Implement new, typed, jit_htonr* interfaces

* include/lightning.h: Split jit_htonr in the new 3 interfaces
	jit_htonr_us, jit_htonr_ui and jit_htonr_ul, the later only
	available on 64 bit. The plain/untyped jit_htonr macro call
	maps to the wordsize one.
	* lib/jit_aarch64-cpu.c,  lib/jit_aarch64-sz.c, lib/jit_aarch64.c,
	lib/jit_alpha-cpu.c, lib/jit_alpha-sz.c, lib/jit_alpha.c,
	lib/jit_arm-cpu.c, lib/jit_arm-sz.c, lib/jit_arm.c,
	lib/jit_hppa-cpu.c, lib/jit_hppa-sz.c, lib/jit_hppa.c,
	lib/jit_ia64-cpu.c, lib/jit_ia64-sz.c, lib/jit_ia64.c,
	lib/jit_mips-cpu.c, lib/jit_mips-sz.c, lib/jit_mips.c,
	lib/jit_ppc-cpu.c, lib/jit_ppc-sz.c, lib/jit_ppc.c,
	lib/jit_s390x-cpu.c, lib/jit_s390x-sz.c, lib/jit_s390x.c,
	lib/jit_sparc-cpu.c, lib/jit_sparc-sz.c, lib/jit_sparc.c,
	lib/jit_x86-cpu.c, lib/jit_x86-sz.c, lib/jit_x86.c:
	Update backends for the new jit_htonr*.
	* check/lightning.c, lib/jit_names.c, lib/lightning.c:
	Update for the new jit_htonr* interfaces.
	* check/Makefile.am: Update for new test cases.
	* check/hton.ok, check/hton.tst: New test cases.

ChangeLog

@@ -1,3 +1,25 @@
+2014-12-25 Paulo Andrade <pcpa@gnu.org>
+
+	* include/lightning.h: Split jit_htonr in the new 3 interfaces
+	jit_htonr_us, jit_htonr_ui and jit_htonr_ul, the later only
+	available on 64 bit. The plain/untyped jit_htonr macro call
+	maps to the wordsize one.
+	* lib/jit_aarch64-cpu.c,  lib/jit_aarch64-sz.c, lib/jit_aarch64.c,
+	lib/jit_alpha-cpu.c, lib/jit_alpha-sz.c, lib/jit_alpha.c,
+	lib/jit_arm-cpu.c, lib/jit_arm-sz.c, lib/jit_arm.c,
+	lib/jit_hppa-cpu.c, lib/jit_hppa-sz.c, lib/jit_hppa.c,
+	lib/jit_ia64-cpu.c, lib/jit_ia64-sz.c, lib/jit_ia64.c,
+	lib/jit_mips-cpu.c, lib/jit_mips-sz.c, lib/jit_mips.c,
+	lib/jit_ppc-cpu.c, lib/jit_ppc-sz.c, lib/jit_ppc.c,
+	lib/jit_s390x-cpu.c, lib/jit_s390x-sz.c, lib/jit_s390x.c,
+	lib/jit_sparc-cpu.c, lib/jit_sparc-sz.c, lib/jit_sparc.c,
+	lib/jit_x86-cpu.c, lib/jit_x86-sz.c, lib/jit_x86.c:
+	Update backends for the new jit_htonr*.
+	* check/lightning.c, lib/jit_names.c, lib/lightning.c:
+	Update for the new jit_htonr* interfaces.
+	* check/Makefile.am: Update for new test cases.
+	* check/hton.ok, check/hton.tst: New test cases.
+
 2014-12-24 Paulo Andrade <pcpa@gnu.org>
 
 	* include/lightning/jit_private.h, include/lightning/jit_x86.h,

check/Makefile.am

@@ -57,6 +57,7 @@ EXTRA_DIST =				\
 	ldstxr-c.tst	ldstxr-c.ok	\
 	ldstxi-c.tst	ldstxi-c.ok	\
 	cvt.tst		cvt.ok		\
+	hton.tst	hton.ok		\
 	branch.tst	branch.ok	\
 	alu.inc				\
 	alu_add.tst	alu_add.ok	\
@@ -104,7 +105,7 @@ base_TESTS =				\
 	ldstr ldsti			\
 	ldstxr ldstxi			\
 	ldstr-c ldstxr-c ldstxi-c	\
-	cvt branch			\
+	cvt hton branch			\
 	alu_add alux_add		\
 	alu_sub alux_sub alu_rsb	\
 	alu_mul alu_div alu_rem		\
@@ -177,7 +178,7 @@ arm_TESTS =					\
 	ldstr.arm ldsti.arm			\
 	ldstxr.arm ldstxi.arm			\
 	ldstr-c.arm ldstxr-c.arm ldstxi-c.arm	\
-	cvt.arm branch.arm			\
+	cvt.arm hton.arm branch.arm		\
 	alu_add.arm alux_add.arm		\
 	alu_sub.arm alux_sub.arm alu_rsb.arm	\
 	alu_mul.arm alu_div.arm alu_rem.arm	\
@@ -201,7 +202,7 @@ swf_TESTS =					\
 	ldstr.swf ldsti.swf			\
 	ldstxr.swf ldstxi.swf			\
 	ldstr-c.swf ldstxr-c.swf ldstxi-c.swf	\
-	cvt.swf branch.swf			\
+	cvt.swf hton.swf branch.swf		\
 	alu_add.swf alux_add.swf		\
 	alu_sub.swf alux_sub.swf alu_rsb.swf	\
 	alu_mul.swf alu_div.swf alu_rem.swf	\
@@ -223,7 +224,7 @@ arm_swf_TESTS =							\
 	ldstr.arm.swf ldsti.arm.swf				\
 	ldstxr.arm.swf ldstxi.arm.swf				\
 	ldstr-c.arm.swf ldstxr-c.arm.swf ldstxi-c.arm.swf	\
-	cvt.arm.swf branch.arm.swf				\
+	cvt.arm.swf hton.arm.swf branch.arm.swf			\
 	alu_add.arm.swf alux_add.arm.swf			\
 	alu_sub.arm.swf alux_sub.arm.swf alu_rsb.arm.swf	\
 	alu_mul.arm.swf alu_div.arm.swf alu_rem.arm.swf		\
@@ -246,7 +247,7 @@ arm4_swf_TESTS =						\
 	ldstr.arm4.swf ldsti.arm4.swf				\
 	ldstxr.arm4.swf ldstxi.arm4.swf				\
 	ldstr-c.arm4.swf ldstxr-c.arm4.swf ldstxi-c.arm4.swf	\
-	cvt.arm4.swf branch.arm4.swf				\
+	cvt.arm4.swf hton.arm4.swf branch.arm4.swf		\
 	alu_add.arm4.swf alux_add.arm4.swf			\
 	alu_sub.arm4.swf alux_sub.arm4.swf alu_rsb.arm4.swf	\
 	alu_mul.arm4.swf alu_div.arm4.swf alu_rem.arm4.swf	\

check/hton.ok

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

check/hton.tst

@@ -0,0 +1,169 @@
+.data	16
+ok:
+.c	"ok\n"
+
+#define us12_i		0x1234
+#define us7f_i		0x7ff7
+#define us80_i		0x8008
+#define usff_i		0xffff
+#define ui12_i		0x01234567
+#define ui7f_i		0x7f7ff7f7
+#define ui80_i		0x80800808
+#define uiff_i		0xffffffff
+#define ul12_i		0x0123456789abcdef
+#define ul7f_i		0x7f7f7f7ff7f7f7f7
+#define ul80_i		0x8080808008080808
+#define ulff_i		0xffffffffffffffff
+
+#if __WORDSIZE == 32
+#  define xus12_i	0xffff1234
+#  define xus7f_i	0x10107ff7
+#  define xus80_i	0x81188008
+#  define xusff_i	0xeaaeffff
+#else
+#  define xus12_i	0xffffffffffff1234
+#  define xus7f_i	0x1010100101017ff7
+#  define xus80_i	0x8181811818818008
+#  define xusff_i	0xeaeaeaaeaeaeffff
+#  define xui12_i	0xffffffff01234567
+#  define xui7f_i	0x101001017f7ff7f7
+#  define xui80_i	0x8181181880800808
+#  define xuiff_i	0xeaeaaeaeffffffff
+#endif
+
+#if __BYTE_ORDER == __LITTLE_ENDIAN
+#  define us12_o	0x3412
+#  define us7f_o	0xf77f
+#  define us80_o	0x0880
+#  define usff_o	0xffff
+#  define ui12_o	0x67452301
+#  define ui7f_o	0xf7f77f7f
+#  define ui80_o	0x08088080
+#  define uiff_o	0xffffffff
+#  define ul12_o	0xefcdab8967452301
+#  define ul7f_o	0xf7f7f7f77f7f7f7f
+#  define ul80_o	0x0808080880808080
+#  define ulff_o	0xffffffffffffffff
+#else
+#  define us12_o	us12_i
+#  define us7f_o	us7f_i
+#  define us80_o	us80_i
+#  define usff_o	usff_i
+#  define ui12_o	ui12_i
+#  define ui7f_o	ui7f_i
+#  define ui80_o	ui80_i
+#  define uiff_o	uiff_i
+#  define ul12_o	ul12_i
+#  define ul7f_o	ul7f_i
+#  define ul80_o	ul80_i
+#  define ulff_o	ulff_i
+#endif
+
+#define HTON4(I, O, T, R0, R1)				\
+	movi %R0 I					\
+	htonr_##T %R1 %R0				\
+	beqi T##R0##R1##I %R1 O				\
+	calli @abort					\
+T##R0##R1##I:
+
+#define HTON3(T, R0, R1)				\
+	HTON4(T##12_i, T##12_o, T, R0, R1)		\
+	HTON4(x##T##12_i, T##12_o, T, R0, R1)		\
+	HTON4(T##7f_i, T##7f_o, T, R0, R1)		\
+	HTON4(x##T##7f_i, T##7f_o, T, R0, R1)		\
+	HTON4(T##80_i, T##80_o, T, R0, R1)		\
+	HTON4(x##T##80_i, T##80_o, T, R0, R1)		\
+	HTON4(T##ff_i, T##ff_o, T, R0, R1)		\
+	HTON4(x##T##ff_i, T##ff_o, T, R0, R1)
+
+#define HTON3x(T, R0, R1)				\
+	HTON4(T##12_i, T##12_o, T, R0, R1)		\
+	HTON4(T##7f_i, T##7f_o, T, R0, R1)		\
+	HTON4(T##80_i, T##80_o, T, R0, R1)		\
+	HTON4(T##ff_i, T##ff_o, T, R0, R1)
+
+#define HTON2(T, V0, V1, V2, R0, R1, R2)		\
+	HTON3(T, V0, V0)				\
+	HTON3(T, V0, V1)				\
+	HTON3(T, V0, V2)				\
+	HTON3(T, V0, R0)				\
+	HTON3(T, V0, R1)				\
+	HTON3(T, V0, R2)				\
+
+#define HTON2x(T, V0, V1, V2, R0, R1, R2)		\
+	HTON3x(T, V0, V0)				\
+	HTON3x(T, V0, V1)				\
+	HTON3x(T, V0, V2)				\
+	HTON3x(T, V0, R0)				\
+	HTON3x(T, V0, R1)				\
+	HTON3x(T, V0, R2)				\
+
+#define HTON1(T, V0, V1, V2, R0, R1, R2)		\
+	HTON2(T, V0, V1, V2, R0, R1, R2)		\
+	HTON2(T, V1, V2, R0, R1, R2, V0)		\
+	HTON2(T, V2, R0, R1, R2, V0, V1)		\
+	HTON2(T, R0, R1, R2, V0, V1, V2)		\
+	HTON2(T, R1, R2, V0, V1, V2, R0)		\
+	HTON2(T, R2, V0, V1, V2, R0, R1)
+
+#define HTON1x(T, V0, V1, V2, R0, R1, R2)		\
+	HTON2x(T, V0, V1, V2, R0, R1, R2)		\
+	HTON2x(T, V1, V2, R0, R1, R2, V0)		\
+	HTON2x(T, V2, R0, R1, R2, V0, V1)		\
+	HTON2x(T, R0, R1, R2, V0, V1, V2)		\
+	HTON2x(T, R1, R2, V0, V1, V2, R0)		\
+	HTON2x(T, R2, V0, V1, V2, R0, R1)
+
+#if __WORDSIZE == 32
+#  define HTON(V0, V1, V2, R0, R1, R2)			\
+	HTON1(us, V0, V1, V2, R0, R1, R2)		\
+	HTON1x(ui, V0, V1, V2, R0, R1, R2)
+#else
+#  define HTON(V0, V1, V2, R0, R1, R2)			\
+	HTON1(us, V0, V1, V2, R0, R1, R2)		\
+	HTON1(ui, V0, V1, V2, R0, R1, R2)		\
+	HTON1x(ul, V0, V1, V2, R0, R1, R2)
+#endif
+
+.code
+	prolog
+	/* simple sequence for easier disassembly reading and encoding check */
+	movi %r0 us12_i
+	htonr_us %r1 %r0
+	beqi us %r1 us12_o
+	calli @abort
+us:
+
+	movi %r0 xus12_i
+	htonr_us %r1 %r0
+	beqi xus %r1 us12_o
+	calli @abort
+xus:
+	movi %r0 ui12_i
+	htonr_ui %r1 %r0
+	beqi ui %r1 ui12_o
+	calli @abort
+ui:
+#if __WORDSIZE == 64
+	movi %r0 xui12_i
+	htonr_ui %r1 %r0
+	beqi xui %r1 ui12_o
+	calli @abort
+xui:
+	movi %r0 ul12_i
+	htonr_ul %r1 %r0
+	beqi ul %r1 ul12_o
+	calli @abort
+ul:
+#endif
+
+	HTON(v0, v1, v2, r0, r1, r2)
+
+	// just to know did not abort
+	prepare
+		pushargi ok
+		ellipsis
+	finishi @printf
+
+	ret
+	epilog

check/lightning.c

@@ -319,6 +319,11 @@ static void extr_s(void);	static void extr_us(void);
 #if __WORDSIZE == 64
 static void extr_i(void);	static void extr_ui(void);
 #endif
+static void htonr_us(void);	static void ntohr_us(void);
+static void htonr_ui(void);	static void ntohr_ui(void);
+#if __WORDSIZE == 64
+static void htonr_ul(void);	static void ntohr_ul(void);
+#endif
 static void htonr(void);	static void ntohr(void);
 static void ldr_c(void);	static void ldi_c(void);
 static void ldr_uc(void);	static void ldi_uc(void);
@@ -621,6 +626,11 @@ static instr_t		  instr_vector[] = {
     entry(extr_s),	entry(extr_us),
 #if __WORDSIZE == 64
     entry(extr_i),	entry(extr_ui),
+#endif
+    entry(htonr_us),	entry(ntohr_us),
+    entry(htonr_ui),	entry(ntohr_ui),
+#if __WORDSIZE == 64
+    entry(htonr_ul),	entry(ntohr_ul),
 #endif
     entry(htonr),	entry(ntohr),
     entry(ldr_c),	entry(ldi_c),
@@ -1437,6 +1447,11 @@ entry_ir_ir(extr_s)		entry_ir_ir(extr_us)
 #if __WORDSIZE == 64
 entry_ir_ir(extr_i)		entry_ir_ir(extr_ui)
 #endif
+entry_ir_ir(htonr_us)		entry_ir_ir(ntohr_us)
+entry_ir_ir(htonr_ui)		entry_ir_ir(ntohr_ui)
+#if __WORDSIZE == 64
+entry_ir_ir(htonr_ul)		entry_ir_ir(ntohr_ul)
+#endif
 entry_ir_ir(htonr)		entry_ir_ir(ntohr)
 entry_ir_ir(ldr_c)		entry_ir_pm(ldi_c)
 entry_ir_ir(ldr_uc)		entry_ir_pm(ldi_uc)

include/lightning.h

@@ -327,9 +327,22 @@ typedef enum {
 #  define jit_extr_ui(u,v)	jit_new_node_ww(jit_code_extr_ui,u,v)
 #endif
     jit_code_extr_i,		jit_code_extr_ui,
-#define jit_htonr(u,v)		jit_new_node_ww(jit_code_htonr,u,v)
+
-#define jit_ntohr(u,v)		jit_new_node_ww(jit_code_htonr,u,v)
+#define jit_htonr_us(u,v)	jit_new_node_ww(jit_code_htonr_us,u,v)
-    jit_code_htonr,
+#define jit_ntohr_us(u,v)	jit_new_node_ww(jit_code_htonr_us,u,v)
+    jit_code_htonr_us,
+#define jit_htonr_ui(u,v)	jit_new_node_ww(jit_code_htonr_ui,u,v)
+#define jit_ntohr_ui(u,v)	jit_new_node_ww(jit_code_htonr_ui,u,v)
+#if __WORDSIZE == 32
+#  define jit_htonr(u,v)	jit_new_node_ww(jit_code_htonr_ui,u,v)
+#  define jit_ntohr(u,v)	jit_new_node_ww(jit_code_htonr_ui,u,v)
+#else
+#define jit_htonr_ul(u,v)	jit_new_node_ww(jit_code_htonr_ul,u,v)
+#define jit_ntohr_ul(u,v)	jit_new_node_ww(jit_code_htonr_ul,u,v)
+#  define jit_htonr(u,v)	jit_new_node_ww(jit_code_htonr_ul,u,v)
+#  define jit_ntohr(u,v)	jit_new_node_ww(jit_code_htonr_ul,u,v)
+#endif
+    jit_code_htonr_ui,		jit_code_htonr_ul,
 
 #define jit_ldr_c(u,v)		jit_new_node_ww(jit_code_ldr_c,u,v)
 #define jit_ldi_c(u,v)		jit_new_node_wp(jit_code_ldi_c,u,v)

lib/jit_aarch64-cpu.c

@@ -656,9 +656,15 @@ static void _stxi_i(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
 #  define stxi_l(i0,r0,r1)		_stxi_l(_jit,i0,r0,r1)
 static void _stxi_l(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
 #  if __BYTE_ORDER == __LITTLE_ENDIAN
-#    define htonr(r0,r1)		REV(r0,r1)
+#  define htonr_us(r0,r1)		_htonr_us(_jit,r0,r1)
+static void _htonr_us(jit_state_t*,jit_int32_t,jit_int32_t);
+#  define htonr_ui(r0,r1)		_htonr_ui(_jit,r0,r1)
+static void _htonr_ui(jit_state_t*,jit_int32_t,jit_int32_t);
+#    define htonr_ul(r0,r1)		REV(r0,r1)
 #  else
-#    define htonr(r0,r1)		movr(r0,r1)
+#    define htonr_us(r0,r1)		extr_us(r0,r1)
+#    define htonr_ui(r0,r1)		extr_ui(r0,r1)
+#    define htonr_ul(r0,r1)		movr(r0,r1)
 #  endif
 #  define extr_c(r0,r1)			SXTB(r0,r1)
 #  define extr_uc(r0,r1)		UXTB(r0,r1)
@@ -1429,6 +1435,22 @@ _xori(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     }
 }
 
+#if __BYTE_ORDER == __LITTLE_ENDIAN
+static void
+_htonr_us(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    htonr_ul(r0, r1);
+    rshi_u(r0, r0, 48);
+}
+
+static void
+_htonr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    htonr_ul(r0, r1);
+    rshi_u(r0, r0, 32);
+}
+#endif
+
 static void
 _ldi_c(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0)
 {

lib/jit_aarch64-sz.c

@@ -84,7 +84,9 @@
     4,	/* extr_us */
     4,	/* extr_i */
     4,	/* extr_ui */
-    4,	/* htonr */
+    8,	/* htonr_us */
+    8,	/* htonr_ui */
+    4,	/* htonr_ul */
     4,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */

lib/jit_aarch64.c

@@ -825,7 +825,9 @@ _emit_code(jit_state_t *_jit)
 		case_wrr(stx, _i);
 		case_rrr(stx, _l);
 		case_wrr(stx, _l);
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
+		case_rr(hton, _ul);
 		case_rr(ext, _c);
 		case_rr(ext, _uc);
 		case_rr(ext, _s);

lib/jit_alpha-cpu.c

@@ -622,10 +622,16 @@ static void _extr_i(jit_state_t*,jit_int32_t,jit_int32_t);
 #  define extr_ui(r0,r1)		_extr_ui(_jit,r0,r1)
 static void _extr_ui(jit_state_t*,jit_int32_t,jit_int32_t);
 #  if __BYTE_ORDER == __LITTLE_ENDIAN
-#    define htonr(r0,r1)		_htonr(_jit,r0,r1)
+#    define htonr_us(r0,r1)		_htonr_us(_jit,r0,r1)
-static void _htonr(jit_state_t*,jit_int32_t,jit_int32_t);
+static void _htonr_us(jit_state_t*,jit_int32_t,jit_int32_t);
+#    define htonr_ui(r0,r1)		_htonr_ui(_jit,r0,r1)
+static void _htonr_ui(jit_state_t*,jit_int32_t,jit_int32_t);
+#    define htonr_ul(r0,r1)		_htonr_ul(_jit,r0,r1)
+static void _htonr_ul(jit_state_t*,jit_int32_t,jit_int32_t);
 #  else
-#    define htonr(r0,r1)		movr(r0,r1)
+#    define htonr_us(r0,r1)		extr_us(r0,r1)
+#    define htonr_ui(r0,r1)		extr_ui(r0,r1)
+#    define htonr_ul(r0,r1)		movr(r0,r1)
 #  endif
 #  define jmpr(r0)			JMP(_R31_REGNO,r0,0)
 #  define jmpi(i0)			_jmpi(_jit,i0)
@@ -2442,7 +2448,45 @@ _extr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 }
 
 static void
-_htonr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+_htonr_us(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    jit_int32_t		t0;
+    t0 = jit_get_reg(jit_class_gpr);
+    EXTBLi(r1, 0, rn(t0));
+    EXTBLi(r1, 1, r0);
+    SLLi(rn(t0), 8, rn(t0));
+    OR(r0, rn(t0), r0);
+    jit_unget_reg(t0);
+}
+
+static void
+_htonr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    jit_int32_t		t0;
+    jit_int32_t		t1;
+    jit_int32_t		t2;
+    jit_int32_t		t3;
+    t0 = jit_get_reg(jit_class_gpr);
+    t1 = jit_get_reg(jit_class_gpr);
+    t2 = jit_get_reg(jit_class_gpr);
+    t3 = jit_get_reg(jit_class_gpr);
+    EXTBLi(r1, 3, rn(t0));
+    INSBLi(r1, 3, rn(t1));
+    SLLi(r1, 8, rn(t2));
+    ZAPNOTi(rn(t2), 4, rn(t2));
+    SRLi(r1, 8, rn(t3));
+    OR(rn(t0), rn(t1), r0);
+    OR(rn(t2), r0, r0);
+    ZAPNOTi(rn(t3), 2, rn(t3));
+    OR(rn(t3), r0, r0);
+    jit_unget_reg(t3);
+    jit_unget_reg(t2);
+    jit_unget_reg(t1);
+    jit_unget_reg(t0);
+}
+
+static void
+_htonr_ul(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
     jit_int32_t		t0;
     jit_int32_t		t1;

lib/jit_alpha-sz.c

@@ -84,7 +84,9 @@
     8,	/* extr_us */
     8,	/* extr_i */
     8,	/* extr_ui */
-    36,	/* htonr */
+    16,	/* htonr_us */
+    36,	/* htonr_ui */
+    36,	/* htonr_ul */
     12,	/* ldr_c */
     40,	/* ldi_c */
     4,	/* ldr_uc */

lib/jit_alpha.c

@@ -850,7 +850,9 @@ _emit_code(jit_state_t *_jit)
 		case_wrr(stx, _i);
 		case_rrr(stx, _l);
 		case_wrr(stx, _l);
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
+		case_rr(hton, _ul);
 		case_rr(ext, _c);
 		case_rr(ext, _uc);
 		case_rr(ext, _s);

lib/jit_arm-cpu.c

@@ -1089,9 +1089,12 @@ static void _stxr_i(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
 #  define stxi_i(r0,r1,i0)		_stxi_i(_jit,r0,r1,i0)
 static void _stxi_i(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
 #  if __BYTE_ORDER == __LITTLE_ENDIAN
-#  define htonr(r0,r1)			_htonr(_jit,r0,r1)
+#  define htonr_us(r0,r1)		_htonr_us(_jit,r0,r1)
-static void _htonr(jit_state_t*,jit_int32_t,jit_int32_t);
+static void _htonr_us(jit_state_t*,jit_int32_t,jit_int32_t);
+#  define htonr_ui(r0,r1)		_htonr_ui(_jit,r0,r1)
+static void _htonr_ui(jit_state_t*,jit_int32_t,jit_int32_t);
 #  else
+#    define htonr_us(r0,r1)		extr_us(r0,r1)
 #    define htonr(r0,r1)		movr(r0,r1)
 #  endif
 #  define extr_c(r0,r1)			_extr_c(_jit,r0,r1)
@@ -3556,9 +3559,37 @@ _stxi_i(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
 }
 
 #  if __BYTE_ORDER == __LITTLE_ENDIAN
+static void
+_htonr_us(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    jit_int32_t		t0;
+    if (jit_thumb_p()) {
+	if ((r0|r1) < 8)
+	    T1_REV(r0, r1);
+	else
+	    T2_REV(r0, r1);
+	rshi_u(r0, r0, 16);
+    }
+    else {
+	if (jit_armv6_p()) {
+	    REV(r0, r1);
+	    rshi_u(r0, r0, 16);
+	}
+	else {
+	    t0 = jit_get_reg(jit_class_gpr);
+	    rshi(rn(t0), r1, 8);
+	    andi(r0, r1, 0xff);
+	    andi(rn(t0), rn(t0), 0xff);
+	    lshi(r0, r0, 8);
+	    orr(r0, r0, rn(t0));
+	    jit_unget_reg(t0);
+	}
+    }
+}
+
 /* inline glibc htonl (without register clobber) */
 static void
-_htonr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+_htonr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
     jit_int32_t		reg;
     if (jit_thumb_p()) {

lib/jit_arm-sz.c

@@ -85,7 +85,9 @@
     4,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    4,	/* htonr */
+    8,	/* htonr_us */
+    4,	/* htonr_ui */
+    0,	/* htonr_ul */
     4,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */
@@ -441,7 +443,9 @@
     4,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    4,	/* htonr */
+    20,	/* htonr_us */
+    16,	/* htonr_ui */
+    0,	/* htonr_ul */
     4,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */

lib/jit_arm.c

@@ -1141,7 +1141,8 @@ _emit_code(jit_state_t *_jit)
 		case_wrr(stx, _s);
 		case_rrr(stx, _i);
 		case_wrr(stx, _i);
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
 		case_rr(ext, _c);
 		case_rr(ext, _uc);
 		case_rr(ext, _s);

lib/jit_hppa-cpu.c

@@ -655,7 +655,8 @@ static jit_word_t _movi_p(jit_state_t*,jit_int32_t,jit_word_t);
 #define extr_s(r0,r1)		EXTRWR(r1,31,16,r0)
 #define extr_us(r0,r1)		EXTRWR_U(r1,31,16,r0)
 #if __BYTE_ORDER == __BIG_ENDIAN
-#  define htonr(r0,r1)		movr(r0,r1)
+#  define htonr_us(r0,r1)	extr_us(r0,r1)
+#  define htonr_ui(r0,r1)	movr(r0,r1)
 #else
 #  error need htonr implementation
 #endif

lib/jit_hppa-sz.c

@@ -84,7 +84,9 @@
     4,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    4,	/* htonr */
+    4,	/* htonr_us */
+    4,	/* htonr_ui */
+    0,	/* htonr_l */
     8,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */

lib/jit_hppa.c

@@ -804,7 +804,8 @@ _emit_code(jit_state_t *_jit)
 		case_rr(ext, _uc);
 		case_rr(ext, _s);
 		case_rr(ext, _us);
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
 		case_rrr(lt,);
 		case_rrw(lt,);
 		case_rrr(lt, _u);

lib/jit_ia64-cpu.c

@@ -1308,9 +1308,15 @@ static void _movi(jit_state_t*,jit_int32_t,jit_word_t);
 #define movi_p(r0,i0)			_movi_p(_jit,r0,i0)
 static jit_word_t _movi_p(jit_state_t*,jit_int32_t,jit_word_t);
 #if __BYTE_ORDER == __LITTLE_ENDIAN
-#  define htonr(r0,r1)			MUX1(r0,r1,MUX_REV)
+#  define htonr_us(r0,r1)		_htonr_us(_jit,r0,r1)
+static void _htonr_us(jit_state_t*,jit_int32_t,jit_int32_t);
+#  define htonr_ui(r0,r1)		_htonr_ui(_jit,r0,r1)
+static void _htonr_ui(jit_state_t*,jit_int32_t,jit_int32_t);
+#  define htonr_ul(r0,r1)		MUX1(r0,r1,MUX_REV)
 #else
-#  define htonr(r0,r1)			movr(r0,r1)
+#  define htonr_us(r0,r1)		extr_us(r0,r1)
+#  define htonr_ui(r0,r1)		extr_ui(r0,r1)
+#  define htonr_ul(r0,r1)		movr(r0,r1)
 #endif
 #define extr_c(r0,r1)			SXT1(r0,r1)
 #define extr_uc(r0,r1)			ZXT1(r0,r1)
@@ -3939,6 +3945,48 @@ _xori(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     }
 }
 
+#if __BYTE_ORDER == __LITTLE_ENDIAN
+static void
+_htonr_us(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    jit_int32_t		t0;
+    t0 = jit_get_reg(jit_class_gpr);
+    rshi(rn(t0), r1, 8);
+    andi(r0, r1, 0xff);
+    andi(rn(t0), rn(t0), 0xff);
+    lshi(r0, r0, 8);
+    orr(r0, r0, rn(t0));
+    jit_unget_reg(t0);
+}
+
+static void
+_htonr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    jit_int32_t		t0;
+    jit_int32_t		t1;
+    jit_int32_t		t2;
+    t0 = jit_get_reg(jit_class_gpr);
+    t1 = jit_get_reg(jit_class_gpr);
+    t2 = jit_get_reg(jit_class_gpr);
+    rshi(rn(t0), r1, 24);
+    rshi(rn(t1), r1, 16);
+    rshi(rn(t2), r1,  8);
+    andi(rn(t0), rn(t0), 0xff);
+    andi(rn(t1), rn(t1), 0xff);
+    andi(rn(t2), rn(t2), 0xff);
+    andi(r0, r1, 0xff);
+    lshi(r0, r0, 24);
+    lshi(rn(t1), rn(t1), 8);
+    orr(r0, r0, rn(t0));
+    lshi(rn(t2), rn(t2), 16);
+    orr(r0, r0, rn(t1));
+    orr(r0, r0, rn(t2));
+    jit_unget_reg(t2);
+    jit_unget_reg(t1);
+    jit_unget_reg(t0);
+}
+#endif
+
 static void
 _lshi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {

lib/jit_ia64-sz.c

@@ -84,7 +84,9 @@
     16,	/* extr_us */
     16,	/* extr_i */
     16,	/* extr_ui */
-    0,	/* htonr */
+    16,	/* htonr_us */
+    16,	/* htonr_ui */
+    16,	/* htonr_ul */
     16,	/* ldr_c */
     32,	/* ldi_c */
     0,	/* ldr_uc */

lib/jit_ia64.c

@@ -945,7 +945,9 @@ _emit_code(jit_state_t *_jit)
 		else
 		    movi(rn(node->u.w), node->v.w);
 		break;
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
+		case_rr(hton, _ul);
 		case_rr(ext, _c);
 		case_rr(ext, _uc);
 		case_rr(ext, _s);

lib/jit_mips-cpu.c

@@ -579,10 +579,22 @@ static void _stxr_l(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 static void _stxi_l(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
 #  endif
 #  if __BYTE_ORDER == __LITTLE_ENDIAN
-#  define htonr(r0,r1)			_htonr(_jit,r0,r1)
+#    define htonr_us(r0,r1)		_htonr_us(_jit,r0,r1)
-static void _htonr(jit_state_t*,jit_int32_t,jit_int32_t);
+static void _htonr_us(jit_state_t*,jit_int32_t,jit_int32_t);
+#    define htonr_ui(r0,r1)		_htonr_ui(_jit,r0,r1)
+static void _htonr_ui(jit_state_t*,jit_int32_t,jit_int32_t);
+#    if __WORDSIZE == 64
+#      define htonr_ul(r0,r1)		_htonr_ul(_jit,r0,r1)
+static void _htonr_ul(jit_state_t*,jit_int32_t,jit_int32_t);
+#    endif
 #  else
-#    define htonr(r0,r1)		movr(r0,r1)
+#    define htonr_us(r0,r1)		extr_us(r0,r1)
+#    if __WORDSIZE == 32
+#      define htonr_ui(r0,r1)		movr(r0,r1)
+#    else
+#      define htonr_ui(r0,r1)		extr_ui(r0,r1)
+#      define htonr_ul(r0,r1)		movr(r0,r1)
+#    endif
 #  endif
 #  define extr_c(r0,r1)			_extr_c(_jit,r0,r1)
 static void _extr_c(jit_state_t*,jit_int32_t,jit_int32_t);
@@ -591,8 +603,8 @@ static void _extr_c(jit_state_t*,jit_int32_t,jit_int32_t);
 static void _extr_s(jit_state_t*,jit_int32_t,jit_int32_t);
 #  define extr_us(r0,r1)		ANDI(r0,r1,0xffff)
 #  if __WORDSIZE == 64
-#  define extr_i(r0,r1)			SLL(r0,r1,0)
+#    define extr_i(r0,r1)		SLL(r0,r1,0)
-#  define extr_ui(r0,r1)		_extr_ui(_jit,r0,r1)
+#    define extr_ui(r0,r1)		_extr_ui(_jit,r0,r1)
 static void _extr_ui(jit_state_t*,jit_int32_t,jit_int32_t);
 #  endif
 #  define ltr(r0,r1,r2)			SLT(r0,r1,r2)
@@ -1657,40 +1669,57 @@ _stxi_l(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
 #endif
 
 #  if __BYTE_ORDER == __LITTLE_ENDIAN
+static void
+_htonr_us(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    jit_int32_t		t0;
+    t0 = jit_get_reg(jit_class_gpr);
+    rshi(rn(t0), r1, 8);
+    andi(r0, r1, 0xff);
+    andi(rn(t0), rn(t0), 0xff);
+    lshi(r0, r0, 8);
+    orr(r0, r0, rn(t0));
+    jit_unget_reg(t0);
+}
+
+static void
+_htonr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    jit_int32_t		t0;
+    jit_int32_t		t1;
+    jit_int32_t		t2;
+    t0 = jit_get_reg(jit_class_gpr);
+    t1 = jit_get_reg(jit_class_gpr);
+    t2 = jit_get_reg(jit_class_gpr);
+    rshi(rn(t0), r1, 24);
+    rshi(rn(t1), r1, 16);
+    rshi(rn(t2), r1,  8);
+    andi(rn(t0), rn(t0), 0xff);
+    andi(rn(t1), rn(t1), 0xff);
+    andi(rn(t2), rn(t2), 0xff);
+    andi(r0, r1, 0xff);
+    lshi(r0, r0, 24);
+    lshi(rn(t1), rn(t1), 8);
+    orr(r0, r0, rn(t0));
+    lshi(rn(t2), rn(t2), 16);
+    orr(r0, r0, rn(t1));
+    orr(r0, r0, rn(t2));
+    jit_unget_reg(t2);
+    jit_unget_reg(t1);
+    jit_unget_reg(t0);
+}
+
 static void
 _htonr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
-    jit_int32_t		rg0;
+    jit_int32_t		reg;
-    jit_int32_t		rg1;
+    reg = jit_get_reg(jit_class_gpr);
-    if (jit_mips2_p()) {
+    rshi_u(rn(reg), r1, 32);
-	WSBH(r0, r1);
+    htonr_ui(r0, r1);
-	ROTR(r0, r0, 16);
+    htonr_ui(rn(reg), rn(reg));
-    }
+    lshi(r0, r0, 32);
-    else {
+    orr(r0, r0, rn(reg));
-	/* FIXME rewrite in a more sane way, but unlikely to be used
+    jit_unget_reg(reg);
-	 * in near time... */
-#  if __WORDSIZE == 64
-#    error htonr only implemented for 64 bit
-#  endif
-	rg0 = jit_get_reg(jit_class_gpr);
-	rg1 = jit_get_reg(jit_class_gpr);
-	LUI(rn(rg0), 0xff00);
-	ORI(rn(rg0), rn(rg0), 0xff00);
-	AND(rn(rg1), r1, rn(rg0));
-	SRL(rn(rg0), rn(rg0), 8);
-	AND(rn(rg0), r1, rn(rg0));
-	SRL(rn(rg1), rn(rg1), 8);
-	SLL(rn(rg0), rn(rg0), 8);
-	OR(r0, rn(rg0), rn(rg1));
-	ANDI(rn(rg1), r0, 0xffff);
-	LUI(rn(rg0), 0xffff);
-	AND(rn(rg0), r0, rn(rg0));
-	SLL(rn(rg1), rn(rg1), 16);
-	SRL(rn(rg0), rn(rg0), 16);
-	OR(r0, rn(rg0), rn(rg1));
-	jit_unget_reg(rg0);
-	jit_unget_reg(rg1);
-    }
 }
 #  endif
 

lib/jit_mips-sz.c

@@ -85,7 +85,9 @@
     4,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    4,	/* htonr */
+    4,	/* htonr_us */
+    4,	/* htonr_ui */
+    0,	/* htonr_ul */
     4,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */
@@ -441,7 +443,9 @@
     4,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    56,	/* htonr */
+    20,	/* htonr_us */
+    52,	/* htonr_ui */
+    0,	/* htonr_ul */
     4,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */
@@ -796,7 +800,9 @@
     4,	/* extr_us */
     4,	/* extr_i */
     8,	/* extr_ui */
-    4,	/* htonr */
+    4,	/* htonr_us */
+    4,	/* htonr_ui */
+    4,	/* htonr_ul */
     4,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */

lib/jit_mips.c

@@ -1094,7 +1094,11 @@ _emit_code(jit_state_t *_jit)
 		case_rrr(stx, _l);
 		case_wrr(stx, _l);
 #endif
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
+#if __WORDSIZE == 64
+		case_rr(hton, _ul);
+#endif
 		case_rr(ext, _c);
 		case_rr(ext, _uc);
 		case_rr(ext, _s);

lib/jit_names.c

@@ -62,7 +62,8 @@ static char *code_name[] = {
     "extr_c",		"extr_uc",
     "extr_s",		"extr_us",
     "extr_i",		"extr_ui",
-    "htonr",
+    "htonr_us",
+    "htonr_ui",		"htonr_ul",
     "ldr_c",		"ldi_c",
     "ldr_uc",		"ldi_uc",
     "ldr_s",		"ldi_s",

lib/jit_ppc-cpu.c

@@ -471,10 +471,22 @@ static jit_word_t _movi_p(jit_state_t*,jit_int32_t,jit_word_t);
 #    define extr_ui(r0,r1)		CLRLDI(r0,r1,32)
 #  endif
 #  if __BYTE_ORDER == __BIG_ENDIAN
-#    define htonr(r0,r1)		movr(r0,r1)
+#    define htonr_us(r0,r1)		extr_us(r0,r1)
+#    if __WORDSIZE == 32
+#      define htonr_ui(r0,r1)		movr(r0,r1)
+#    else
+#      define htonr_ui(r0,r1)		extr_ui(r0,r1)
+#      define htonr_ul(r0,r1)		movr(r0,r1)
+#    endif
 #  else
-#    define htonr(r0,r1)		_htonr(_jit,r0,r1)
+#    define htonr_us(r0,r1)		_htonr_us(_jit,r0,r1)
-static void _htonr(jit_state_t*,jit_int32_t,jit_int32_t);
+static void _htonr_us(jit_state_t*,jit_int32_t,jit_int32_t);
+#    define htonr_ui(r0,r1)		_htonr_ui(_jit,r0,r1)
+static void _htonr_ui(jit_state_t*,jit_int32_t,jit_int32_t);
+#    if __WORDSIZE == 64
+#      define htonr_ul(r0,r1)		_htonr_ul(_jit,r0,r1)
+static void _htonr_ul(jit_state_t*,jit_int32_t,jit_int32_t);
+#    endif
 #  endif
 #  define addr(r0,r1,r2)		ADD(r0,r1,r2)
 #  define addi(r0,r1,i0)		_addi(_jit,r0,r1,i0)
@@ -1048,31 +1060,44 @@ _movi_p(jit_state_t *_jit, jit_int32_t r0, jit_word_t i0)
 
 #  if __BYTE_ORDER == __LITTLE_ENDIAN
 static void
-_htonr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+_htonr_us(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    jit_int32_t		t0;
+    t0 = jit_get_reg(jit_class_gpr);
+    rshi(rn(t0), r1, 8);
+    andi(r0, r1, 0xff);
+    andi(rn(t0), rn(t0), 0xff);
+    lshi(r0, r0, 8);
+    orr(r0, r0, rn(t0));
+    jit_unget_reg(t0);
+}
+
+static void
+_htonr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
     jit_int32_t		reg;
-#    if __WORDSIZE == 64
-    jit_int32_t		top;
-    top = jit_get_reg(jit_class_gpr);
-    rshi_u(rn(top), r1, 32);
-#    endif
     reg = jit_get_reg(jit_class_gpr);
     ROTLWI(rn(reg), r1, 8);
     RLWIMI(rn(reg), r1, 24, 0, 7);
     RLWIMI(rn(reg), r1, 24, 16, 23);
-#    if __WORDSIZE == 32
     CLRLDI(r0, rn(reg), 32);
-#    else
-    lshi(r0, rn(reg), 32);
-    ROTLWI(rn(reg), rn(top), 8);
-    RLWIMI(rn(reg), rn(top), 24, 0, 7);
-    RLWIMI(rn(reg), rn(top), 24, 16, 23);
-    orr(r0, r0, rn(reg));
-    movr(r0, rn(top));
-    jit_unget_reg(top);
-#    endif
     jit_unget_reg(reg);
 }
+
+#    if __WORDSIZE == 64
+static void
+_htonr_ul(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    jit_int32_t		reg;
+    reg = jit_get_reg(jit_class_gpr);
+    rshi_u(rn(reg), r1, 32);
+    htonr_ui(r0, r1);
+    htonr_ui(rn(reg), rn(reg));
+    lshi(r0, r0, 32);
+    orr(r0, r0, rn(reg));
+    jit_unget_reg(reg);
+}
+#    endif
 #  endif
 
 static void

lib/jit_ppc-sz.c

@@ -85,7 +85,9 @@
     4,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    4,	/* htonr */
+    4,	/* htonr_us */
+    4,	/* htonr_ui */
+    0,	/* htonr_ul */
     8,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */
@@ -441,7 +443,9 @@
     4,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    4,	/* htonr */
+    4,	/* htonr_us */
+    4,	/* htonr_ui */
+    0,	/* htonr_ul */
     8,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */
@@ -797,7 +801,15 @@
     4,	/* extr_us */
     4,	/* extr_i */
     4,	/* extr_ui */
-    4,	/* htonr */
+#  if __BYTE_ORDER == __BIG_ENDIAN
+    4,	/* htonr_us */
+    4,	/* htonr_ui */
+    4,	/* htonr_ul */
+#else
+    20,	/* htonr_us */
+    16,	/* htonr_ui */
+    44,	/* htonr_ul */
+#endif
     8,	/* ldr_c */
     28,	/* ldi_c */
     4,	/* ldr_uc */

lib/jit_ppc.c

@@ -949,7 +949,11 @@ _emit_code(jit_state_t *_jit)
 		case_rr(ext, _i);
 		case_rr(ext, _ui);
 #  endif
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
+#  if __WORDSIZE == 64
+		case_rr(hton, _ul);
+#  endif
 		case_rr(neg,);
 		case_rr(com,);
 		case_rr(mov,);

lib/jit_s390x-cpu.c

@@ -1004,7 +1004,9 @@ static void _ori(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 static void _xorr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define xori(r0,r1,i0)		_xori(_jit,r0,r1,i0)
 static void _xori(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
-#  define htonr(r0,r1)			movr(r0,r1)
+#  define htonr_us(r0,r1)		extr_us(r0,r1)
+#  define htonr_ui(r0,r1)		extr_ui(r0,r1)
+#  define htonr_ul(r0,r1)		movr(r0,r1)
 #  define extr_c(r0,r1)			LGBR(r0,r1)
 #  define extr_uc(r0,r1)		LLGCR(r0,r1)
 #  define extr_s(r0,r1)			LGHR(r0,r1)

lib/jit_s390x-sz.c

@@ -84,7 +84,9 @@
     4,	/* extr_us */
     4,	/* extr_i */
     4,	/* extr_ui */
-    4,	/* htonr */
+    4,	/* htonr_us */
+    4,	/* htonr_ui */
+    4,	/* htonr_ul */
     6,	/* ldr_c */
     18,	/* ldi_c */
     6,	/* ldr_uc */

lib/jit_s390x.c

@@ -808,7 +808,9 @@ _emit_code(jit_state_t *_jit)
 		case_wrr(stx, _i);
 		case_rrr(stx, _l);
 		case_wrr(stx, _l);
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
+		case_rr(hton, _ul);
 		case_rr(ext, _c);
 		case_rr(ext, _uc);
 		case_rr(ext, _s);

lib/jit_sparc-cpu.c

@@ -508,7 +508,8 @@ static void _xori(jit_state_t*, jit_int32_t, jit_int32_t, jit_word_t);
 #  define rshi(r0, r1, i0)		SRAI(r1, i0, r0)
 #  define rshr_u(r0, r1, r2)		SRL(r1, r2, r0)
 #  define rshi_u(r0, r1, i0)		SRLI(r1, i0, r0)
-#  define htonr(r0,r1)			movr(r0,r1)
+#  define htonr_us(r0,r1)		extr_us(r0,r1)
+#  define htonr_ui(r0,r1)		movr(r0,r1)
 #  define extr_c(r0,r1)			_extr_c(_jit,r0,r1)
 static void _extr_c(jit_state_t*,jit_int32_t,jit_int32_t);
 #  define extr_uc(r0,r1)		andi(r0, r1, 0xff)

lib/jit_sparc-sz.c

@@ -84,7 +84,9 @@
     8,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    4,	/* htonr */
+    4,	/* htonr_us */
+    4,	/* htonr_ui */
+    0,	/* htonr_ul */
     4,	/* ldr_c */
     12,	/* ldi_c */
     4,	/* ldr_uc */

lib/jit_sparc.c

@@ -830,7 +830,8 @@ _emit_code(jit_state_t *_jit)
 		case_wrr(stx, _s);
 		case_rrr(stx, _i);
 		case_wrr(stx, _i);
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
 		case_rr(ext, _c);
 		case_rr(ext, _uc);
 		case_rr(ext, _s);

lib/jit_x86-cpu.c

@@ -371,8 +371,14 @@ static void _movir(jit_state_t*,jit_int32_t,jit_int32_t);
 #    define movir_u(r0, r1)		_movir_u(_jit, r0, r1)
 static void _movir_u(jit_state_t*,jit_int32_t,jit_int32_t);
 #  endif
-#define htonr(r0, r1)			_htonr(_jit, r0, r1)
+#  define htonr_us(r0, r1)		_htonr_us(_jit, r0, r1)
-static void _htonr(jit_state_t*,jit_int32_t,jit_int32_t);
+static void _htonr_us(jit_state_t*,jit_int32_t,jit_int32_t);
+#  define htonr_ui(r0, r1)		_htonr_ui(_jit, r0, r1)
+static void _htonr_ui(jit_state_t*,jit_int32_t,jit_int32_t);
+#  if __X64 && !__X64_32
+#define htonr_ul(r0, r1)		_htonr_ul(_jit, r0, r1)
+static void _htonr_ul(jit_state_t*,jit_int32_t,jit_int32_t);
+#endif
 #  define extr_c(r0, r1)		_extr_c(_jit, r0, r1)
 static void _extr_c(jit_state_t*,jit_int32_t,jit_int32_t);
 #  define extr_uc(r0, r1)		_extr_uc(_jit, r0, r1)
@@ -438,7 +444,6 @@ static void _ldxi_s(jit_state_t*, jit_int32_t, jit_int32_t, jit_word_t);
 static void _ldxr_us(jit_state_t*, jit_int32_t, jit_int32_t, jit_int32_t);
 #  define ldxi_us(r0, r1, i0)		_ldxi_us(_jit, r0, r1, i0)
 static void _ldxi_us(jit_state_t*, jit_int32_t, jit_int32_t, jit_word_t);
-
 #  if __X32 || !__X64_32
 #    define ldxr_i(r0, r1, r2)		_ldxr_i(_jit, r0, r1, r2)
 static void _ldxr_i(jit_state_t*, jit_int32_t, jit_int32_t, jit_int32_t);
@@ -2203,14 +2208,36 @@ _movir_u(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 #endif
 
 static void
-_htonr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+_htonr_us(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    extr_us(r0, r1);
+    ic(0x66);
+    rex(0, 0, _NOREG, _NOREG, r0);
+    ic(0xc1);
+    mrm(0x03, X86_ROR, r7(r0));
+    ic(8);
+}
+
+static void
+_htonr_ui(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
     movr(r0, r1);
-    rex(0, WIDE, _NOREG, _NOREG, r0);
+    rex(0, 0, _NOREG, _NOREG, r0);
     ic(0x0f);
     ic(0xc8 | r7(r0));
 }
 
+#if __X64 && !__X64_32
+static void
+_htonr_ul(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
+{
+    movr(r0, r1);
+    rex(0, 1, _NOREG, _NOREG, r0);
+    ic(0x0f);
+    ic(0xc8 | r7(r0));
+}
+#endif
+
 static void
 _extr_c(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {

lib/jit_x86-sz.c

@@ -84,7 +84,9 @@
     3,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    4,	/* htonr */
+    4,	/* htonr_us */
+    4,	/* htonr_ui */
+    0,	/* htonr_ul */
     3,	/* ldr_c */
     7,	/* ldi_c */
     3,	/* ldr_uc */
@@ -439,7 +441,9 @@
     4,	/* extr_us */
     3,	/* extr_i */
     3,	/* extr_ui */
-    6,	/* htonr */
+    9,	/* htonr_us */
+    6,	/* htonr_ui */
+    6,	/* htonr_ul */
     4,	/* ldr_c */
     15,	/* ldi_c */
     4,	/* ldr_uc */
@@ -793,7 +797,9 @@
     4,	/* extr_us */
     0,	/* extr_i */
     0,	/* extr_ui */
-    6,	/* htonr */
+    9,	/* htonr_us */
+    6,	/* htonr_ui */
+    0,	/* htonr_ul */
     5,	/* ldr_c */
     9,	/* ldi_c */
     5,	/* ldr_uc */
@@ -1146,7 +1152,9 @@
     4,	/* extr_us */
     3,	/* extr_i */
     3,	/* extr_ui */
-    6,	/* htonr */
+    9,	/* htonr_us */
+    6,	/* htonr_ui */
+    6,	/* htonr_ul */
     5,	/* ldr_c */
     9,	/* ldi_c */
     5,	/* ldr_uc */

lib/jit_x86.c

@@ -1376,7 +1376,11 @@ _emit_code(jit_state_t *_jit)
 		else
 		    movi(rn(node->u.w), node->v.w);
 		break;
-		case_rr(hton,);
+		case_rr(hton, _us);
+		case_rr(hton, _ui);
+#if __X64 && !__X64_32
+		case_rr(hton, _ul);
+#endif
 		case_rr(ext, _c);
 		case_rr(ext, _uc);
 		case_rr(ext, _s);

lib/lightning.c

@@ -1237,7 +1237,8 @@ _jit_classify(jit_state_t *_jit, jit_code_t code)
 	case jit_code_extr_us:	case jit_code_extr_i:	case jit_code_extr_ui:
 	case jit_code_truncr_f_i:			case jit_code_truncr_f_l:
 	case jit_code_truncr_d_i:			case jit_code_truncr_d_l:
-	case jit_code_htonr:	case jit_code_ldr_c:	case jit_code_ldr_uc:
+	case jit_code_htonr_us:	case jit_code_htonr_ui:	case jit_code_htonr_ul:
+	case jit_code_ldr_c:	case jit_code_ldr_uc:
 	case jit_code_ldr_s:	case jit_code_ldr_us:	case jit_code_ldr_i:
 	case jit_code_ldr_ui:	case jit_code_ldr_l:	case jit_code_negr_f:
 	case jit_code_absr_f:	case jit_code_sqrtr_f:	case jit_code_movr_f: