Add new varargs test and correct related test case failures in all ports.

ChangeLog

@@ -1,3 +1,24 @@
+2012-12-13 Paulo Andrade <pcpa@gnu.org>
+
+	* check/varargs.ok, check/varargs.tst: New test cases implementing
+	simple varargs calls with a large amount of arguments to exercise
+	excess arguments on stack.
+
+	* include/lightning.h: Include config.h if HAVE_CONFIG_H is
+	defined.
+
+	* lib/jit_arm.c: Allocate a fpr register, not a gpr one for
+	temporary when pushing varargs arguments in the stack.
+
+	* lib/jit_arm-swf.c: Correct code changing the wrong offset
+	in jit_absr_d and jit_negr_d in software float.
+
+	* lib/jit_mips.c: Correct calculation of offsets of arguments
+	on stack.
+
+	* lib/jit_ppc.c: Correct bogus logic for "next" offset of arguments
+	on stack and adjust for fixed offset of stack arguments.
+
 2012-12-12 Paulo Andrade <pcpa@gnu.org>
 
 	* include/lightning.h, lib/jit_arm.c, lib/jit_mips.c,

check/Makefile.am

@@ -57,6 +57,7 @@ EXTRA_DIST =				\
 	alu_rsh.tst	alu_rsh.ok	\
 	alu_com.tst	alu_com.ok	\
 	alu_neg.tst	alu_neg.ok	\
+	varargs.tst	varargs.ok	\
 	check.sh	run-test	\
 	all.tst
 
@@ -71,7 +72,8 @@ TESTS = 3to2 add allocai		\
 	alu_mul alu_div alu_rem		\
 	alu_and alu_or alu_xor		\
 	alu_lsh alu_rsh			\
-	alu_com alu_neg
+	alu_com alu_neg			\
+	varargs
 
 CLEANFILES = $(TESTS)
 

check/varargs.ok

@@ -0,0 +1,3 @@
+0 1 2 3 4 5 6 7 8 9
+0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0
+0 0.0 1 1.0 2 2.0 3 3.0 4 4.0 5 5.0 6 6.0 7 7.0 8 8.0 9 9.0

check/varargs.tst

@@ -0,0 +1,283 @@
+.data	1024
+ifmt:
+.c	"%d %d %d %d %d %d %d %d %d %d\n"
+.align	4
+ichk:
+.i	9 8 7 6 5 4 3 2 1 0
+dfmt:
+.c	"%.1f %.1f %.1f %.1f %.1f %.1f %.1f %.1f %.1f %.1f\n"
+lfmt:
+.c	"%lf %lf %lf %lf %lf %lf %lf %lf %lf %lf\n"
+dchk:
+.align	8
+.d	9.0 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0
+idfmt:
+.c	"%d %.1f %d %.1f %d %.1f %d %.1f %d %.1f %d %.1f %d %.1f %d %.1f %d %.1f %d %.1f\n"
+ldfmt:
+.c	"%d %lf %d %lf %d %lf %d %lf %d %lf %d %lf %d %lf %d %lf %d %lf %d %lf\n"
+buff:
+.align	4
+.size	256
+
+.code
+	jmpi main
+
+main:
+	prolog
+
+	/*
+		sprintf(buff, "%d %d %d %d %d %d %d %d %d %d\n",
+			0, 1, 2, 3, 4, 5, 6, 7, 8, 9);
+	 */
+	prepare
+		pushargi buff
+		pushargi ifmt
+		ellipsis
+		pushargi 0
+		pushargi 1
+		pushargi 2
+		pushargi 3
+		pushargi 4
+		pushargi 5
+		pushargi 6
+		pushargi 7
+		pushargi 8
+		pushargi 9
+	finishi @sprintf
+
+	/*
+		sscanf(buff, "%d %d %d %d %d %d %d %d %d %d\n",
+		       ichk+0, ichk+1, ichk+2, ichk+3, ichk+4,
+		       ichk+5, ichk+6, ichk+7, ichk+8, ichk+9);
+	 */
+	movi %v0 ichk
+	prepare
+		pushargi buff
+		pushargi ifmt
+		ellipsis
+		pushargr %v0		/* 0 */
+		addi %v0 %v0 4
+		pushargr %v0		/* 1 */
+		addi %v0 %v0 4
+		pushargr %v0		/* 2 */
+		addi %v0 %v0 4
+		pushargr %v0		/* 3 */
+		addi %v0 %v0 4
+		pushargr %v0		/* 4 */
+		addi %v0 %v0 4
+		pushargr %v0		/* 5 */
+		addi %v0 %v0 4
+		pushargr %v0		/* 6 */
+		addi %v0 %v0 4
+		pushargr %v0		/* 7 */
+		addi %v0 %v0 4
+		pushargr %v0		/* 8 */
+		addi %v0 %v0 4
+		pushargr %v0		/* 9 */
+	finishi @sscanf
+
+	movi %v0 ichk
+	movi %r0 0
+loopi:
+	ldr_i %r1 %v0
+	beqr nexti %r0 %r1
+	calli @abort
+nexti:
+	addi %r0 %r0 1
+	bgei outi %r0 10
+	addi %v0 %v0 4
+	jmpi loopi
+outi:
+
+	prepare
+		pushargi buff
+		ellipsis
+	finishi @printf
+
+	/*
+		sprintf(buff,
+			"%.1f %.1f %.1f %.1f %.1f "
+			"%.1f %.1f %.1f %.1f %.1f\n",
+			0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0);
+	 */
+	prepare
+		pushargi buff
+		pushargi dfmt
+		ellipsis
+		pushargi_d 0.0
+		pushargi_d 1.0
+		pushargi_d 2.0
+		pushargi_d 3.0
+		pushargi_d 4.0
+		pushargi_d 5.0
+		pushargi_d 6.0
+		pushargi_d 7.0
+		pushargi_d 8.0
+		pushargi_d 9.0
+	finishi @sprintf
+
+	/*
+		sscanf(buff, "%lf %lf %lf %lf %lf %lf %lf %lf %lf %lf\n",
+		       dchk+0, dchk+1, dchk+2, dchk+3, dchk+4,
+		       dchk+5, dchk+6, dchk+7, dchk+8, dchk+9);
+	 */
+	movi %v0 dchk
+	prepare
+		pushargi buff
+		pushargi lfmt
+		ellipsis
+		pushargr %v0		/* 0 */
+		addi %v0 %v0 8
+		pushargr %v0		/* 1 */
+		addi %v0 %v0 8
+		pushargr %v0		/* 2 */
+		addi %v0 %v0 8
+		pushargr %v0		/* 3 */
+		addi %v0 %v0 8
+		pushargr %v0		/* 4 */
+		addi %v0 %v0 8
+		pushargr %v0		/* 5 */
+		addi %v0 %v0 8
+		pushargr %v0		/* 6 */
+		addi %v0 %v0 8
+		pushargr %v0		/* 7 */
+		addi %v0 %v0 8
+		pushargr %v0		/* 8 */
+		addi %v0 %v0 8
+		pushargr %v0		/* 9 */
+	finishi @sscanf
+
+	movi %v0 dchk
+	movi_d %f0 0.0
+loopd:
+	ldr_d %f1 %v0
+	beqr_d nextd %f0 %f1
+	calli @abort
+nextd:
+	addi_d %f0 %f0 1.0
+	bgei_d outd %f0 10.0
+	addi %v0 %v0 8
+	jmpi loopd
+outd:
+
+	prepare
+		pushargi buff
+		ellipsis
+	finishi @printf
+
+	/*
+	    sprintf(buff,
+		   "%d %.1f %d %.1f %d %.1f %d %.1f %d %.1f "
+		   "%d %.1f %d %.1f %d %.1f %d %.1f %d %.1f\n",
+		   0, 0.0, 1, 1.0, 2, 2.0, 3, 3.0, 4, 4.0,
+		   5, 5.0, 6, 6.0, 7, 7.0, 8, 8.0, 9, 9.0);
+	 */
+	prepare
+		pushargi buff
+		pushargi idfmt
+		ellipsis
+		pushargi 0
+		pushargi_d 0.0
+		pushargi 1
+		pushargi_d 1.0
+		pushargi 2
+		pushargi_d 2.0
+		pushargi 3
+		pushargi_d 3.0
+		pushargi 4
+		pushargi_d 4.0
+		pushargi 5
+		pushargi_d 5.0
+		pushargi 6
+		pushargi_d 6.0
+		pushargi 7
+		pushargi_d 7.0
+		pushargi 8
+		pushargi_d 8.0
+		pushargi 9
+		pushargi_d 9.0
+	finishi @sprintf
+
+	/*
+		sscanf(buff,
+		      "%d %lf %d %lf %d %lf %d %lf %d %lf "
+		      "%d %lf %d %lf %d %lf %d %lf %d %lf\n",
+		      ichk+0, dchk+0, ichk+1, dchk+1, ichk+2,
+		      dchk+2, ichk+3, dchk+3, ichk+4, dchk+4,
+		      ichk+5, dchk+5, ichk+6, dchk+6, ichk+7,
+		      dchk+7, ichk+8, dchk+8, ichk+9, dchk+9);
+	 */
+	movi %v0 ichk
+	movi %v1 dchk
+	prepare
+		pushargi buff
+		pushargi ldfmt
+		ellipsis
+		pushargr %v0		/* 0 */
+		addi %v0 %v0 4
+		pushargr %v1
+		addi %v1 %v1 8
+		pushargr %v0		/* 1 */
+		addi %v0 %v0 4
+		pushargr %v1
+		addi %v1 %v1 8
+		pushargr %v0		/* 2 */
+		addi %v0 %v0 4
+		pushargr %v1
+		addi %v1 %v1 8
+		pushargr %v0		/* 3 */
+		addi %v0 %v0 4
+		pushargr %v1
+		addi %v1 %v1 8
+		pushargr %v0		/* 4 */
+		addi %v0 %v0 4
+		pushargr %v1
+		addi %v1 %v1 8
+		pushargr %v0		/* 5 */
+		addi %v0 %v0 4
+		pushargr %v1
+		addi %v1 %v1 8
+		pushargr %v0		/* 6 */
+		addi %v0 %v0 4
+		pushargr %v1
+		addi %v1 %v1 8
+		pushargr %v0		/* 7 */
+		addi %v0 %v0 4
+		pushargr %v1
+		addi %v1 %v1 8
+		pushargr %v0		/* 8 */
+		addi %v0 %v0 4
+		pushargr %v1
+		addi %v1 %v1 8
+		pushargr %v0		/* 9 */
+		pushargr %v1
+	finishi @sscanf
+
+	movi %v0 ichk
+	movi %v1 dchk
+	movi %r0 0
+	movi_d %f0 0.0
+loopid:
+	ldr_i %r1 %v0
+	beqr checkd %r0 %r1
+	calli @abort
+checkd:
+	ldr_d %f1 %v1
+	beqr_d nextid %f0 %f1
+	calli @abort
+nextid:
+	addi %r0 %r0 1
+	addi_d %f0 %f0 1.0
+	bgei outid %r0 10
+	addi %v0 %v0 4
+	addi %v1 %v1 8
+	jmpi loopid
+outid:
+
+	prepare
+		pushargi buff
+		ellipsis
+	finishi @printf
+
+	ret
+	epilog

include/lightning.h

@@ -18,6 +18,10 @@
 #ifndef _lightning_h
 #define _lightning_h
 
+#if HAVE_CONFIG_H
+# include "config.h"
+#endif
+
 #include <unistd.h>
 #include <stdlib.h>
 #include <stdint.h>

lib/jit_arm-swf.c

@@ -1740,25 +1740,25 @@ _swf_absr_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
     if (jit_fpr_p(r1)) {
 	if (jit_fpr_p(r0) && !jit_thumb_p() && jit_armv5e_p() &&
 	    r0 != r1 && (reg = jit_get_reg_pair()) != JIT_NOREG) {
-	    LDRDIN(rn(reg), _FP_REGNO, swf_off(r1) + 8);
+	    LDRDIN(rn(reg), _FP_REGNO, swf_off(r1) + 4);
 	    swf_bici(rn(reg), rn(reg), 0x80000000);
-	    STRDIN(rn(reg), _FP_REGNO, swf_off(r0) + 8);
+	    STRDIN(rn(reg), _FP_REGNO, swf_off(r0) + 4);
 	    jit_unget_reg_pair(reg);
 	}
 	else {
 	    reg = jit_get_reg(jit_class_gpr);
-	    swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 8);
+	    swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 4);
 	    swf_bici(rn(reg), rn(reg), 0x80000000);
 	    if (jit_fpr_p(r0)) {
-		swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 8);
-		if (r0 != r1) {
-		    swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 4);
 		swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 4);
+		if (r0 != r1) {
+		    swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 8);
+		    swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 8);
 		}
 	    }
 	    else {
 		movr(r0, rn(reg));
-		swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 4);
+		swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 8);
 		movr(r0 + 1, rn(reg));
 	    }
 	    jit_unget_reg(reg);
@@ -1768,9 +1768,9 @@ _swf_absr_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 	reg = jit_get_reg(jit_class_gpr);
 	movr(rn(reg), r1);
 	swf_bici(rn(reg), rn(reg), 0x80000000);
-	swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 8);
-	movr(rn(reg), r1 + 1);
 	swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 4);
+	movr(rn(reg), r1 + 1);
+	swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 8);
 	jit_unget_reg(reg);
     }
     else {
@@ -1812,25 +1812,25 @@ _swf_negr_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
     if (jit_fpr_p(r1)) {
 	if (jit_fpr_p(r0) && !jit_thumb_p() && jit_armv5e_p() &&
 	    r0 != r1 && (reg = jit_get_reg_pair()) != JIT_NOREG) {
-	    LDRDIN(rn(reg), _FP_REGNO, swf_off(r1) + 8);
+	    LDRDIN(rn(reg), _FP_REGNO, swf_off(r1) + 4);
 	    EORI(rn(reg), rn(reg), encode_arm_immediate(0x80000000));
-	    STRDIN(rn(reg), _FP_REGNO, swf_off(r0) + 8);
+	    STRDIN(rn(reg), _FP_REGNO, swf_off(r0) + 4);
 	    jit_unget_reg_pair(reg);
 	}
 	else {
 	    reg = jit_get_reg(jit_class_gpr);
-	    swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 8);
+	    swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 4);
 	    xori(rn(reg), rn(reg), 0x80000000);
 	    if (jit_fpr_p(r0)) {
-		swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 8);
-		if (r0 != r1) {
-		    swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 4);
 		swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 4);
+		if (r0 != r1) {
+		    swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 8);
+		    swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 8);
 		}
 	    }
 	    else {
 		movr(r0, rn(reg));
-		swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 4);
+		swf_ldrin(rn(reg), _FP_REGNO, swf_off(r1) + 8);
 		movr(r0 + 1, rn(reg));
 	    }
 	    jit_unget_reg(reg);
@@ -1840,9 +1840,9 @@ _swf_negr_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 	reg = jit_get_reg(jit_class_gpr);
 	movr(rn(reg), r1);
 	xori(rn(reg), rn(reg), 0x80000000);
-	swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 8);
-	movr(rn(reg), r1 + 1);
 	swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 4);
+	movr(rn(reg), r1 + 1);
+	swf_strin(rn(reg), _FP_REGNO, swf_off(r0) + 8);
 	jit_unget_reg(reg);
     }
     else {

lib/jit_arm.c

@@ -563,7 +563,7 @@ _jit_pushargi_f(jit_state_t *_jit, jit_float32_t u)
 	    return;
 	}
     }
-    regno = jit_get_reg(jit_class_gpr);
+    regno = jit_get_reg(jit_class_fpr);
     jit_movi_f(regno, u);
     jit_stxi_f(_jit->function->call.size, JIT_SP, regno);
     jit_unget_reg(regno);
@@ -624,7 +624,7 @@ _jit_pushargi_d(jit_state_t *_jit, jit_float64_t u)
     }
     if (_jit->function->call.size & 7)
 	_jit->function->call.size += 4;
-    regno = jit_get_reg(jit_class_gpr);
+    regno = jit_get_reg(jit_class_fpr);
     jit_movi_d(regno, u);
     jit_stxi_d(_jit->function->call.size, JIT_SP, regno);
     jit_unget_reg(regno);

lib/jit_mips.c

@@ -421,17 +421,17 @@ _jit_getarg_d(jit_state_t *_jit, jit_int32_t u, jit_int32_t v)
 void
 _jit_pushargr(jit_state_t *_jit, jit_int32_t u)
 {
+    jit_word_t		offset;
+
     assert(_jit->function);
-    if (_jit->function->call.argf) {
+    offset = _jit->function->call.size >> 2;
-	_jit->function->call.argi = _jit->function->call.argf;
+    if (_jit->function->call.argf)
 	_jit->function->call.argf = 0;
-    }
+    if (offset < 4)
-    if (_jit->function->call.argi < 4) {
+	jit_movr(_A0 - offset, u);
-	jit_movr(_A0 - _jit->function->call.argi, u);
-	++_jit->function->call.argi;
-    }
     else
 	jit_stxi(_jit->function->call.size, JIT_SP, u);
+    _jit->function->call.argi = offset + 1;
     _jit->function->call.size += sizeof(jit_word_t);
 }
 
@@ -439,40 +439,51 @@ void
 _jit_pushargi(jit_state_t *_jit, jit_word_t u)
 {
     jit_int32_t		regno;
+    jit_word_t		offset;
 
     assert(_jit->function);
-    if (_jit->function->call.argf) {
+    offset = _jit->function->call.size >> 2;
-	_jit->function->call.argi = _jit->function->call.argf;
+    if (_jit->function->call.argf)
 	_jit->function->call.argf = 0;
-    }
+    if (offset < 4)
-    if (_jit->function->call.argi < 4) {
+	jit_movi(_A0 - offset, u);
-	jit_movi(_A0 - _jit->function->call.argi, u);
-	++_jit->function->call.argi;
-    }
     else {
 	regno = jit_get_reg(jit_class_gpr);
 	jit_movi(regno, u);
 	jit_stxi(_jit->function->call.size, JIT_SP, regno);
 	jit_unget_reg(regno);
     }
+    _jit->function->call.argi = offset + 1;
     _jit->function->call.size += sizeof(jit_word_t);
 }
 
 void
 _jit_pushargr_f(jit_state_t *_jit, jit_int32_t u)
 {
+    jit_word_t		offset;
+
     assert(_jit->function);
-    if (_jit->function->call.argi) {
+    offset = _jit->function->call.size >> 2;
-	jit_new_node_ww(jit_code_pushargr_f,
+    if (offset < 3) {
-			_A0 - _jit->function->call.argi, u);
+	if (offset & 1) {
-	_jit->function->call.argi += 2;
+	    ++offset;
+	    _jit->function->call.size += 4;
 	}
-    else if (_jit->function->call.argf < 4) {
+	if (_jit->function->call.argi)
-	jit_movr_f(_F12 - (_jit->function->call.argf >> 1), u);
+	    jit_new_node_ww(jit_code_pushargr_f, _A0 - offset, u);
-	_jit->function->call.argf += 2;
+	else
+	    jit_movr_f(_F12 - (offset >> 1), u);
     }
     else
 	jit_stxi_f(_jit->function->call.size, JIT_SP, u);
+    if (offset < 3) {
+	if (!_jit->function->call.argi)
+	    _jit->function->call.argf = offset + 2;
+	else
+	    _jit->function->call.argi = offset + 2;
+    }
+    else
+	_jit->function->call.argi = offset + 1;
     _jit->function->call.size += sizeof(jit_float32_t);
 }
 
@@ -480,17 +491,19 @@ void
 _jit_pushargi_f(jit_state_t *_jit, jit_float32_t u)
 {
     jit_int32_t		regno;
+    jit_word_t		offset;
 
-    if (_jit->function->call.argi) {
+    assert(_jit->function);
-	if (_jit->function->call.argi & 1)
+    offset = _jit->function->call.size >> 2;
-	    ++_jit->function->call.argi;
+    if (offset < 3) {
-	jit_new_node_wf(jit_code_pushargi_f,
+	if (offset & 1) {
-			_A0 - _jit->function->call.argi, u);
+	    ++offset;
-	_jit->function->call.argi += 2;
+	    _jit->function->call.size += 4;
 	}
-    else if (_jit->function->call.argf < 4) {
+	if (_jit->function->call.argi)
-	jit_movi_f(_F12 - (_jit->function->call.argf >> 1), u);
+	    jit_new_node_ww(jit_code_pushargi_f, _A0 - offset, u);
-	_jit->function->call.argf += 2;
+	else
+	    jit_movi_f(_F12 - (offset >> 1), u);
     }
     else {
 	assert(_jit->function);
@@ -499,26 +512,38 @@ _jit_pushargi_f(jit_state_t *_jit, jit_float32_t u)
 	jit_stxi_f(_jit->function->call.size, JIT_SP, regno);
 	jit_unget_reg(regno);
     }
+    if (offset < 3) {
+	if (!_jit->function->call.argi)
+	    _jit->function->call.argf = offset + 2;
+	else
+	    _jit->function->call.argi = offset + 2;
+    }
+    else
+	_jit->function->call.argi = offset + 1;
     _jit->function->call.size += sizeof(jit_float32_t);
 }
 
 void
 _jit_pushargr_d(jit_state_t *_jit, jit_int32_t u)
 {
+    jit_word_t		offset;
+
     assert(_jit->function);
-    if (_jit->function->call.argi) {
+    if (_jit->function->call.size & 7)
-	if (_jit->function->call.argi & 1)
+	_jit->function->call.size += 4;
-	    ++_jit->function->call.argi;
+    offset = _jit->function->call.size >> 2;
-	jit_new_node_ww(jit_code_pushargr_d,
+    if (offset < 3) {
-			_A0 - _jit->function->call.argi, u);
+	if (_jit->function->call.argi)
-	_jit->function->call.argi += 2;
+	    jit_new_node_ww(jit_code_pushargr_d, _A0 - offset, u);
-    }
+	else
-    else if (_jit->function->call.argf < 4) {
+	    jit_movr_d(_F12 - (offset >> 1), u);
-	jit_movr_d(_F12 - (_jit->function->call.argf >> 1), u);
-	_jit->function->call.argf += 2;
     }
     else
 	jit_stxi_d(_jit->function->call.size, JIT_SP, u);
+    if (offset < 3 && !_jit->function->call.argi)
+	_jit->function->call.argf = offset + 2;
+    else
+	_jit->function->call.argi = offset + 2;
     _jit->function->call.size += sizeof(jit_float64_t);
 }
 
@@ -526,25 +551,28 @@ void
 _jit_pushargi_d(jit_state_t *_jit, jit_float64_t u)
 {
     jit_int32_t		regno;
+    jit_word_t		offset;
 
-    if (_jit->function->call.argi) {
+    assert(_jit->function);
-	if (_jit->function->call.argi & 1)
+    if (_jit->function->call.size & 7)
-	    ++_jit->function->call.argi;
+	_jit->function->call.size += 4;
-	jit_new_node_wd(jit_code_pushargi_d,
+    offset = _jit->function->call.size >> 2;
-			_A0 - _jit->function->call.argi, u);
+    if (offset < 3) {
-	_jit->function->call.argi += 2;
+	if (_jit->function->call.argi)
-    }
+	    jit_new_node_ww(jit_code_pushargi_d, _A0 - offset, u);
-    else if (_jit->function->call.argf < 4) {
+	else
-	jit_movi_d(_F12 - (_jit->function->call.argf >> 1), u);
+	    jit_movi_d(_F12 - (offset >> 1), u);
-	_jit->function->call.argf += 2;
     }
     else {
-	assert(_jit->function);
 	regno = jit_get_reg(jit_class_fpr);
-	jit_movi_f(regno, u);
+	jit_movi_d(regno, u);
-	jit_stxi_f(_jit->function->call.size, JIT_SP, regno);
+	jit_stxi_d(_jit->function->call.size, JIT_SP, regno);
 	jit_unget_reg(regno);
     }
+    if (offset < 3 && !_jit->function->call.argi)
+	_jit->function->call.argf = offset + 2;
+    else
+	_jit->function->call.argi = offset + 2;
     _jit->function->call.size += sizeof(jit_float64_t);
 }
 

lib/jit_ppc.c

@@ -379,6 +379,8 @@ _jit_pushargr(jit_state_t *_jit, jit_int32_t u)
 	++_jit->function->call.argi;
     }
     else {
+	if (!_jit->function->call.size)
+	    _jit->function->call.size = 56;
 	jit_stxi(_jit->function->call.size, JIT_SP, u);
 	_jit->function->call.size += sizeof(jit_word_t);
     }
@@ -394,6 +396,8 @@ _jit_pushargi(jit_state_t *_jit, jit_word_t u)
 	++_jit->function->call.argi;
     }
     else {
+	if (!_jit->function->call.size)
+	    _jit->function->call.size = 56;
 	regno = jit_get_reg(jit_class_gpr);
 	jit_movi(regno, u);
 	jit_stxi(_jit->function->call.size, JIT_SP, regno);
@@ -424,7 +428,7 @@ _jit_pushargr_d(jit_state_t *_jit, jit_int32_t u)
 	if (!(_jit->function->call.kind & jit_call_varargs))
 	    return;
     }
-    if (_jit->function->call.argi < 6) {
+    if (_jit->function->call.argi < 8) {
 	jit_stxi_d(-8, JIT_FP, u);
 	jit_ldxi(JIT_RA0 - _jit->function->call.argi, JIT_FP, -8);
 	_jit->function->call.argi++;
@@ -432,8 +436,10 @@ _jit_pushargr_d(jit_state_t *_jit, jit_int32_t u)
 	_jit->function->call.argi++;
     }
     else {
+	if (!_jit->function->call.size)
+	    _jit->function->call.size = 56;
 	jit_stxi_d(_jit->function->call.size, JIT_SP, u);
-	_jit->function->call.size += (sizeof(jit_float64_t) + 8) & -8;
+	_jit->function->call.size += sizeof(jit_float64_t);
     }
 }
 
@@ -451,7 +457,7 @@ _jit_pushargi_d(jit_state_t *_jit, jit_float64_t u)
     }
     regno = jit_get_reg(jit_class_fpr);
     jit_movi_d(regno, u);
-    if (_jit->function->call.argi < 6) {
+    if (_jit->function->call.argi < 8) {
 	jit_stxi_d(-8, JIT_FP, regno);
 	jit_ldxi(JIT_RA0 - _jit->function->call.argi, JIT_FP, -8);
 	_jit->function->call.argi++;
@@ -459,8 +465,10 @@ _jit_pushargi_d(jit_state_t *_jit, jit_float64_t u)
 	_jit->function->call.argi++;
     }
     else {
+	if (!_jit->function->call.size)
+	    _jit->function->call.size = 56;
 	jit_stxi_d(_jit->function->call.size, JIT_SP, regno);
-	_jit->function->call.size += (sizeof(jit_float64_t) + 8) & -8;
+	_jit->function->call.size += sizeof(jit_float64_t);
     }
     jit_unget_reg(regno);
 }