	pushargi 5	IDVPUSH6()
+#define IDVPUSH6()	pushargi_d 6	IDVPUSH7()
+#define IDVPUSH7()	pushargi 7	IDVPUSH8()
+#define IDVPUSH8()	pushargi_d 8	IDVPUSH9()
+#define IDVPUSH9()	pushargi 9	IDVPUSH10()
+#define IDVPUSH10()	pushargi_d 10
+#define DIPUSH1()	pushargi_d 1
+#define DIPUSH2()	DIPUSH1()	pushargi 2
+#define DIPUSH3()	DIPUSH2()	pushargi_d 3
+#define DIPUSH4()	DIPUSH3()	pushargi 4
+#define DIPUSH5()	DIPUSH4()	pushargi_d 5
+#define DIPUSH6()	DIPUSH5()	pushargi 6
+#define DIPUSH7()	DIPUSH6()	pushargi_d 7
+#define DIPUSH8()	DIPUSH7()	pushargi 8
+#define DIPUSH9()	DIPUSH8()	pushargi_d 9
+#define DIVPUSH1()	pushargi_d 1	DIVPUSH2()
+#define DIVPUSH2()	pushargi 2	DIVPUSH3()
+#define DIVPUSH3()	pushargi_d 3	DIVPUSH4()
+#define DIVPUSH4()	pushargi 4	DIVPUSH5()
+#define DIVPUSH5()	pushargi_d 5	DIVPUSH6()
+#define DIVPUSH6()	pushargi 6	DIVPUSH7()
+#define DIVPUSH7()	pushargi_d 7	DIVPUSH8()
+#define DIVPUSH8()	pushargi 8	DIVPUSH9()
+#define DIVPUSH9()	pushargi_d 9	DIVPUSH10()
+#define DIVPUSH10()	pushargi 10
 
 main:
 	prolog
 	prepare
 		ellipsis
-		pushargi 1
+		VPUSH1()
-		pushargi_d 2
+	finishi _iiiiiiiiii
-	finishi varargs
+	prepare
+		PUSH1()
+		ellipsis
+		VPUSH2()
+	finishi i_iiiiiiiii
+	prepare
+		PUSH2()
+		ellipsis
+		VPUSH3()
+	finishi ii_iiiiiiii
+	prepare
+		PUSH3()
+		ellipsis
+		VPUSH4()
+	finishi iii_iiiiiii
+	prepare
+		PUSH4()
+		ellipsis
+		VPUSH5()
+	finishi iiii_iiiiii
+	prepare
+		PUSH5()
+		ellipsis
+		VPUSH6()
+	finishi iiiii_iiiii
+	prepare
+		PUSH6()
+		ellipsis
+		VPUSH7()
+	finishi iiiiii_iiii
+	prepare
+		PUSH7()
+		ellipsis
+		VPUSH8()
+	finishi iiiiiii_iii
+	prepare
+		PUSH8()
+		ellipsis
+		VPUSH9()
+	finishi iiiiiiii_ii
+	prepare
+		PUSH9()
+		ellipsis
+		VPUSH10()
+	finishi iiiiiiiii_i
+	prepare
+		ellipsis
+		VPUSHD1()
+	finishi _dddddddddd
+	prepare
+		PUSHD1()
+		ellipsis
+		VPUSHD2()
+	finishi d_ddddddddd
+	prepare
+		PUSHD2()
+		ellipsis
+		VPUSHD3()
+	finishi dd_dddddddd
+	prepare
+		PUSHD3()
+		ellipsis
+		VPUSHD4()
+	finishi ddd_ddddddd
+	prepare
+		PUSHD4()
+		ellipsis
+		VPUSHD5()
+	finishi dddd_dddddd
+	prepare
+		PUSHD5()
+		ellipsis
+		VPUSHD6()
+	finishi ddddd_ddddd
+	prepare
+		PUSHD6()
+		ellipsis
+		VPUSHD7()
+	finishi dddddd_dddd
+	prepare
+		PUSHD7()
+		ellipsis
+		VPUSHD8()
+	finishi ddddddd_ddd
+	prepare
+		PUSHD8()
+		ellipsis
+		VPUSHD9()
+	finishi dddddddd_dd
+	prepare
+		PUSHD9()
+		ellipsis
+		VPUSHD10()
+	finishi ddddddddd_d
+	prepare
+		ellipsis
+		IDVPUSH1()
+	finishi _ididididid
+	prepare
+		IDPUSH1()
+		ellipsis
+		IDVPUSH2()
+	finishi i_didididid
+	prepare
+		IDPUSH2()
+		ellipsis
+		IDVPUSH3()
+	finishi id_idididid
+	prepare
+		IDPUSH3()
+		ellipsis
+		IDVPUSH4()
+	finishi idi_dididid
+	prepare
+		IDPUSH4()
+		ellipsis
+		IDVPUSH5()
+	finishi idid_ididid
+	prepare
+		IDPUSH5()
+		ellipsis
+		IDVPUSH6()
+	finishi ididi_didid
+	prepare
+		IDPUSH6()
+		ellipsis
+		IDVPUSH7()
+	finishi ididid_idid
+	prepare
+		IDPUSH7()
+		ellipsis
+		IDVPUSH8()
+	finishi idididi_did
+	prepare
+		IDPUSH8()
+		ellipsis
+		IDVPUSH9()
+	finishi idididid_id
+	prepare
+		IDPUSH9()
+		ellipsis
+		IDVPUSH10()
+	finishi ididididi_d
+	prepare
+		ellipsis
+		DIVPUSH1()
+	finishi _dididididi
+	prepare
+		DIPUSH1()
+		ellipsis
+		DIVPUSH2()
+	finishi d_ididididi
+	prepare
+		DIPUSH2()
+		ellipsis
+		DIVPUSH3()
+	finishi di_didididi
+	prepare
+		DIPUSH3()
+		ellipsis
+		DIVPUSH4()
+	finishi did_idididi
+	prepare
+		DIPUSH4()
+		ellipsis
+		DIVPUSH5()
+	finishi didi_dididi
+	prepare
+		DIPUSH5()
+		ellipsis
+		DIVPUSH6()
+	finishi didid_ididi
+	prepare
+		DIPUSH6()
+		ellipsis
+		DIVPUSH7()
+	finishi dididi_didi
+	prepare
+		DIPUSH7()
+		ellipsis
+		DIVPUSH8()
+	finishi dididid_idi
+	prepare
+		DIPUSH8()
+		ellipsis
+		DIVPUSH9()
+	finishi didididi_di
+	prepare
+		DIPUSH9()
+		ellipsis
+		DIVPUSH10()
+	finishi didididid_i
+	prepare
+		pushargi ok
+		ellipsis
+	finishi @printf
 	ret
 	epilog

lib/jit_arm-cpu.c

@@ -3778,42 +3778,25 @@ _prolog(jit_state_t *_jit, jit_node_t *node)
 	BX(_R12_REGNO);
 	if (!_jitc->thumb)
 	    _jitc->thumb = _jit->pc.w;
-
-	/* If the jit function is varargs, do a slightly more
-	 * costly prolog to save first the 4 argument registers. */
 	if (jit_cpu.abi) {
-	    if (_jitc->function->self.call & jit_call_varargs)
+	    T2_PUSH(0xf);
-		T2_PUSH(0xf);
 	    T2_PUSH(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
 	    VPUSH_F64(_D8_REGNO, 8);
-	    if (!(_jitc->function->self.call & jit_call_varargs))
-		T2_PUSH(0xf);
 	}
 	else {
-	    if (_jitc->function->self.call & jit_call_varargs) {
+	    T2_PUSH(0xf);
-		T2_PUSH(0xf);
+	    T2_PUSH(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
-		T2_PUSH(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
-	    }
-	    else
-		T2_PUSH(0x3ff|(1<<_FP_REGNO)|(1<<_LR_REGNO));
 	}
     }
     else {
 	if (jit_cpu.abi) {
-	    if (_jitc->function->self.call & jit_call_varargs)
+	    PUSH(0xf);
-		PUSH(0xf);
 	    PUSH(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
 	    VPUSH_F64(_D8_REGNO, 8);
-	    if (!(_jitc->function->self.call & jit_call_varargs))
-		PUSH(0xf);
 	}
 	else {
-	    if (_jitc->function->self.call & jit_call_varargs) {
+	    PUSH(0xf);
-		PUSH(0xf);
+	    PUSH(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
-		PUSH(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
-	    }
-	    else
-		PUSH(0x3ff|(1<<_FP_REGNO)|(1<<_LR_REGNO));
 	}
     }
     movr(_FP_REGNO, _SP_REGNO);
@@ -3833,35 +3816,18 @@ _epilog(jit_state_t *_jit, jit_node_t *node)
     if (_jitc->function->assume_frame)
 	return;
 
-    /* If the jit function is varargs, need a different
+    movr(_SP_REGNO, _FP_REGNO);
-     * epilog, that also does not directly restore the
-     * pc, but instead branch to the lr, after correcting
-     * the stack. */
-    if (_jitc->function->self.call & jit_call_varargs)
-	movr(_SP_REGNO, _FP_REGNO);
-    else
-	addi(_SP_REGNO, _FP_REGNO, 16);
     if (jit_cpu.abi)
 	VPOP_F64(_D8_REGNO, 8);
-    if (jit_thumb_p()) {
+    if (jit_thumb_p())
-	if (!(_jitc->function->self.call & jit_call_varargs))
+	T2_POP(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
-	    T2_POP(0x3f0|(1<<_FP_REGNO)|(1<<_PC_REGNO));
+    else
-	else
+	POP(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
-	    T2_POP(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
+    addi(_SP_REGNO, _SP_REGNO, 16);
-    }
+    if (jit_thumb_p())
-    else {
+	T1_BX(_LR_REGNO);
-	if (!(_jitc->function->self.call & jit_call_varargs))
+    else
-	    POP(0x3f0|(1<<_FP_REGNO)|(1<<_PC_REGNO));
+	BX(_LR_REGNO);
-	else
-	    POP(0x3f0|(1<<_FP_REGNO)|(1<<_LR_REGNO));
-    }
-    if (_jitc->function->self.call & jit_call_varargs) {
-	addi(_SP_REGNO, _SP_REGNO, 16);
-	if (jit_thumb_p())
-	    T1_BX(_LR_REGNO);
-	else
-	    BX(_LR_REGNO);
-    }
     if (jit_thumb_p() && (_jit->pc.w & 2))
 	T1_NOP();
 }
@@ -3869,45 +3835,26 @@ _epilog(jit_state_t *_jit, jit_node_t *node)
 static void
 _vastart(jit_state_t *_jit, jit_int32_t r0)
 {
-    jit_int32_t		reg;
-
     assert(_jitc->function->self.call & jit_call_varargs);
 
-    /* Return jit_va_list_t in the register argument */
-    addi(r0, _FP_REGNO, _jitc->function->vaoff);
-    reg = jit_get_reg(jit_class_gpr);
-
     /* Initialize stack pointer to the first stack argument.
      * The -16 is to account for the 4 argument registers
      * always saved, and _jitc->function->vagp is to account
      * for declared arguments. */
-    addi(rn(reg), _FP_REGNO, _jitc->function->self.size -
+    addi(r0, _FP_REGNO, _jitc->function->self.size -
 	 16 + _jitc->function->vagp);
-    stxi(offsetof(jit_va_list_t, stack), r0, rn(reg));
-
-    jit_unget_reg(reg);
 }
 
 static void
 _vaarg(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
-    jit_int32_t		reg;
-
     assert(_jitc->function->self.call & jit_call_varargs);
 
-    reg = jit_get_reg(jit_class_gpr);
-
-    /* Load stack pointer. */
-    ldxi(rn(reg), r1, offsetof(jit_va_list_t, stack));
-
     /* Load argument. */
-    ldr(r0, rn(reg));
+    ldr(r0, r1);
 
     /* Update stack pointer. */
-    addi(rn(reg), rn(reg), sizeof(jit_word_t));
+    addi(r1, r1, sizeof(jit_word_t));
-    stxi(offsetof(jit_va_list_t, stack), r1, rn(reg));
-
-    jit_unget_reg(reg);
 }
 
 static void

lib/jit_arm-swf.c

@@ -2620,27 +2620,21 @@ _swf_stxi_d(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
 static void
 _swf_vaarg_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
-    jit_int32_t		rg0, rg1;
+    jit_int32_t		reg;
 
     assert(_jitc->function->self.call & jit_call_varargs);
 
-    rg0 = jit_get_reg(jit_class_gpr);
+    /* Adjust pointer. */
-
+    reg = jit_get_reg(jit_class_gpr);
-    /* Load stack pointer. */
+    andi(rn(reg), r1, 7);
-    ldxi(rn(rg0), r1, offsetof(jit_va_list_t, stack));
+    addr(r1, r1, rn(reg));
-    rg1 = jit_get_reg(jit_class_gpr);
+    jit_unget_reg(reg);
-    andi(rn(rg1), rn(rg0), 7);
-    addr(rn(rg0), rn(rg0), rn(rg1));
-    jit_unget_reg(rg1);
 
     /* Load argument. */
-    swf_ldr_d(r0, rn(rg0));
+    swf_ldr_d(r0, r1);
 
     /* Update stack pointer. */
-    addi(rn(rg0), rn(rg0), sizeof(jit_float64_t));
+    addi(r1, r1, sizeof(jit_float64_t));
-    stxi(offsetof(jit_va_list_t, stack), r1, rn(rg0));
-
-    jit_unget_reg(rg0);
 }
 
 #endif

lib/jit_arm-vfp.c

@@ -2307,27 +2307,21 @@ _vfp_stxi_d(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
 static void
 _vfp_vaarg_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
-    jit_int32_t		rg0, rg1;
+    jit_int32_t		reg;
 
     assert(_jitc->function->self.call & jit_call_varargs);
 
-    rg0 = jit_get_reg(jit_class_gpr);
+    /* Adjust pointer. */
-
+    reg = jit_get_reg(jit_class_gpr);
-    /* Load stack pointer. */
+    andi(rn(reg), r1, 7);
-    ldxi(rn(rg0), r1, offsetof(jit_va_list_t, stack));
+    addr(r1, r1, rn(reg));
-    rg1 = jit_get_reg(jit_class_gpr);
+    jit_unget_reg(reg);
-    andi(rn(rg1), rn(rg0), 7);
-    addr(rn(rg0), rn(rg0), rn(rg1));
-    jit_unget_reg(rg1);
 
     /* Load argument. */
-    vfp_ldr_d(r0, rn(rg0));
+    vfp_ldr_d(r0, r1);
 
     /* Update stack pointer. */
-    addi(rn(rg0), rn(rg0), sizeof(jit_float64_t));
+    addi(r1, r1, sizeof(jit_float64_t));
-    stxi(offsetof(jit_va_list_t, stack), r1, rn(rg0));
-
-    jit_unget_reg(rg0);
 }
 #  undef dbopi
 #  undef fbopi

lib/jit_arm.c

@@ -34,6 +34,11 @@
 
 #define jit_fpr_p(rn)			((rn) > 15)
 
+#define arg_base()							\
+    (stack_framesize - 16 + (jit_cpu.abi ? 64 : 0))
+#define arg_offset(n)							\
+    ((n) < 4 ? arg_base() + ((n) << 2) : (n))
+
 /* Assume functions called never match jit instruction set, that is
  * libc, gmp, mpfr, etc functions are in thumb mode and jit is in
  * arm mode, what may cause a crash upon return of that function
@@ -51,9 +56,7 @@ typedef union _jit_thumb_t {
     jit_int16_t		s[2];
 } jit_thumb_t;
 
-typedef struct jit_va_list {
+typedef jit_pointer_t	jit_va_list;
-    jit_pointer_t	stack;
-} jit_va_list_t;
 
 /*
  * Prototypes
@@ -420,9 +423,6 @@ _jit_ellipsis(jit_state_t *_jit)
 	assert(!(_jitc->function->self.call & jit_call_varargs));
 	_jitc->function->self.call |= jit_call_varargs;
 
-	/* Allocate va_list like object in the stack. */
-	_jitc->function->vaoff = jit_allocai(sizeof(jit_va_list_t));
-
 	/* First 4 stack addresses are always spilled r0-r3 */
 	if (jit_arg_reg_p(_jitc->function->self.argi))
 	    _jitc->function->vagp = _jitc->function->self.argi * 4;
@@ -504,7 +504,7 @@ _jit_getarg_c(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
     assert(v->code == jit_code_arg);
     if (jit_swf_p())
-	jit_ldxi_c(u, JIT_FP, v->u.w < 4 ? v->u.w << 2 : v->u.w);
+	jit_ldxi_c(u, JIT_FP, arg_offset(v->u.w));
     else if (jit_arg_reg_p(v->u.w))
 	jit_extr_c(u, JIT_RA0 - v->u.w);
     else
@@ -516,7 +516,7 @@ _jit_getarg_uc(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
     assert(v->code == jit_code_arg);
     if (jit_swf_p())
-	jit_ldxi_uc(u, JIT_FP, v->u.w < 4 ? v->u.w << 2 : v->u.w);
+	jit_ldxi_uc(u, JIT_FP, arg_offset(v->u.w));
     else if (jit_arg_reg_p(v->u.w))
 	jit_extr_uc(u, JIT_RA0 - v->u.w);
     else
@@ -528,7 +528,7 @@ _jit_getarg_s(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
     assert(v->code == jit_code_arg);
     if (jit_swf_p())
-	jit_ldxi_s(u, JIT_FP, v->u.w < 4 ? v->u.w << 2 : v->u.w);
+	jit_ldxi_s(u, JIT_FP, arg_offset(v->u.w));
     else if (jit_arg_reg_p(v->u.w))
 	jit_extr_s(u, JIT_RA0 - v->u.w);
     else
@@ -540,7 +540,7 @@ _jit_getarg_us(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
     assert(v->code == jit_code_arg);
     if (jit_swf_p())
-	jit_ldxi_us(u, JIT_FP, v->u.w < 4 ? v->u.w << 2 : v->u.w);
+	jit_ldxi_us(u, JIT_FP, arg_offset(v->u.w));
     else if (jit_arg_reg_p(v->u.w))
 	jit_extr_us(u, JIT_RA0 - v->u.w);
     else
@@ -552,7 +552,7 @@ _jit_getarg_i(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
     assert(v->code == jit_code_arg);
     if (jit_swf_p())
-	jit_ldxi_i(u, JIT_FP, v->u.w < 4 ? v->u.w << 2 : v->u.w);
+	jit_ldxi_i(u, JIT_FP, arg_offset(v->u.w));
     else if (jit_arg_reg_p(v->u.w))
 	jit_movr(u, JIT_RA0 - v->u.w);
     else
@@ -564,7 +564,7 @@ _jit_putargr(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
     assert(v->code == jit_code_arg);
     if (jit_swf_p())
-	jit_stxi(v->u.w < 4 ? v->u.w << 2 : v->u.w, JIT_FP, u);
+	jit_stxi(arg_offset(v->u.w), JIT_FP, u);
     else if (jit_arg_reg_p(v->u.w))
 	jit_movr(JIT_RA0 - v->u.w, u);
     else
@@ -579,7 +579,7 @@ _jit_putargi(jit_state_t *_jit, jit_word_t u, jit_node_t *v)
     if (jit_swf_p()) {
 	regno = jit_get_reg(jit_class_gpr);
 	jit_movi(regno, u);
-	jit_stxi(v->u.w < 4 ? v->u.w << 2 : v->u.w, JIT_FP, regno);
+	jit_stxi(arg_offset(v->u.w), JIT_FP, regno);
 	jit_unget_reg(regno);
     }
     else if (jit_arg_reg_p(v->u.w))
@@ -603,7 +603,7 @@ _jit_getarg_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 	    jit_ldxi_f(u, JIT_FP, v->u.w);
     }
     else if (jit_swf_p())
-	jit_ldxi_f(u, JIT_FP, v->u.w < 4 ? v->u.w << 2 : v->u.w);
+	jit_ldxi_f(u, JIT_FP, arg_offset(v->u.w));
     else {
 	if (jit_arg_reg_p(v->u.w))
 	    jit_movr_w_f(u, JIT_RA0 - v->u.w);
@@ -623,7 +623,7 @@ _jit_putargr_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 	    jit_stxi_f(v->u.w, JIT_FP, u);
     }
     else if (jit_swf_p())
-	jit_stxi_f(v->u.w < 4 ? v->u.w << 2 : v->u.w, JIT_FP, u);
+	jit_stxi_f(arg_offset(v->u.w), JIT_FP, u);
     else {
 	if (jit_arg_reg_p(v->u.w))
 	    jit_movr_f_w(JIT_RA0 - v->u.w, u);
@@ -650,7 +650,7 @@ _jit_putargi_f(jit_state_t *_jit, jit_float32_t u, jit_node_t *v)
     else if (jit_swf_p()) {
 	regno = jit_get_reg(jit_class_fpr);
 	jit_movi_f(regno, u);
-	jit_stxi_f(v->u.w < 4 ? v->u.w << 2 : v->u.w, JIT_FP, regno);
+	jit_stxi_f(arg_offset(v->u.w), JIT_FP, regno);
 	jit_unget_reg(regno);
     }
     else {
@@ -675,7 +675,7 @@ _jit_getarg_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 	    jit_ldxi_d(u, JIT_FP, v->u.w);
     }
     else if (jit_swf_p())
-	jit_ldxi_d(u, JIT_FP, v->u.w < 4 ? v->u.w << 2 : v->u.w);
+	jit_ldxi_d(u, JIT_FP, arg_offset(v->u.w));
     else {
 	if (jit_arg_reg_p(v->u.w))
 	    jit_movr_ww_d(u, JIT_RA0 - v->u.w, JIT_RA0 - (v->u.w + 1));
@@ -695,7 +695,7 @@ _jit_putargr_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 	    jit_stxi_d(v->u.w, JIT_FP, u);
     }
     else if (jit_swf_p())
-	jit_stxi_d(v->u.w < 4 ? v->u.w << 2 : v->u.w, JIT_FP, u);
+	jit_stxi_d(arg_offset(v->u.w), JIT_FP, u);
     else {
 	if (jit_arg_reg_p(v->u.w))
 	    jit_movr_d_ww(JIT_RA0 - v->u.w, JIT_RA0 - (v->u.w + 1), u);
@@ -722,7 +722,7 @@ _jit_putargi_d(jit_state_t *_jit, jit_float64_t u, jit_node_t *v)
     else if (jit_swf_p()) {
 	regno = jit_get_reg(jit_class_fpr);
 	jit_movi_d(regno, u);
-	jit_stxi_d(v->u.w < 4 ? v->u.w << 2 : v->u.w, JIT_FP, regno);
+	jit_stxi_d(arg_offset(v->u.w), JIT_FP, regno);
 	jit_unget_reg(regno);
     }
     else {

lib/jit_x86-cpu.c

@@ -3654,6 +3654,10 @@ _epilog(jit_state_t *_jit, jit_node_t *node)
 static void
 _vastart(jit_state_t *_jit, jit_int32_t r0)
 {
+#if __X32 || __CYGWIN__
+    assert(_jitc->function->self.call & jit_call_varargs);
+    addi(r0, _RBP_REGNO, _jitc->function->self.size);
+#else
     jit_int32_t		reg;
 
     assert(_jitc->function->self.call & jit_call_varargs);
@@ -3662,7 +3666,6 @@ _vastart(jit_state_t *_jit, jit_int32_t r0)
     addi(r0, _RBP_REGNO, _jitc->function->vaoff);
     reg = jit_get_reg(jit_class_gpr);
 
-#if __X64 && !__CYGWIN__
     /* Initialize gp offset in the save area. */
     movi(rn(reg), _jitc->function->vagp);
     stxi_i(offsetof(jit_va_list_t, gpoff), r0, rn(reg));
@@ -3670,35 +3673,35 @@ _vastart(jit_state_t *_jit, jit_int32_t r0)
     /* Initialize fp offset in the save area. */
     movi(rn(reg), _jitc->function->vafp);
     stxi_i(offsetof(jit_va_list_t, fpoff), r0, rn(reg));
-#endif
 
     /* Initialize overflow pointer to the first stack argument. */
     addi(rn(reg), _RBP_REGNO, _jitc->function->self.size);
     stxi(offsetof(jit_va_list_t, over), r0, rn(reg));
 
-#if __X64 && !__CYGWIN__
     /* Initialize register save area pointer. */
     addi(rn(reg), r0, first_gp_offset);
     stxi(offsetof(jit_va_list_t, save), r0, rn(reg));
-#endif
 
     jit_unget_reg(reg);
+#endif
 }
 
 static void
 _vaarg(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
+#if __X32 || __CYGWIN__
+    assert(_jitc->function->self.call & jit_call_varargs);
+    ldr(r0, r1);
+    addi(r1, r1, va_gp_increment);
+#else
     jit_int32_t		rg0;
-#if __X64 && !__CYGWIN__
     jit_int32_t		rg1;
     jit_word_t		ge_code;
     jit_word_t		lt_code;
-#endif
 
     assert(_jitc->function->self.call & jit_call_varargs);
 
     rg0 = jit_get_reg(jit_class_gpr);
-#if __X64 && !__CYGWIN__
     rg1 = jit_get_reg(jit_class_gpr);
 
     /* Load the gp offset in save area in the first temporary. */
@@ -3728,7 +3731,6 @@ _vaarg(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 
     /* Where to land if argument is in overflow area. */
     patch_rel_char(ge_code, _jit->pc.w);
-#endif
 
     /* Load overflow pointer. */
     ldxi(rn(rg0), r1, offsetof(jit_va_list_t, over));
@@ -3740,12 +3742,11 @@ _vaarg(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
     addi(rn(rg0), rn(rg0), va_gp_increment);
     stxi(offsetof(jit_va_list_t, over), r1, rn(rg0));
 
-#if __X64 && !__CYGWIN__
     /* Where to land if argument is in save area. */
     patch_rel_char(lt_code, _jit->pc.w);
-#endif
 
     jit_unget_reg(rg0);
+#endif
 }
 
 /* The x87 boolean argument tells if will put the result in a x87
@@ -3753,17 +3754,22 @@ _vaarg(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 static void
 _vaarg_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_bool_t x87)
 {
+#if __X32 || __CYGWIN__
+    assert(_jitc->function->self.call & jit_call_varargs);
+    if (x87)
+	x87_ldr_d(r0, r1);
+    else
+	sse_ldr_d(r0, r1);
+    addi(r1, r1, 8);
+#else
     jit_int32_t		rg0;
-#if __X64 && !__CYGWIN__
     jit_int32_t		rg1;
     jit_word_t		ge_code;
     jit_word_t		lt_code;
-#endif
 
     assert(_jitc->function->self.call & jit_call_varargs);
 
     rg0 = jit_get_reg(jit_class_gpr);
-#if __X64 && !__CYGWIN__
     rg1 = jit_get_reg(jit_class_gpr);
 
     /* Load the fp offset in save area in the first temporary. */
@@ -3796,7 +3802,6 @@ _vaarg_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_bool_t x87)
 
     /* Where to land if argument is in overflow area. */
     patch_rel_char(ge_code, _jit->pc.w);
-#endif
 
     /* Load overflow pointer. */
     ldxi(rn(rg0), r1, offsetof(jit_va_list_t, over));
@@ -3811,12 +3816,11 @@ _vaarg_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_bool_t x87)
     addi(rn(rg0), rn(rg0), 8);
     stxi(offsetof(jit_va_list_t, over), r1, rn(rg0));
 
-#if __X64 && !__CYGWIN__
     /* Where to land if argument is in save area. */
     patch_rel_char(lt_code, _jit->pc.w);
-#endif
 
     jit_unget_reg(rg0);
+#endif
 }
 
 static void

lib/jit_x86-sz.c

@@ -9,7 +9,7 @@
     0,	/* #name */
     0,	/* #note */
     3,	/* label */
-    26,	/* prolog */
+    34,	/* prolog */
     0,	/* arg */
     3,	/* addr */
     6,	/* addi */
@@ -23,7 +23,7 @@
     6,	/* subci */
     6,	/* subxr */
     5,	/* subxi */
-    10,	/* rsbi */
+    8,	/* rsbi */
     5,	/* mulr */
     7,	/* muli */
     20,	/* qmulr */
@@ -84,7 +84,7 @@
     3,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    4,	/* htonr_us */
+    7,	/* htonr_us */
     4,	/* htonr_ui */
     0,	/* htonr_ul */
     3,	/* ldr_c */
@@ -102,13 +102,13 @@
     0,	/* ldr_l */
     0,	/* ldi_l */
     4,	/* ldxr_c */
-    4,	/* ldxi_c */
+    7,	/* ldxi_c */
     4,	/* ldxr_uc */
-    4,	/* ldxi_uc */
+    7,	/* ldxi_uc */
     4,	/* ldxr_s */
-    4,	/* ldxi_s */
+    7,	/* ldxi_s */
     4,	/* ldxr_us */
-    4,	/* ldxi_us */
+    7,	/* ldxi_us */
     3,	/* ldxr_i */
     6,	/* ldxi_i */
     0,	/* ldxr_ui */
@@ -126,7 +126,7 @@
     11,	/* stxr_c */
     11,	/* stxi_c */
     4,	/* stxr_s */
-    4,	/* stxi_s */
+    7,	/* stxi_s */
     3,	/* stxr_i */
     6,	/* stxi_i */
     0,	/* stxr_l */
@@ -173,15 +173,15 @@
     9,	/* bxsubi_u */
     2,	/* jmpr */
     5,	/* jmpi */
-    4,	/* callr */
+    2,	/* callr */
-    7,	/* calli */
+    5,	/* calli */
     24,	/* epilog */
     0,	/* arg_f */
     8,	/* addr_f */
     19,	/* addi_f */
     12,	/* subr_f */
     19,	/* subi_f */
-    19,	/* rsbi_f */
+    21,	/* rsbi_f */
     8,	/* mulr_f */
     19,	/* muli_f */
     12,	/* divr_f */
@@ -226,11 +226,11 @@
     4,	/* ldr_f */
     8,	/* ldi_f */
     5,	/* ldxr_f */
-    5,	/* ldxi_f */
+    8,	/* ldxi_f */
     6,	/* str_f */
     10,	/* sti_f */
     7,	/* stxr_f */
-    7,	/* stxi_f */
+    8,	/* stxi_f */
     10,	/* bltr_f */
     23,	/* blti_f */
     10,	/* bler_f */
@@ -264,7 +264,7 @@
     26,	/* addi_d */
     12,	/* subr_d */
     26,	/* subi_d */
-    26,	/* rsbi_d */
+    30,	/* rsbi_d */
     8,	/* mulr_d */
     26,	/* muli_d */
     12,	/* divr_d */
@@ -313,7 +313,7 @@
     6,	/* str_d */
     10,	/* sti_d */
     7,	/* stxr_d */
-    7,	/* stxi_d */
+    8,	/* stxi_d */
     10,	/* bltr_d */
     28,	/* blti_d */
     10,	/* bler_d */
@@ -353,9 +353,9 @@
     0,	/* movi_d_w */
     10,	/* x86_retval_f */
     10,	/* x86_retval_d */
-    0,	/* va_start */
+    3,	/* va_start */
-    0,	/* va_arg */
+    5,	/* va_arg */
-    0,	/* va_arg_d */
+    7,	/* va_arg_d */
     0,	/* va_end */
 #endif /* __X32 */
 
@@ -721,7 +721,7 @@
 #else
 
 #  if __X64_32
-#define JIT_INSTR_MAX 44
+#define JIT_INSTR_MAX 108
     0,	/* data */
     0,	/* live */
     3,	/* align */
@@ -730,7 +730,7 @@
     0,	/* #name */
     0,	/* #note */
     3,	/* label */
-    39,	/* prolog */
+    108,	/* prolog */
     0,	/* arg */
     5,	/* addr */
     7,	/* addi */
@@ -822,15 +822,15 @@
     0,	/* ldi_ui */
     0,	/* ldr_l */
     0,	/* ldi_l */
-    6,	/* ldxr_c */
+    9,	/* ldxr_c */
     7,	/* ldxi_c */
-    6,	/* ldxr_uc */
+    9,	/* ldxr_uc */
     7,	/* ldxi_uc */
-    6,	/* ldxr_s */
+    9,	/* ldxr_s */
     7,	/* ldxi_s */
-    6,	/* ldxr_us */
+    9,	/* ldxr_us */
     7,	/* ldxi_us */
-    5,	/* ldxr_i */
+    8,	/* ldxr_i */
     7,	/* ldxi_i */
     0,	/* ldxr_ui */
     0,	/* ldxi_ui */
@@ -844,11 +844,11 @@
     8,	/* sti_i */
     0,	/* str_l */
     0,	/* sti_l */
-    8,	/* stxr_c */
+    12,	/* stxr_c */
     7,	/* stxi_c */
-    6,	/* stxr_s */
+    10,	/* stxr_s */
     7,	/* stxi_s */
-    5,	/* stxr_i */
+    9,	/* stxr_i */
     6,	/* stxi_i */
     0,	/* stxr_l */
     0,	/* stxi_l */
@@ -896,7 +896,7 @@
     5,	/* jmpi */
     3,	/* callr */
     9,	/* calli */
-    34,	/* epilog */
+    35,	/* epilog */
     0,	/* arg_f */
     10,	/* addr_f */
     21,	/* addi_f */
@@ -946,11 +946,11 @@
     11,	/* movi_f */
     6,	/* ldr_f */
     10,	/* ldi_f */
-    7,	/* ldxr_f */
+    11,	/* ldxr_f */
     9,	/* ldxi_f */
     6,	/* str_f */
     10,	/* sti_f */
-    7,	/* stxr_f */
+    11,	/* stxr_f */
     9,	/* stxi_f */
     10,	/* bltr_f */
     21,	/* blti_f */
@@ -1029,11 +1029,11 @@
     23,	/* movi_d */
     6,	/* ldr_d */
     10,	/* ldi_d */
-    7,	/* ldxr_d */
+    11,	/* ldxr_d */
     9,	/* ldxi_d */
     6,	/* str_d */
     10,	/* sti_d */
-    7,	/* stxr_d */
+    11,	/* stxr_d */
     9,	/* stxi_d */
     11,	/* bltr_d */
     34,	/* blti_d */
@@ -1074,22 +1074,22 @@
     0,	/* movi_d_w */
     0,	/* x86_retval_f */
     0,	/* x86_retval_d */
-    0,	/* va_start */
+    41,	/* va_start */
-    0,	/* va_arg */
+    45,	/* va_arg */
-    0,	/* va_arg_d */
+    54,	/* va_arg_d */
     0,	/* va_end */
 #  else
 
-#define JIT_INSTR_MAX 43
+#define JIT_INSTR_MAX 115
     0,	/* data */
     0,	/* live */
-    7,	/* align */
+    6,	/* align */
     0,	/* save */
     0,	/* load */
     0,	/* #name */
     0,	/* #note */
     7,	/* label */
-    43,	/* prolog */
+    115,	/* prolog */
     0,	/* arg */
     5,	/* addr */
     13,	/* addi */
@@ -1103,7 +1103,7 @@
     13,	/* subci */
     9,	/* subxr */
     7,	/* subxi */
-    19,	/* rsbi */
+    16,	/* rsbi */
     7,	/* mulr */
     14,	/* muli */
     20,	/* qmulr */
@@ -1182,17 +1182,17 @@
     4,	/* ldr_l */
     8,	/* ldi_l */
     6,	/* ldxr_c */
-    5,	/* ldxi_c */
+    8,	/* ldxi_c */
     6,	/* ldxr_uc */
-    5,	/* ldxi_uc */
+    8,	/* ldxi_uc */
     6,	/* ldxr_s */
-    5,	/* ldxi_s */
+    8,	/* ldxi_s */
     6,	/* ldxr_us */
-    5,	/* ldxi_us */
+    8,	/* ldxi_us */
     5,	/* ldxr_i */
     7,	/* ldxi_i */
     5,	/* ldxr_ui */
-    4,	/* ldxi_ui */
+    6,	/* ldxi_ui */
     5,	/* ldxr_l */
     7,	/* ldxi_l */
     4,	/* str_c */
@@ -1204,9 +1204,9 @@
     4,	/* str_l */
     8,	/* sti_l */
     5,	/* stxr_c */
-    4,	/* stxi_c */
+    6,	/* stxi_c */
     6,	/* stxr_s */
-    5,	/* stxi_s */
+    7,	/* stxi_s */
     5,	/* stxr_i */
     6,	/* stxi_i */
     5,	/* stxr_l */
@@ -1251,22 +1251,22 @@
     10,	/* bxsubi */
     9,	/* bxsubr_u */
     10,	/* bxsubi_u */
-    0,	/* jmpr */
+    3,	/* jmpr */
     5,	/* jmpi */
     3,	/* callr */
     13,	/* calli */
-    37,	/* epilog */
+    38,	/* epilog */
     0,	/* arg_f */
     10,	/* addr_f */
-    20,	/* addi_f */
+    21,	/* addi_f */
     15,	/* subr_f */
-    20,	/* subi_f */
+    21,	/* subi_f */
-    20,	/* rsbi_f */
+    30,	/* rsbi_f */
     10,	/* mulr_f */
-    20,	/* muli_f */
+    21,	/* muli_f */
     15,	/* divr_f */
-    20,	/* divi_f */
+    21,	/* divi_f */
-    14,	/* negr_f */
+    15,	/* negr_f */
     15,	/* absr_f */
     5,	/* sqrtr_f */
     11,	/* ltr_f */
@@ -1306,13 +1306,13 @@
     6,	/* ldr_f */
     10,	/* ldi_f */
     7,	/* ldxr_f */
-    6,	/* ldxi_f */
+    9,	/* ldxi_f */
     6,	/* str_f */
     10,	/* sti_f */
     7,	/* stxr_f */
-    7,	/* stxi_f */
+    9,	/* stxi_f */
     10,	/* bltr_f */
-    20,	/* blti_f */
+    21,	/* blti_f */
     10,	/* bler_f */
     25,	/* blei_f */
     12,	/* beqr_f */
@@ -1344,12 +1344,12 @@
     25,	/* addi_d */
     15,	/* subr_d */
     25,	/* subi_d */
-    25,	/* rsbi_d */
+    30,	/* rsbi_d */
     10,	/* mulr_d */
     25,	/* muli_d */
     15,	/* divr_d */
     25,	/* divi_d */
-    21,	/* negr_d */
+    22,	/* negr_d */
     16,	/* absr_d */
     5,	/* sqrtr_d */
     12,	/* ltr_d */
@@ -1389,11 +1389,11 @@
     6,	/* ldr_d */
     10,	/* ldi_d */
     7,	/* ldxr_d */
-    6,	/* ldxi_d */
+    9,	/* ldxi_d */
     6,	/* str_d */
     10,	/* sti_d */
     7,	/* stxr_d */
-    7,	/* stxi_d */
+    9,	/* stxi_d */
     11,	/* bltr_d */
     26,	/* blti_d */
     11,	/* bler_d */
@@ -1433,9 +1433,9 @@
     0,	/* movi_d_w */
     0,	/* x86_retval_f */
     0,	/* x86_retval_d */
-    0,	/* va_start */
+    42,	/* va_start */
-    0,	/* va_arg */
+    41,	/* va_arg */
-    0,	/* va_arg_d */
+    50,	/* va_arg_d */
     0,	/* va_end */
 #endif /* __CYGWIN__ */
 #  endif /* __X64_32 */

lib/jit_x86.c

@@ -61,13 +61,16 @@
 #  define REAL_WORDSIZE			8
 #endif
 
+/*
+ * Types
+ */
+#if __X32 || __CYGWIN__
+typedef jit_pointer_t jit_va_list_t;
+#else
 typedef struct jit_va_list {
-#if __X64 && !__CYGWIN__
     jit_int32_t		gpoff;
     jit_int32_t		fpoff;
-#endif
     jit_pointer_t	over;
-#if __X64 && !__CYGWIN__
     jit_pointer_t	save;
     /* Declared explicitly as int64 for the x32 abi */
     jit_int64_t		rdi;
@@ -92,8 +95,8 @@ typedef struct jit_va_list {
     jit_float64_t	_up6;
     jit_float64_t	xmm7;
     jit_float64_t	_up7;
-#endif
 } jit_va_list_t;
+#endif
 
 /*
  * Prototypes
@@ -564,12 +567,12 @@ _jit_ellipsis(jit_state_t *_jit)
 	assert(!(_jitc->function->self.call & jit_call_varargs));
 	_jitc->function->self.call |= jit_call_varargs;
 
+#if __X64 && !__CYGWIN__
 	/* Allocate va_list like object in the stack.
 	 * If applicable, with enough space to save all argument
 	 * registers, and use fixed offsets for them. */
 	_jitc->function->vaoff = jit_allocai(sizeof(jit_va_list_t));
 
-#if __X64 && !__CYGWIN__
 	/* Initialize gp offset in save area. */
 	if (jit_arg_reg_p(_jitc->function->self.argi))
 	    _jitc->function->vagp = _jitc->function->self.argi * 8;