diff --git a/ChangeLog b/ChangeLog
index ebff1f877..3e6e0dbe8 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,31 @@
+2014-09-24 Paulo Andrade <pcpa@gnu.org>
+
+        * check/lightning.c: Add a hopefully temporary kludge to not use
+	sprintf and sscanf returned by dlsym. This is required to pass
+	the varargs test.
+
+        * include/lightning/jit_private.h: Use symbolic name for first
+	integer register argument, as this is different in sysv and
+	win64 abi.
+
+        * include/lightning/jit_x86.h: Add conditionals and definitions
+	for Windows x64 (under __CYGWIN__ preprocessor conditional).
+
+        * lib/jit_x86-cpu.c: Correct one instruction encoding bug, that
+	was working by accident. Only use rax to rdx for some byte
+	operations to work on compatibility mode (that is, to generate
+	the proper encoding, instead of actually generating encoding
+	for high byte registers, e.g. %bh).
+	Add proper prolog and epilog for windows x64.
+
+        * lib/jit_x86-sse.c: Correct a swapped rex prefix for float
+	operations.
+
+        * lib/jit_x86.c: Adjust to support Windows x64 abi.
+
+	* check/check.x87.nodata.sh: New file, previously used but that
+	was missing git.
+
 2014-09-07 Paulo Andrade <pcpa@gnu.org>
 
 	* lib/lightning.c: Mark all registers advertised as live, as
diff --git a/check/check.x87.nodata.sh b/check/check.x87.nodata.sh
new file mode 100755
index 000000000..1582e9f62
--- /dev/null
+++ b/check/check.x87.nodata.sh
@@ -0,0 +1,15 @@
+#!/bin/sh
+test=`basename $0 | sed -e 's|\.x87.nodata$||'`
+./lightning -data=0 -mx87=1 $srcdir/$test.tst | tr -d \\r > $test.out
+if test $? != 0; then
+  exit $?
+fi
+
+cmp -s $srcdir/$test.ok $test.out
+result=$?
+if test $result != 0; then
+    diff $srcdir/$test.ok $test.out
+    rm $test.out
+    exit 1
+fi
+rm $test.out
diff --git a/check/lightning.c b/check/lightning.c
index 72555be12..47e39a3c3 100644
--- a/check/lightning.c
+++ b/check/lightning.c
@@ -2384,9 +2384,20 @@ dynamic(void)
     char	*string;
     (void)identifier('@');
     if ((label = get_label_by_name(parser.string)) == NULL) {
-	value = dlsym(DL_HANDLE, parser.string + 1);
-	if ((string = dlerror()))
-	    error("%s", string);
+#if __CYGWIN__
+	/* FIXME kludge to pass varargs test case, otherwise,
+	 * will not print/scan float values */
+	if (strcmp(parser.string + 1, "sprintf") == 0)
+	    value = sprintf;
+	else if (strcmp(parser.string + 1, "sscanf") == 0)
+	    value = sscanf;
+	else
+#endif
+	{
+	    value = dlsym(DL_HANDLE, parser.string + 1);
+	    if ((string = dlerror()))
+		error("%s", string);
+	}
 	label = new_label(label_kind_dynamic, parser.string, value);
     }
     parser.type = type_p;
diff --git a/include/lightning/jit_private.h b/include/lightning/jit_private.h
index 9b50d1f86..b620a1b06 100644
--- a/include/lightning/jit_private.h
+++ b/include/lightning/jit_private.h
@@ -54,7 +54,11 @@
 #    define JIT_FRET		_ST0
 typedef jit_uint32_t		jit_regset_t;
 #  else
-#    define JIT_RA0		_RDI
+#    if __CYGWIN__
+#      define JIT_RA0		_RCX
+#    else
+#      define JIT_RA0		_RDI
+#    endif
 #    define JIT_FA0		_XMM0
 #    define JIT_FRET		_XMM0
 typedef jit_uint64_t		jit_regset_t;
diff --git a/include/lightning/jit_x86.h b/include/lightning/jit_x86.h
index bf33125c2..b96eab607 100644
--- a/include/lightning/jit_x86.h
+++ b/include/lightning/jit_x86.h
@@ -60,37 +60,89 @@ typedef enum {
     _XMM0,	_XMM1,	_XMM2,	_XMM3,	_XMM4,	_XMM5,	_XMM6,	 _XMM7,
 #  define jit_sse_reg_p(reg)	((reg) >= _XMM0 && (reg) <= _XMM7)
 #else
-#  define jit_arg_reg_p(i)	((i) >= 0 && (i) < 6)
-#  define jit_r(i)		(_RAX + (i))
-#  define jit_r_num()		4
-#  define jit_v(i)		(_RBX + (i))
-#  define jit_v_num()		4
-#  define jit_arg_f_reg_p(i)	((i) >= 0 && (i) < 8)
-#  define jit_f(index)		(_XMM8 + (index))
-#  define jit_f_num()		8
-#  define JIT_R0		_RAX
-#  define JIT_R1		_R10
-#  define JIT_R2		_R11
-#  define JIT_R3		_R12
+#  if __CYGWIN__
+#    define jit_arg_reg_p(i)	((i) >= 0 && (i) < 4)
+#    define jit_r(i)		(_RAX + (i))
+#    define jit_r_num()		3
+#    define jit_v(i)		(_RBX + (i))
+#    define jit_v_num()		7
+#    define jit_arg_f_reg_p(i)	jit_arg_reg_p(i)
+#    define jit_f(index)	(_XMM4 + (index))
+#    define jit_f_num()		12
+#    define JIT_R0		_RAX
+#    define JIT_R1		_R10
+#    define JIT_R2		_R11
+#    define JIT_V0		_RBX
+#    define JIT_V1		_RDI
+#    define JIT_V2		_RSI
+#    define JIT_V3		_R12
+#    define JIT_V4		_R13
+#    define JIT_V5		_R14
+#    define JIT_V6		_R15
+    /* Volatile - Return value register */
+    _RAX,
+    /* Volatile */
+    _R10,	_R11,
+    /* Nonvolatile */
+    _RBX,	_RDI,	_RSI,
+    _R12,	_R13,	_R14,	_R15,
+    /* Volatile - Integer arguments (4 to 1) */
+    _R9,	_R8,	_RDX,	_RCX,
+    /* Nonvolatile */
+    _RSP,	_RBP,
+#    define JIT_F0		_XMM4
+#    define JIT_F1		_XMM5
+#    define JIT_F2		_XMM6
+#    define JIT_F3		_XMM7
+#    define JIT_F4		_XMM8
+#    define JIT_F5		_XMM9
+#    define JIT_F6		_XMM10
+#    define JIT_F7		_XMM11
+#    define JIT_F8		_XMM12
+#    define JIT_F9		_XMM13
+#    define JIT_F10		_XMM14
+#    define JIT_F11		_XMM15
+    /* Volatile */
+    _XMM4,	_XMM5,
+    /* Nonvolatile */
+    _XMM6,	_XMM7,	_XMM8,	_XMM9,	_XMM10,
+    _XMM11,	_XMM12, _XMM13,	_XMM14,	_XMM15,
+    /* Volatile - FP arguments (4 to 1) */
+    _XMM3,	_XMM2,	_XMM1,	_XMM0,
+#    define jit_sse_reg_p(reg)	((reg) >= _XMM4 && (reg) <= _XMM0)
+#  else
+#    define jit_arg_reg_p(i)	((i) >= 0 && (i) < 6)
+#    define jit_r(i)		(_RAX + (i))
+#    define jit_r_num()		4
+#    define jit_v(i)		(_RBX + (i))
+#    define jit_v_num()		4
+#    define jit_arg_f_reg_p(i)	((i) >= 0 && (i) < 8)
+#    define jit_f(index)	(_XMM8 + (index))
+#    define jit_f_num()		8
+#    define JIT_R0		_RAX
+#    define JIT_R1		_R10
+#    define JIT_R2		_R11
+#    define JIT_R3		_R12
     _RAX,	_R10,	_R11,	_R12,
-#  define JIT_V0		_RBX
-#  define JIT_V1		_R13
-#  define JIT_V2		_R14
-#  define JIT_V3		_R15
+#    define JIT_V0		_RBX
+#    define JIT_V1		_R13
+#    define JIT_V2		_R14
+#    define JIT_V3		_R15
     _RBX,	_R13,	_R14,	_R15,
     _R9,	_R8,	_RCX,	_RDX,	_RSI,	_RDI,
     _RSP,	_RBP,
-#  define JIT_F0		_XMM8
-#  define JIT_F1		_XMM9
-#  define JIT_F2		_XMM10
-#  define JIT_F3		_XMM11
-#  define JIT_F4		_XMM12
-#  define JIT_F5		_XMM13
-#  define JIT_F6		_XMM14
-#  define JIT_F7		_XMM15
+#    define JIT_F0		_XMM8
+#    define JIT_F1		_XMM9
+#    define JIT_F2		_XMM10
+#    define JIT_F3		_XMM11
+#    define JIT_F4		_XMM12
+#    define JIT_F5		_XMM13
+#    define JIT_F6		_XMM14
+#    define JIT_F7		_XMM15
     _XMM8,	_XMM9,	_XMM10,	_XMM11,	_XMM12,	_XMM13,	_XMM14,	_XMM15,
     _XMM7,	_XMM6,	_XMM5,	_XMM4,	_XMM3,	_XMM2,	_XMM1,	_XMM0,
-#  define jit_sse_reg_p(reg)	((reg) >= _XMM8 && (reg) <= _XMM0)
+#    define jit_sse_reg_p(reg)	((reg) >= _XMM8 && (reg) <= _XMM0)
+#  endif
 #endif
     _ST0,	_ST1,	_ST2,	_ST3,	_ST4,	_ST5,	_ST6,	_ST7,
 #  define JIT_NOREG		_NOREG
diff --git a/lib/jit_x86-cpu.c b/lib/jit_x86-cpu.c
index ce67c20b1..c24ad16e8 100644
--- a/lib/jit_x86-cpu.c
+++ b/lib/jit_x86-cpu.c
@@ -39,10 +39,15 @@
 #    define can_sign_extend_int_p(im)					\
 	(((im) >= 0 && (long)(im) <=  0x7fffffffL) ||			\
 	 ((im) <  0 && (long)(im) >  -0x80000000L))
-#  define can_zero_extend_int_p(im)					\
-    ((im) >= 0 && (im) < 0x80000000L)
-#    define fits_uint32_p(im)		((im & 0xffffffff00000000L) == 0)
-#    define reg8_p(rn)			1
+#    define can_zero_extend_int_p(im)					\
+	((im) >= 0 && (im) < 0x80000000L)
+#    define fits_uint32_p(im)		(((im) & 0xffffffff00000000L) == 0)
+#    if __CYGWIN__
+#      define reg8_p(rn)						\
+	(r7(rn) >= _RAX_REGNO && r7(rn) <= _RBX_REGNO)
+#    else
+#      define reg8_p(rn)		1
+#    endif
 #  endif
 #  define _RAX_REGNO			0
 #  define _RCX_REGNO			1
@@ -60,8 +65,8 @@
 #  define _R13_REGNO			13
 #  define _R14_REGNO			14
 #  define _R15_REGNO			15
-#  define r7(reg)			(reg & 7)
-#  define r8(reg)			(reg & 15)
+#  define r7(reg)			((reg) & 7)
+#  define r8(reg)			((reg) & 15)
 #  define _SCL1				0x00
 #  define _SCL2				0x01
 #  define _SCL4				0x02
@@ -634,6 +639,9 @@ static void _prolog(jit_state_t*, jit_node_t*);
 static void _epilog(jit_state_t*, jit_node_t*);
 #  define patch_at(node, instr, label)	_patch_at(_jit, node, instr, label)
 static void _patch_at(jit_state_t*, jit_node_t*, jit_word_t, jit_word_t);
+#  if __WORDSIZE == 64 && !defined(HAVE_FFSL)
+static int ffsl(long);
+#  endif
 #endif
 
 #if CODE
@@ -1871,7 +1879,7 @@ _ci(jit_state_t *_jit,
 	reg = jit_get_reg(jit_class_gpr|jit_class_rg8);
 	ixorr(rn(reg), rn(reg));
 	icmpi(r1, i0);
-	cc(code, reg);
+	cc(code, rn(reg));
 	movr(r0, rn(reg));
 	jit_unget_reg(reg);
     }
@@ -2596,6 +2604,7 @@ _str_c(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
     else {
 	reg = jit_get_reg(jit_class_gpr|jit_class_rg8);
 	movr(rn(reg), r1);
+	rex(0, 0, rn(reg), _NOREG, r0);
 	ic(0x88);
 	rx(rn(reg), 0, r0, _NOREG, _SCL1);
 	jit_unget_reg(reg);
@@ -2615,6 +2624,7 @@ _sti_c(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0)
 	else {
 	    reg = jit_get_reg(jit_class_gpr|jit_class_rg8);
 	    movr(rn(reg), r0);
+	    rex(0, 0, rn(reg), _NOREG, _NOREG);
 	    ic(0x88);
 	    rx(rn(reg), i0, _NOREG, _NOREG, _SCL1);
 	    jit_unget_reg(reg);
@@ -2719,6 +2729,7 @@ _stxr_c(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
     else {
 	reg = jit_get_reg(jit_class_gpr|jit_class_rg8);
 	movr(rn(reg), r2);
+	rex(0, 0, rn(reg), r1, r0);
 	ic(0x88);
 	rx(rn(reg), 0, r0, r1, _SCL1);
 	jit_unget_reg(reg);
@@ -2738,6 +2749,7 @@ _stxi_c(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
 	else {
 	    reg = jit_get_reg(jit_class_gpr|jit_class_rg8);
 	    movr(rn(reg), r1);
+	    rex(0, 0, rn(reg), _NOREG, r0);
 	    ic(0x88);
 	    rx(rn(reg), i0, r0, _NOREG, _SCL1);
 	    jit_unget_reg(reg);
@@ -3310,14 +3322,18 @@ _jmpi(jit_state_t *_jit, jit_word_t i0)
 static void
 _prolog(jit_state_t *_jit, jit_node_t *node)
 {
-#if __WORDSIZE == 32
-    _jitc->function->stack = (((_jitc->function->self.alen -
-			       _jitc->function->self.aoff) + 15) & -16) + 12;
+#if __WORDSIZE == 64 && __CYGWIN__
+    _jitc->function->stack = (((/* first 32 bytes must be allocated */
+				(_jitc->function->self.alen > 32 ?
+				 _jitc->function->self.alen : 32) -
+				/* align stack at 16 bytes */
+				_jitc->function->self.aoff) + 15) & -16) +
+	stack_adjust;
 #else
     _jitc->function->stack = (((_jitc->function->self.alen -
-			       _jitc->function->self.aoff) + 15) & -16) + 8;
+			       _jitc->function->self.aoff) + 15) & -16) +
+	stack_adjust;
 #endif
-
     /* callee save registers */
     subi(_RSP_REGNO, _RSP_REGNO, stack_framesize - sizeof(jit_word_t));
 #if __WORDSIZE == 32
@@ -3328,6 +3344,42 @@ _prolog(jit_state_t *_jit, jit_node_t *node)
     if (jit_regset_tstbit(&_jitc->function->regset, _RBX))
 	stxi( 4, _RSP_REGNO, _RBX_REGNO);
 #else
+#  if __CYGWIN__
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM15))
+	sse_stxi_d(136, _RSP_REGNO, _XMM15_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM14))
+	sse_stxi_d(128, _RSP_REGNO, _XMM14_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM13))
+	sse_stxi_d(120, _RSP_REGNO, _XMM13_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM12))
+	sse_stxi_d(112, _RSP_REGNO, _XMM12_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM11))
+	sse_stxi_d(104, _RSP_REGNO, _XMM11_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM10))
+	sse_stxi_d(96, _RSP_REGNO, _XMM10_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM9))
+	sse_stxi_d(88, _RSP_REGNO, _XMM9_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM8))
+	sse_stxi_d(80, _RSP_REGNO, _XMM8_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM7))
+	sse_stxi_d(72, _RSP_REGNO, _XMM7_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM6))
+	sse_stxi_d(64, _RSP_REGNO, _XMM6_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _R15))
+	stxi(56, _RSP_REGNO, _R15_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _R14))
+	stxi(48, _RSP_REGNO, _R14_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _R13))
+	stxi(40, _RSP_REGNO, _R13_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _R12))
+	stxi(32, _RSP_REGNO, _R12_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _RSI))
+	stxi(24, _RSP_REGNO, _RSI_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _RDI))
+	stxi(16, _RSP_REGNO, _RDI_REGNO);
+    if (jit_regset_tstbit(&_jitc->function->regset, _RBX))
+	stxi( 8, _RSP_REGNO, _RBX_REGNO);
+#  else
     if (jit_regset_tstbit(&_jitc->function->regset, _RBX))
 	stxi(40, _RSP_REGNO, _RBX_REGNO);
     if (jit_regset_tstbit(&_jitc->function->regset, _R12))
@@ -3338,6 +3390,7 @@ _prolog(jit_state_t *_jit, jit_node_t *node)
 	stxi(16, _RSP_REGNO, _R14_REGNO);
     if (jit_regset_tstbit(&_jitc->function->regset, _R15))
 	stxi( 8, _RSP_REGNO, _R15_REGNO);
+#  endif
 #endif
     stxi(0, _RSP_REGNO, _RBP_REGNO);
     movr(_RBP_REGNO, _RSP_REGNO);
@@ -3359,6 +3412,42 @@ _epilog(jit_state_t *_jit, jit_node_t *node)
     if (jit_regset_tstbit(&_jitc->function->regset, _RBX))
 	ldxi(_RBX_REGNO, _RSP_REGNO,  4);
 #else
+#  if __CYGWIN__
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM15))
+	sse_ldxi_d(_XMM15_REGNO, _RSP_REGNO, 136);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM14))
+	sse_ldxi_d(_XMM14_REGNO, _RSP_REGNO, 128);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM13))
+	sse_ldxi_d(_XMM13_REGNO, _RSP_REGNO, 120);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM12))
+	sse_ldxi_d(_XMM12_REGNO, _RSP_REGNO, 112);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM11))
+	sse_ldxi_d(_XMM11_REGNO, _RSP_REGNO, 104);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM10))
+	sse_ldxi_d(_XMM10_REGNO, _RSP_REGNO, 96);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM9))
+	sse_ldxi_d(_XMM9_REGNO, _RSP_REGNO, 88);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM8))
+	sse_ldxi_d(_XMM8_REGNO, _RSP_REGNO, 80);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM7))
+	sse_ldxi_d(_XMM7_REGNO, _RSP_REGNO, 72);
+    if (jit_regset_tstbit(&_jitc->function->regset, _XMM6))
+	sse_ldxi_d(_XMM6_REGNO, _RSP_REGNO, 64);
+    if (jit_regset_tstbit(&_jitc->function->regset, _R15))
+	ldxi(_R15_REGNO, _RSP_REGNO, 56);
+    if (jit_regset_tstbit(&_jitc->function->regset, _R14))
+	ldxi(_R14_REGNO, _RSP_REGNO, 48);
+    if (jit_regset_tstbit(&_jitc->function->regset, _R13))
+	ldxi(_R13_REGNO, _RSP_REGNO, 40);
+    if (jit_regset_tstbit(&_jitc->function->regset, _R12))
+	ldxi(_R12_REGNO, _RSP_REGNO, 32);
+    if (jit_regset_tstbit(&_jitc->function->regset, _RSI))
+	ldxi(_RSI_REGNO, _RSP_REGNO, 24);
+    if (jit_regset_tstbit(&_jitc->function->regset, _RDI))
+	ldxi(_RDI_REGNO, _RSP_REGNO, 16);
+    if (jit_regset_tstbit(&_jitc->function->regset, _RBX))
+	ldxi(_RBX_REGNO, _RSP_REGNO,  8);
+#  else
     if (jit_regset_tstbit(&_jitc->function->regset, _RBX))
 	ldxi(_RBX_REGNO, _RSP_REGNO, 40);
     if (jit_regset_tstbit(&_jitc->function->regset, _R12))
@@ -3369,6 +3458,7 @@ _epilog(jit_state_t *_jit, jit_node_t *node)
 	ldxi(_R14_REGNO, _RSP_REGNO, 16);
     if (jit_regset_tstbit(&_jitc->function->regset, _R15))
 	ldxi(_R15_REGNO, _RSP_REGNO,  8);
+#  endif
 #endif
     ldxi(_RBP_REGNO, _RSP_REGNO, 0);
     addi(_RSP_REGNO, _RSP_REGNO, stack_framesize - sizeof(jit_word_t));
@@ -3392,4 +3482,23 @@ _patch_at(jit_state_t *_jit, jit_node_t *node,
 	    break;
     }
 }
+
+#  if __WORDSIZE == 64 && !defined(HAVE_FFSL)
+static int
+ffsl(long i)
+{
+    int		bit;
+#    if __CYGWIN__
+    /* Bug workaround */
+    if ((int)i == (int)0x80000000)
+	bit = 32;
+    else
+#    endif
+    if ((bit = ffs((int)i)) == 0) {
+	if ((bit = ffs((int)((unsigned long)i >> 32))))
+	    bit += 32;
+    }
+    return (bit);
+}
+#  endif
 #endif
diff --git a/lib/jit_x86-sse.c b/lib/jit_x86-sse.c
index edac0fcba..d3bd7aeea 100644
--- a/lib/jit_x86-sse.c
+++ b/lib/jit_x86-sse.c
@@ -18,6 +18,16 @@
  */
 
 #if PROTO
+#  define _XMM6_REGNO			6
+#  define _XMM7_REGNO			7
+#  define _XMM8_REGNO			8
+#  define _XMM9_REGNO			9
+#  define _XMM10_REGNO			10
+#  define _XMM11_REGNO			11
+#  define _XMM12_REGNO			12
+#  define _XMM13_REGNO			13
+#  define _XMM14_REGNO			14
+#  define _XMM15_REGNO			15
 #define X86_SSE_MOV			0x10
 #define X86_SSE_MOV1			0x11
 #define X86_SSE_MOVLP			0x12
@@ -457,7 +467,7 @@ _sse_b##name##i_##type(jit_state_t *_jit,				\
 static void
 _sser(jit_state_t *_jit, jit_int32_t c, jit_int32_t r0, jit_int32_t r1)
 {
-    rex(0, 0, r1, 0, r0);
+    rex(0, 0, r0, 0, r1);
     ic(0x0f);
     ic(c);
     mrm(0x03, r7(r0), r7(r1));
diff --git a/lib/jit_x86.c b/lib/jit_x86.c
index 5eaad68e9..dc386bea2 100644
--- a/lib/jit_x86.c
+++ b/lib/jit_x86.c
@@ -24,13 +24,17 @@
 #define rn(reg)				(jit_regno(_rvs[jit_regno(reg)].spec))
 
 #if __WORDSIZE == 32
-#  define stack_alignment		4
 #  define stack_framesize		20
-#    define CVT_OFFSET			-12
+#  define stack_adjust			12
+#  define CVT_OFFSET			-12
 #else
-#  define stack_alignment		8
-#  define stack_framesize		56
-#    define CVT_OFFSET			-8
+#  if __CYGWIN__
+#    define stack_framesize		152
+#  else
+#    define stack_framesize		56
+#  endif
+#  define stack_adjust			8
+#  define CVT_OFFSET			-8
 #endif
 
 /*
@@ -83,6 +87,40 @@ jit_register_t		_rvs[] = {
     { rc(fpr) | 5,			"st(5)" },
     { rc(fpr) | 6,			"st(6)" },
     { rc(fpr) | 7,			"st(7)" },
+#else
+#  if __CYGWIN__
+    { rc(gpr) | rc(rg8) | 0,		"%rax" },
+    { rc(gpr) | rc(rg8) | rc(rg8) | 10,	"%r10" },
+    { rc(gpr) | rc(rg8) | rc(rg8) | 11,	"%r11" },
+    { rc(sav) | rc(rg8) | rc(gpr) | 3,	"%rbx" },
+    { rc(sav) | rc(gpr) | 7,		"%rdi" },
+    { rc(sav) | rc(gpr) | 6,		"%rsi" },
+    { rc(sav) | rc(gpr) | 12,		"%r12" },
+    { rc(sav) | rc(gpr) | 13,		"%r13" },
+    { rc(sav) | rc(gpr) | 14,		"%r14" },
+    { rc(sav) | rc(gpr) | 15,		"%r15" },
+    { rc(arg) | rc(rg8) | rc(gpr) | 9,	"%r9" },
+    { rc(arg) | rc(rg8) | rc(gpr) | 8,	"%r8" },
+    { rc(arg) | rc(rg8) | rc(gpr) | 2,	"%rdx" },
+    { rc(arg) | rc(rg8) | rc(gpr) | 1,	"%rcx" },
+    { rc(sav) | 4,			"%rsp" },
+    { rc(sav) | 5,			"%rbp" },
+    { rc(xpr) | rc(fpr) | 4,		"%xmm4" },
+    { rc(xpr) | rc(fpr) | 5,		"%xmm5" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 6,	"%xmm6" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 7,	"%xmm7" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 8,	"%xmm8" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 9,	"%xmm9" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 10,	"%xmm10" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 11,	"%xmm11" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 12,	"%xmm12" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 13,	"%xmm13" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 14,	"%xmm14" },
+    { rc(sav) | rc(xpr) | rc(fpr) | 15,	"%xmm15" },
+    { rc(xpr) | rc(arg) | rc(fpr) | 3,	"%xmm3" },
+    { rc(xpr) | rc(arg) | rc(fpr) | 2,	"%xmm2" },
+    { rc(xpr) | rc(arg) | rc(fpr) | 1,	"%xmm1" },
+    { rc(xpr) | rc(arg) | rc(fpr) | 0,	"%xmm0" },
 #else
     /* %rax is a pseudo flag argument for varargs functions */
     { rc(arg) | rc(gpr) | rc(rg8) | 0,	"%rax" },
@@ -117,6 +155,7 @@ jit_register_t		_rvs[] = {
     { rc(xpr) | rc(arg) | rc(fpr) | 2,	"%xmm2" },
     { rc(xpr) | rc(arg) | rc(fpr) | 1,	"%xmm1" },
     { rc(xpr) | rc(arg) | rc(fpr) | 0,	"%xmm0" },
+#  endif
     { rc(fpr) | 0,			"st(0)" },
     { rc(fpr) | 1,			"st(1)" },
     { rc(fpr) | 2,			"st(2)" },
@@ -429,8 +468,12 @@ _jit_arg(jit_state_t *_jit)
 
     assert(_jitc->function);
 #if __WORDSIZE == 64
-    if (_jitc->function->self.argi < 6)
+    if (jit_arg_reg_p(_jitc->function->self.argi)) {
 	offset = _jitc->function->self.argi++;
+#  if __CYGWIN__
+	_jitc->function->self.size += sizeof(jit_word_t);
+#  endif
+    }
     else
 #endif
     {
@@ -446,7 +489,11 @@ _jit_arg_reg_p(jit_state_t *_jit, jit_int32_t offset)
 #if __WORDSIZE == 32
     return (0);
 #else
+#  if __CYGWIN__
+    return (offset >= 0 && offset < 4);
+#  else
     return (offset >= 0 && offset < 6);
+#  endif
 #endif
 }
 
@@ -457,8 +504,15 @@ _jit_arg_f(jit_state_t *_jit)
 
     assert(_jitc->function);
 #if __WORDSIZE == 64
-    if (_jitc->function->self.argf < 8)
+#  if __CYGWIN__
+    if (jit_arg_reg_p(_jitc->function->self.argi)) {
+	offset = _jitc->function->self.argi++;
+	_jitc->function->self.size += sizeof(jit_word_t);
+    }
+#  else
+    if (jit_arg_f_reg_p(_jitc->function->self.argf))
 	offset = _jitc->function->self.argf++;
+#  endif
     else
 #endif
     {
@@ -478,7 +532,11 @@ _jit_arg_f_reg_p(jit_state_t *_jit, jit_int32_t offset)
 #if __WORDSIZE == 32
     return (0);
 #else
+#  if __CYGWIN__
+    return (offset >= 0 && offset < 4);
+#  else
     return (offset >= 0 && offset < 8);
+#  endif
 #endif
 }
 
@@ -489,8 +547,15 @@ _jit_arg_d(jit_state_t *_jit)
 
     assert(_jitc->function);
 #if __WORDSIZE == 64
-    if (_jitc->function->self.argf < 8)
+#  if __CYGWIN__
+    if (jit_arg_reg_p(_jitc->function->self.argi)) {
+	offset = _jitc->function->self.argi++;
+	_jitc->function->self.size += sizeof(jit_word_t);
+    }
+#  else
+    if (jit_arg_f_reg_p(_jitc->function->self.argf))
 	offset = _jitc->function->self.argf++;
+#  endif
     else
 #endif
     {
@@ -510,8 +575,8 @@ void
 _jit_getarg_c(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
 #if __WORDSIZE == 64
-    if (v->u.w < 6)
-	jit_extr_c(u, _RDI - v->u.w);
+    if (jit_arg_reg_p(v->u.w))
+	jit_extr_c(u, JIT_RA0 - v->u.w);
     else
 #endif
 	jit_ldxi_c(u, _RBP, v->u.w);
@@ -521,8 +586,8 @@ void
 _jit_getarg_uc(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
 #if __WORDSIZE == 64
-    if (v->u.w < 6)
-	jit_extr_uc(u, _RDI - v->u.w);
+    if (jit_arg_reg_p(v->u.w))
+	jit_extr_uc(u, JIT_RA0 - v->u.w);
     else
 #endif
 	jit_ldxi_uc(u, _RBP, v->u.w);
@@ -532,8 +597,8 @@ void
 _jit_getarg_s(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
 #if __WORDSIZE == 64
-    if (v->u.w < 6)
-	jit_extr_s(u, _RDI - v->u.w);
+    if (jit_arg_reg_p(v->u.w))
+	jit_extr_s(u, JIT_RA0 - v->u.w);
     else
 #endif
 	jit_ldxi_s(u, _RBP, v->u.w);
@@ -543,8 +608,8 @@ void
 _jit_getarg_us(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
 #if __WORDSIZE == 64
-    if (v->u.w < 6)
-	jit_extr_us(u, _RDI - v->u.w);
+    if (jit_arg_reg_p(v->u.w))
+	jit_extr_us(u, JIT_RA0 - v->u.w);
     else
 #endif
 	jit_ldxi_us(u, _RBP, v->u.w);
@@ -554,8 +619,8 @@ void
 _jit_getarg_i(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
 #if __WORDSIZE == 64
-    if (v->u.w < 6)
-	jit_extr_i(u, _RDI - v->u.w);
+    if (jit_arg_reg_p(v->u.w))
+	jit_extr_i(u, JIT_RA0 - v->u.w);
     else
 #endif
 	jit_ldxi_i(u, _RBP, v->u.w);
@@ -565,8 +630,8 @@ _jit_getarg_i(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 void
 _jit_getarg_ui(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
-    if (v->u.w < 6)
-	jit_extr_ui(u, _RDI - v->u.w);
+    if (jit_arg_reg_p(v->u.w))
+	jit_extr_ui(u, JIT_RA0 - v->u.w);
     else
 	jit_ldxi_ui(u, _RBP, v->u.w);
 }
@@ -574,8 +639,8 @@ _jit_getarg_ui(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 void
 _jit_getarg_l(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
-    if (v->u.w < 6)
-	jit_movr(u, _RDI - v->u.w);
+    if (jit_arg_reg_p(v->u.w))
+	jit_movr(u, JIT_RA0 - v->u.w);
     else
 	jit_ldxi_l(u, _RBP, v->u.w);
 }
@@ -585,7 +650,7 @@ void
 _jit_getarg_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
 #if __WORDSIZE == 64
-    if (v->u.w < 8)
+    if (jit_arg_f_reg_p(v->u.w))
 	jit_movr_f(u, _XMM0 - v->u.w);
     else
 #endif
@@ -596,7 +661,7 @@ void
 _jit_getarg_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 {
 #if __WORDSIZE == 64
-    if (v->u.w < 8)
+    if (jit_arg_f_reg_p(v->u.w))
 	jit_movr_d(u, _XMM0 - v->u.w);
     else
 #endif
@@ -608,9 +673,12 @@ _jit_pushargr(jit_state_t *_jit, jit_int32_t u)
 {
     assert(_jitc->function);
 #if __WORDSIZE == 64
-    if (_jitc->function->call.argi < 6) {
-	jit_movr(_RDI - _jitc->function->call.argi, u);
+    if (jit_arg_reg_p(_jitc->function->call.argi)) {
+	jit_movr(JIT_RA0 - _jitc->function->call.argi, u);
 	++_jitc->function->call.argi;
+#  if __CYGWIN__
+	_jitc->function->call.size += sizeof(jit_word_t);
+#  endif
     }
     else
 #endif
@@ -627,9 +695,12 @@ _jit_pushargi(jit_state_t *_jit, jit_word_t u)
 
     assert(_jitc->function);
 #if __WORDSIZE == 64
-    if (_jitc->function->call.argi < 6) {
-	jit_movi(_RDI - _jitc->function->call.argi, u);
+    if (jit_arg_reg_p(_jitc->function->call.argi)) {
+	jit_movi(JIT_RA0 - _jitc->function->call.argi, u);
 	++_jitc->function->call.argi;
+#  if __CYGWIN__
+	_jitc->function->call.size += sizeof(jit_word_t);
+#  endif
     }
     else
 #endif
@@ -647,10 +718,24 @@ _jit_pushargr_f(jit_state_t *_jit, jit_int32_t u)
 {
     assert(_jitc->function);
 #if __WORDSIZE == 64
-    if (_jitc->function->call.argf < 8) {
+#  if __CYGWIN__
+    if (jit_arg_reg_p(_jitc->function->call.argi)) {
+	jit_movr_f(_XMM0 - _jitc->function->call.argi, u);
+	if (_jitc->function->call.call & jit_call_varargs) {
+	    jit_stxi_f(_jitc->function->call.size, _RSP,
+		       _XMM0 - _jitc->function->call.argi);
+	    jit_ldxi_i(JIT_RA0 - _jitc->function->call.argi, _RSP,
+		       _jitc->function->call.size);
+	}
+	++_jitc->function->call.argi;
+	_jitc->function->call.size += sizeof(jit_word_t);
+    }
+#  else
+    if (jit_arg_f_reg_p(_jitc->function->self.argf)) {
 	jit_movr_f(_XMM0 - _jitc->function->call.argf, u);
 	++_jitc->function->call.argf;
     }
+#  endif
     else
 #endif
     {
@@ -666,10 +751,24 @@ _jit_pushargi_f(jit_state_t *_jit, jit_float32_t u)
 
     assert(_jitc->function);
 #if __WORDSIZE == 64
-    if (_jitc->function->call.argf < 8) {
+#  if __CYGWIN__
+    if (jit_arg_reg_p(_jitc->function->call.argi)) {
+	jit_movi_f(_XMM0 - _jitc->function->call.argi, u);
+	if (_jitc->function->call.call & jit_call_varargs) {
+	    jit_stxi_f(_jitc->function->call.size, _RSP,
+		       _XMM0 - _jitc->function->call.argi);
+	    jit_ldxi_i(JIT_RA0 - _jitc->function->call.argi, _RSP,
+		       _jitc->function->call.size);
+	}
+	++_jitc->function->call.argi;
+	_jitc->function->call.size += sizeof(jit_word_t);
+    }
+#  else
+    if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 	jit_movi_f(_XMM0 - _jitc->function->call.argf, u);
 	++_jitc->function->call.argf;
     }
+#  endif
     else
 #endif
     {
@@ -686,10 +785,24 @@ _jit_pushargr_d(jit_state_t *_jit, jit_int32_t u)
 {
     assert(_jitc->function);
 #if __WORDSIZE == 64
-    if (_jitc->function->call.argf < 8) {
+#  if __CYGWIN__
+    if (jit_arg_reg_p(_jitc->function->call.argi)) {
+	jit_movr_d(_XMM0 - _jitc->function->call.argi, u);
+	if (_jitc->function->call.call & jit_call_varargs) {
+	    jit_stxi_d(_jitc->function->call.size, _RSP,
+		       _XMM0 - _jitc->function->call.argi);
+	    jit_ldxi_l(JIT_RA0 - _jitc->function->call.argi, _RSP,
+		       _jitc->function->call.size);
+	}
+	++_jitc->function->call.argi;
+	_jitc->function->call.size += sizeof(jit_word_t);
+    }
+#  else
+    if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 	jit_movr_d(_XMM0 - _jitc->function->call.argf, u);
 	++_jitc->function->call.argf;
     }
+#  endif
     else
 #endif
     {
@@ -705,10 +818,24 @@ _jit_pushargi_d(jit_state_t *_jit, jit_float64_t u)
 
     assert(_jitc->function);
 #if __WORDSIZE == 64
-    if (_jitc->function->call.argf < 8) {
+#  if __CYGWIN__
+    if (jit_arg_reg_p(_jitc->function->call.argi)) {
+	jit_movi_d(_XMM0 - _jitc->function->call.argi, u);
+	if (_jitc->function->call.call & jit_call_varargs) {
+	    jit_stxi_d(_jitc->function->call.size, _RSP,
+		       _XMM0 - _jitc->function->call.argi);
+	    jit_ldxi_l(JIT_RA0 - _jitc->function->call.argi, _RSP,
+		       _jitc->function->call.size);
+	}
+	++_jitc->function->call.argi;
+	_jitc->function->call.size += sizeof(jit_word_t);
+    }
+#  else
+    if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 	jit_movi_d(_XMM0 - _jitc->function->call.argf, u);
 	++_jitc->function->call.argf;
     }
+#  endif
     else
 #endif
     {
@@ -729,7 +856,7 @@ _jit_regarg_p(jit_state_t *_jit, jit_node_t *node, jit_int32_t regno)
     spec = jit_class(_rvs[regno].spec);
     if (spec & jit_class_arg) {
 	if (spec & jit_class_gpr) {
-	    regno = _RDI - regno;
+	    regno = JIT_RA0 - regno;
 	    if (regno >= 0 && regno < node->v.w)
 		return (1);
 	}
@@ -754,6 +881,7 @@ _jit_finishr(jit_state_t *_jit, jit_int32_t r0)
     if (_jitc->function->self.alen < _jitc->function->call.size)
 	_jitc->function->self.alen = _jitc->function->call.size;
 #if __WORDSIZE == 64
+#  if !__CYGWIN__
     if (_jitc->function->call.call & jit_call_varargs) {
 	if (jit_regno(reg) == _RAX) {
 	    reg = jit_get_reg(jit_class_gpr);
@@ -766,6 +894,7 @@ _jit_finishr(jit_state_t *_jit, jit_int32_t r0)
 	if (reg != r0)
 	    jit_unget_reg(reg);
     }
+#  endif
 #endif
     call = jit_callr(reg);
     call->v.w = _jitc->function->call.argi;