From 960280decd66fed85d9d64d2442008420c0cf73f Mon Sep 17 00:00:00 2001
From: pcpa <paulo.cesar.pereira.de.andrade@gmail.com>
Date: Sat, 18 Oct 2014 11:31:18 -0300
Subject: [PATCH] Implement the jit_rsb* interface.

	* check/alu_rsb.ok, check/alu_rsb.tst: New files implementing
	tests for jit_rsb*.

	* check/Makefile.am, check/lightning.c, include/lightning.h,
	lib/jit_aarch64-cpu.c, lib/jit_aarch64-fpu.c, lib/jit_aarch64-sz.c,
	lib/jit_aarch64.c, lib/jit_alpha-cpu.c, lib/jit_alpha-fpu.c,
	lib/jit_alpha-sz.c, lib/jit_alpha.c, lib/jit_arm-cpu.c,
	lib/jit_arm-swf.c, lib/jit_arm-sz.c, lib/jit_arm-vfp.c,
	lib/jit_arm.c, lib/jit_hppa-cpu.c, lib/jit_hppa-fpu.c,
	lib/jit_hppa-sz.c, lib/jit_hppa.c, lib/jit_ia64-cpu.c,
	lib/jit_ia64-fpu.c, lib/jit_ia64-sz.c, lib/jit_ia64.c,
	lib/jit_mips-cpu.c, lib/jit_mips-fpu.c, lib/jit_mips-sz.c,
	lib/jit_mips.c, lib/jit_names.c, lib/jit_ppc-cpu.c,
	lib/jit_ppc-fpu.c, lib/jit_ppc-sz.c, lib/jit_ppc.c,
	lib/jit_s390x-cpu.c, lib/jit_s390x-fpu.c, lib/jit_s390x-sz.c,
	lib/jit_s390x.c, lib/jit_sparc-cpu.c, lib/jit_sparc-fpu.c,
	lib/jit_sparc-sz.c, lib/jit_sparc.c, lib/jit_x86-cpu.c,
	lib/jit_x86-sse.c, lib/jit_x86-sz.c, lib/jit_x86-x87.c,
	lib/jit_x86.c, lib/lightning.c: Implement jit_rsb*. This
	was a missing lightning 1.x interface, that on most
	backends is synthesized, but on a few backends (hppa and ia64),
	it can generate better code as on those there is, or the
	only instruction with an immediate is in "rsb" format
	(left operand).
---
 ChangeLog             | 27 +++++++++++++++++
 check/Makefile.am     | 13 +++++----
 check/alu_rsb.ok      |  1 +
 check/alu_rsb.tst     | 49 +++++++++++++++++++++++++++++++
 check/lightning.c     |  9 ++++++
 include/lightning.h   | 10 ++++++-
 lib/jit_aarch64-cpu.c |  9 ++++++
 lib/jit_aarch64-fpu.c |  8 +++++
 lib/jit_aarch64-sz.c  |  3 ++
 lib/jit_aarch64.c     |  3 ++
 lib/jit_alpha-cpu.c   |  9 ++++++
 lib/jit_alpha-fpu.c   |  8 +++++
 lib/jit_alpha-sz.c    |  3 ++
 lib/jit_alpha.c       |  3 ++
 lib/jit_arm-cpu.c     |  9 ++++++
 lib/jit_arm-swf.c     | 68 +++++++++++++++++++++++++++++++++++++++++++
 lib/jit_arm-sz.c      |  6 ++++
 lib/jit_arm-vfp.c     |  8 +++++
 lib/jit_arm.c         |  3 ++
 lib/jit_hppa-cpu.c    | 18 +++++++++++-
 lib/jit_hppa-fpu.c    |  8 +++++
 lib/jit_hppa-sz.c     |  3 ++
 lib/jit_hppa.c        |  3 ++
 lib/jit_ia64-cpu.c    | 18 ++++++++++++
 lib/jit_ia64-fpu.c    |  8 +++++
 lib/jit_ia64-sz.c     |  3 ++
 lib/jit_ia64.c        |  3 ++
 lib/jit_mips-cpu.c    |  9 ++++++
 lib/jit_mips-fpu.c    |  8 +++++
 lib/jit_mips-sz.c     |  9 ++++++
 lib/jit_mips.c        |  3 ++
 lib/jit_names.c       |  1 +
 lib/jit_ppc-cpu.c     | 17 ++++++++---
 lib/jit_ppc-fpu.c     | 16 +++++-----
 lib/jit_ppc-sz.c      |  9 ++++++
 lib/jit_ppc.c         |  3 ++
 lib/jit_s390x-cpu.c   |  9 ++++++
 lib/jit_s390x-fpu.c   |  6 ++++
 lib/jit_s390x-sz.c    |  3 ++
 lib/jit_s390x.c       |  3 ++
 lib/jit_sparc-cpu.c   | 11 ++++++-
 lib/jit_sparc-fpu.c   | 45 +++++++++++++++++++++-------
 lib/jit_sparc-sz.c    |  3 ++
 lib/jit_sparc.c       |  3 ++
 lib/jit_x86-cpu.c     |  9 ++++++
 lib/jit_x86-sse.c     | 10 +++++++
 lib/jit_x86-sz.c      |  9 ++++++
 lib/jit_x86-x87.c     |  9 ++++++
 lib/jit_x86.c         |  3 ++
 lib/lightning.c       | 27 +++++++++--------
 50 files changed, 494 insertions(+), 44 deletions(-)
 create mode 100644 check/alu_rsb.ok
 create mode 100644 check/alu_rsb.tst

diff --git a/ChangeLog b/ChangeLog
index eef40a4ec..36ceff9fa 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,30 @@
+2014-10-17 Paulo Andrade <pcpa@gnu.org>
+
+	* check/alu_rsb.ok, check/alu_rsb.tst: New files implementing
+	tests for jit_rsb*.
+
+	* check/Makefile.am, check/lightning.c, include/lightning.h,
+	lib/jit_aarch64-cpu.c, lib/jit_aarch64-fpu.c, lib/jit_aarch64-sz.c,
+	lib/jit_aarch64.c, lib/jit_alpha-cpu.c, lib/jit_alpha-fpu.c,
+	lib/jit_alpha-sz.c, lib/jit_alpha.c, lib/jit_arm-cpu.c,
+	lib/jit_arm-swf.c, lib/jit_arm-sz.c, lib/jit_arm-vfp.c,
+	lib/jit_arm.c, lib/jit_hppa-cpu.c, lib/jit_hppa-fpu.c,
+	lib/jit_hppa-sz.c, lib/jit_hppa.c, lib/jit_ia64-cpu.c,
+	lib/jit_ia64-fpu.c, lib/jit_ia64-sz.c, lib/jit_ia64.c,
+	lib/jit_mips-cpu.c, lib/jit_mips-fpu.c, lib/jit_mips-sz.c,
+	lib/jit_mips.c, lib/jit_names.c, lib/jit_ppc-cpu.c,
+	lib/jit_ppc-fpu.c, lib/jit_ppc-sz.c, lib/jit_ppc.c,
+	lib/jit_s390x-cpu.c, lib/jit_s390x-fpu.c, lib/jit_s390x-sz.c,
+	lib/jit_s390x.c, lib/jit_sparc-cpu.c, lib/jit_sparc-fpu.c,
+	lib/jit_sparc-sz.c, lib/jit_sparc.c, lib/jit_x86-cpu.c,
+	lib/jit_x86-sse.c, lib/jit_x86-sz.c, lib/jit_x86-x87.c,
+	lib/jit_x86.c, lib/lightning.c: Implement jit_rsb*. This
+	was a missing lightning 1.x interface, that on most
+	backends is synthesized, but on a few backends (hppa and ia64),
+	it can generate better code as on those there is, or the
+	only instruction with an immediate is in "rsb" format
+	(left operand).
+
 2014-10-17 Paulo Andrade <pcpa@gnu.org>
 
 	* lib/jit_names.c: New file with single definition of string
diff --git a/check/Makefile.am b/check/Makefile.am
index 5e0b21b3b..59a4f7a1c 100644
--- a/check/Makefile.am
+++ b/check/Makefile.am
@@ -60,6 +60,7 @@ EXTRA_DIST =				\
 	alux_add.tst	alux_add.ok	\
 	alu_sub.tst	alu_sub.ok	\
 	alux_sub.tst	alux_sub.ok	\
+	alu_rsb.tst	alu_rsb.ok	\
 	alu_mul.tst	alu_mul.ok	\
 	alu_div.tst	alu_div.ok	\
 	alu_rem.tst	alu_rem.ok	\
@@ -99,7 +100,7 @@ base_TESTS =				\
 	ldstr-c ldstxr-c ldstxi-c	\
 	cvt branch			\
 	alu_add alux_add		\
-	alu_sub alux_sub		\
+	alu_sub alux_sub alu_rsb	\
 	alu_mul alu_div alu_rem		\
 	alu_and alu_or alu_xor		\
 	alu_lsh alu_rsh			\
@@ -126,7 +127,7 @@ x87_TESTS =					\
 	ldstr-c.x87 ldstxr-c.x87 ldstxi-c.x87	\
 	cvt.x87 branch.x87			\
 	alu_add.x87 alux_add.x87		\
-	alu_sub.x87 alux_sub.x87		\
+	alu_sub.x87 alux_sub.x87 alu_rsb.x87	\
 	alu_mul.x87 alu_div.x87 alu_rem.x87	\
 	alu_and.x87 alu_or.x87 alu_xor.x87	\
 	alu_lsh.x87 alu_rsh.x87			\
@@ -148,7 +149,7 @@ x87_nodata_TESTS =							\
 	ldstr-c.x87.nodata ldstxr-c.x87.nodata ldstxi-c.x87.nodata	\
 	cvt.x87.nodata branch.x87.nodata				\
 	alu_add.x87.nodata alux_add.x87.nodata				\
-	alu_sub.x87.nodata alux_sub.x87.nodata				\
+	alu_sub.x87.nodata alux_sub.x87.nodata alu_rsb.x87.nodata	\
 	alu_mul.x87.nodata alu_div.x87.nodata alu_rem.x87.nodata	\
 	alu_and.x87.nodata alu_or.x87.nodata alu_xor.x87.nodata		\
 	alu_lsh.x87.nodata alu_rsh.x87.nodata				\
@@ -172,7 +173,7 @@ arm_TESTS =					\
 	ldstr-c.arm ldstxr-c.arm ldstxi-c.arm	\
 	cvt.arm branch.arm			\
 	alu_add.arm alux_add.arm		\
-	alu_sub.arm alux_sub.arm		\
+	alu_sub.arm alux_sub.arm alu_rsb.arm	\
 	alu_mul.arm alu_div.arm alu_rem.arm	\
 	alu_and.arm alu_or.arm alu_xor.arm	\
 	alu_lsh.arm alu_rsh.arm			\
@@ -196,7 +197,7 @@ swf_TESTS =					\
 	ldstr-c.swf ldstxr-c.swf ldstxi-c.swf	\
 	cvt.swf branch.swf			\
 	alu_add.swf alux_add.swf		\
-	alu_sub.swf alux_sub.swf		\
+	alu_sub.swf alux_sub.swf alu_rsb.swf	\
 	alu_mul.swf alu_div.swf alu_rem.swf	\
 	alu_and.swf alu_or.swf alu_xor.swf	\
 	alu_lsh.swf alu_rsh.swf			\
@@ -220,7 +221,7 @@ nodata_TESTS =						\
 	ldstr-c.nodata ldstxr-c.nodata ldstxi-c.nodata	\
 	cvt.nodata branch.nodata			\
 	alu_add.nodata alux_add.nodata			\
-	alu_sub.nodata alux_sub.nodata			\
+	alu_sub.nodata alux_sub.nodata alu_rsb.nodata	\
 	alu_mul.nodata alu_div.nodata alu_rem.nodata	\
 	alu_and.nodata alu_or.nodata alu_xor.nodata	\
 	alu_lsh.nodata alu_rsh.nodata			\
diff --git a/check/alu_rsb.ok b/check/alu_rsb.ok
new file mode 100644
index 000000000..9766475a4
--- /dev/null
+++ b/check/alu_rsb.ok
@@ -0,0 +1 @@
+ok
diff --git a/check/alu_rsb.tst b/check/alu_rsb.tst
new file mode 100644
index 000000000..00e08c22f
--- /dev/null
+++ b/check/alu_rsb.tst
@@ -0,0 +1,49 @@
+#include "alu.inc"
+
+.code
+	prolog
+
+#define RSB(N, I0, I1, V)	ALU(N, , rsb, I0, I1, V)
+
+	RSB(0,	1,	    0x7fffffff,	0x7ffffffe)
+	RSB(2,	1,	    0x80000000,	0x7fffffff)
+	RSB(3,	0x7fffffff, 0x80000000,	1)
+	RSB(4,	0xffffffff, 0xffffffff,	0)
+	RSB(5,	0x7fffffff, 0xffffffff, 0x80000000)
+	RSB(6,	0, 0x7fffffff,		0x7fffffff)
+#if __WORDSIZE == 32
+	RSB(7,	0x7fffffff,	     1, 0x80000002)
+	RSB(8,	0x80000000,	     1, 0x80000001)
+	RSB(9,  0x80000000, 0x7fffffff, 0xffffffff)
+	RSB(10, 0xffffffff, 0x7fffffff, 0x80000000)
+	RSB(11,	0x7fffffff,	     0,	0x80000001)
+#else
+	RSB(7,	        0x7fffffff,		    1,	0xffffffff80000002)
+	RSB(8,  0xffffffff80000000,		    1,		0x80000001)
+	RSB(9,  0xffffffff80000000,	   0x7fffffff,		0xffffffff)
+	RSB(10, 0xffffffffffffffff, 0xffffffff7fffffff, 0xffffffff80000000)
+	RSB(11,	        0x7fffffff,		     0, 0xffffffff80000001)
+	RSB(12,	1,		    0x7fffffffffffffff, 0x7ffffffffffffffe)
+	RSB(13,	0x7fffffffffffffff,		     1, 0x8000000000000002)
+	RSB(14,			 1, 0x8000000000000000, 0x7fffffffffffffff)
+	RSB(15,	0x8000000000000000,		     1, 0x8000000000000001)
+	RSB(16, 0x8000000000000000, 0x7fffffffffffffff, 0xffffffffffffffff)
+	RSB(17, 0x7fffffffffffffff, 0x8000000000000000, 1)
+	RSB(18, 0xffffffffffffffff, 0x7fffffffffffffff, 0x8000000000000000)
+	RSB(19, 0x7fffffffffffffff, 0xffffffffffffffff, 0x8000000000000000)
+	RSB(20, 0xffffffffffffffff, 0xffffffffffffffff, 0)
+#endif
+
+#undef RSB
+#define RSB(N, T, I0, I1, V)	FOP(N, T, rsb, I0, I1, V)
+	RSB(0, _f,	    0.5,	-0.5,	-1.0)
+	RSB(1, _f,	    0.75,	 0.25,	-0.5)
+	RSB(0, _d,	    0.5,	-0.5,	-1.0)
+	RSB(1, _d,	    0.75,	 0.25,	-0.5)
+
+	prepare
+		pushargi ok
+		ellipsis
+	finishi @printf
+	ret
+	epilog
diff --git a/check/lightning.c b/check/lightning.c
index cf08dea7c..5f179fbed 100644
--- a/check/lightning.c
+++ b/check/lightning.c
@@ -286,6 +286,7 @@ static void addcr(void);	static void addci(void);
 static void subr(void);		static void subi(void);
 static void subxr(void);	static void subxi(void);
 static void subcr(void);	static void subci(void);
+static void rsbr(void);		static void rsbi(void);
 static void mulr(void);		static void muli(void);
 static void qmulr(void);	static void qmuli(void);
 static void qmulr_u(void);	static void qmuli_u(void);
@@ -392,6 +393,7 @@ static void arg_f(void);
 static void getarg_f(void);
 static void addr_f(void);	static void addi_f(void);
 static void subr_f(void);	static void subi_f(void);
+static void rsbr_f(void);	static void rsbi_f(void);
 static void mulr_f(void);	static void muli_f(void);
 static void divr_f(void);	static void divi_f(void);
 static void negr_f(void);	static void absr_f(void);
@@ -442,6 +444,7 @@ static void arg_d(void);
 static void getarg_d(void);
 static void addr_d(void);	static void addi_d(void);
 static void subr_d(void);	static void subi_d(void);
+static void rsbr_d(void);	static void rsbi_d(void);
 static void mulr_d(void);	static void muli_d(void);
 static void divr_d(void);	static void divi_d(void);
 static void negr_d(void);	static void absr_d(void);
@@ -586,6 +589,7 @@ static instr_t		  instr_vector[] = {
     entry(subr),	entry(subi),
     entry(subxr),	entry(subxi),
     entry(subcr),	entry(subci),
+    entry(rsbr),	entry(rsbi),
     entry(mulr),	entry(muli),
     entry(qmulr),	entry(qmuli),
     entry(qmulr_u),	entry(qmuli_u),
@@ -692,6 +696,7 @@ static instr_t		  instr_vector[] = {
     entry(getarg_f),
     entry(addr_f),	entry(addi_f),
     entry(subr_f),	entry(subi_f),
+    entry(rsbr_f),	entry(rsbi_f),
     entry(mulr_f),	entry(muli_f),
     entry(divr_f),	entry(divi_f),
     entry(negr_f),	entry(absr_f),
@@ -742,6 +747,7 @@ static instr_t		  instr_vector[] = {
     entry(getarg_d),
     entry(addr_d),	entry(addi_d),
     entry(subr_d),	entry(subi_d),
+    entry(rsbr_d),	entry(rsbi_d),
     entry(mulr_d),	entry(muli_d),
     entry(divr_d),	entry(divi_d),
     entry(negr_d),	entry(absr_d),
@@ -1351,6 +1357,7 @@ entry_ir_ir_ir(addcr)		entry_ir_ir_im(addci)
 entry_ir_ir_ir(subr)		entry_ir_ir_im(subi)
 entry_ir_ir_ir(subxr)		entry_ir_ir_im(subxi)
 entry_ir_ir_ir(subcr)		entry_ir_ir_im(subci)
+entry_ir_ir_ir(rsbr)		entry_ir_ir_im(rsbi)
 entry_ir_ir_ir(mulr)		entry_ir_ir_im(muli)
 entry_ir_ir_ir_ir(qmulr)	entry_ir_ir_ir_im(qmuli)
 entry_ir_ir_ir_ir(qmulr_u)	entry_ir_ir_ir_im(qmuli_u)
@@ -1504,6 +1511,7 @@ entry_ca(arg_f)
 entry_fa(getarg_f)
 entry_fr_fr_fr(addr_f)		entry_fr_fr_fm(addi_f)
 entry_fr_fr_fr(subr_f)		entry_fr_fr_fm(subi_f)
+entry_fr_fr_fr(rsbr_f)		entry_fr_fr_fm(rsbi_f)
 entry_fr_fr_fr(mulr_f)		entry_fr_fr_fm(muli_f)
 entry_fr_fr_fr(divr_f)		entry_fr_fr_fm(divi_f)
 entry_fr_fr(negr_f)		entry_fr_fr(absr_f)
@@ -1554,6 +1562,7 @@ entry_ca(arg_d)
 entry_fa(getarg_d)
 entry_fr_fr_fr(addr_d)		entry_fr_fr_dm(addi_d)
 entry_fr_fr_fr(subr_d)		entry_fr_fr_dm(subi_d)
+entry_fr_fr_fr(rsbr_d)		entry_fr_fr_dm(rsbi_d)
 entry_fr_fr_fr(mulr_d)		entry_fr_fr_dm(muli_d)
 entry_fr_fr_fr(divr_d)		entry_fr_fr_dm(divi_d)
 entry_fr_fr(negr_d)		entry_fr_fr(absr_d)
diff --git a/include/lightning.h b/include/lightning.h
index 2e87905be..926d0f18c 100644
--- a/include/lightning.h
+++ b/include/lightning.h
@@ -228,7 +228,9 @@ typedef enum {
 #define jit_subxr(u,v,w)	jit_new_node_www(jit_code_subxr,u,v,w)
 #define jit_subxi(u,v,w)	jit_new_node_www(jit_code_subxi,u,v,w)
     jit_code_subxr,		jit_code_subxi,
-
+#define jit_rsbr(u,v,w)		jit_subr(u,w,v)
+#define jit_rsbi(u,v,w)		jit_new_node_www(jit_code_rsbi,u,v,w)
+    jit_code_rsbi,
 #define jit_mulr(u,v,w)		jit_new_node_www(jit_code_mulr,u,v,w)
 #define jit_muli(u,v,w)		jit_new_node_www(jit_code_muli,u,v,w)
     jit_code_mulr,		jit_code_muli,
@@ -533,6 +535,9 @@ typedef enum {
 #define jit_subr_f(u,v,w)	jit_new_node_www(jit_code_subr_f,u,v,w)
 #define jit_subi_f(u,v,w)	jit_new_node_wwf(jit_code_subi_f,u,v,w)
     jit_code_subr_f,		jit_code_subi_f,
+#define jit_rsbr_f(u,v,w)	jit_subr_f(u,w,v)
+#define jit_rsbi_f(u,v,w)	jit_new_node_wwf(jit_code_rsbi_f,u,v,w)
+    jit_code_rsbi_f,
 #define jit_mulr_f(u,v,w)	jit_new_node_www(jit_code_mulr_f,u,v,w)
 #define jit_muli_f(u,v,w)	jit_new_node_wwf(jit_code_muli_f,u,v,w)
     jit_code_mulr_f,		jit_code_muli_f,
@@ -675,6 +680,9 @@ typedef enum {
 #define jit_subr_d(u,v,w)	jit_new_node_www(jit_code_subr_d,u,v,w)
 #define jit_subi_d(u,v,w)	jit_new_node_wwd(jit_code_subi_d,u,v,w)
     jit_code_subr_d,		jit_code_subi_d,
+#define jit_rsbr_d(u,v,w)	jit_subr_d(u,w,v)
+#define jit_rsbi_d(u,v,w)	jit_new_node_wwd(jit_code_rsbi_d,u,v,w)
+    jit_code_rsbi_d,
 #define jit_mulr_d(u,v,w)	jit_new_node_www(jit_code_mulr_d,u,v,w)
 #define jit_muli_d(u,v,w)	jit_new_node_wwd(jit_code_muli_d,u,v,w)
     jit_code_mulr_d,		jit_code_muli_d,
diff --git a/lib/jit_aarch64-cpu.c b/lib/jit_aarch64-cpu.c
index fd25f7f30..29834199b 100644
--- a/lib/jit_aarch64-cpu.c
+++ b/lib/jit_aarch64-cpu.c
@@ -520,6 +520,8 @@ static void _subci(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #  define subxr(r0,r1,r2)		SBCS(r0,r1,r2)
 #  define subxi(r0,r1,i0)		_subxi(_jit,r0,r1,i0)
 static void _subxi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
+#  define rsbi(r0, r1, i0)		_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #  define mulr(r0,r1,r2)		MUL(r0,r1,r2)
 #  define muli(r0,r1,i0)		_muli(_jit,r0,r1,i0)
 static void _muli(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
@@ -1121,6 +1123,13 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     jit_unget_reg(reg);
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    subi(r0, r1, i0);
+    negr(r0, r0);
+}
+
 static void
 _muli(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {
diff --git a/lib/jit_aarch64-fpu.c b/lib/jit_aarch64-fpu.c
index ad20714ce..ea8c1cd52 100644
--- a/lib/jit_aarch64-fpu.c
+++ b/lib/jit_aarch64-fpu.c
@@ -86,6 +86,9 @@ static void _addi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t);
 #  define subr_f(r0,r1,r2)		FSUBS(r0,r1,r2)
 #  define subi_f(r0,r1,i0)		_subi_f(_jit,r0,r1,i0)
 static void _subi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t);
+#  define rsbr_f(r0, r1, r2)		subr_f(r0, r2, r1)
+#  define rsbi_f(r0, r1, i0)		_rsbi_f(_jit, r0, r1, i0)
+static void _rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t);
 #  define mulr_f(r0,r1,r2)		FMULS(r0,r1,r2)
 #  define muli_f(r0,r1,i0)		_muli_f(_jit,r0,r1,i0)
 static void _muli_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t);
@@ -198,6 +201,9 @@ static void _addi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t);
 #  define subr_d(r0,r1,r2)		FSUBD(r0,r1,r2)
 #  define subi_d(r0,r1,i0)		_subi_d(_jit,r0,r1,i0)
 static void _subi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t);
+#  define rsbr_d(r0, r1, r2)		subr_d(r0, r2, r1)
+#  define rsbi_d(r0, r1, i0)		_rsbi_d(_jit, r0, r1, i0)
+static void _rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t);
 #  define mulr_d(r0,r1,r2)		FMULD(r0,r1,r2)
 #  define muli_d(r0,r1,i0)		_muli_d(_jit,r0,r1,i0)
 static void _muli_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t);
@@ -420,6 +426,7 @@ _truncr_d_i(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 
 fopi(add)
 fopi(sub)
+fopi(rsb)
 fopi(mul)
 fopi(div)
 
@@ -636,6 +643,7 @@ fbopi(ltgt)
 
 dopi(add)
 dopi(sub)
+dopi(rsb)
 dopi(mul)
 dopi(div)
 
diff --git a/lib/jit_aarch64-sz.c b/lib/jit_aarch64-sz.c
index 98a30f322..bfb096045 100644
--- a/lib/jit_aarch64-sz.c
+++ b/lib/jit_aarch64-sz.c
@@ -23,6 +23,7 @@
     12,	/* subci */
     4,	/* subxr */
     8,	/* subxi */
+    20,	/* rsbi */
     4,	/* mulr */
     12,	/* muli */
     12,	/* qmulr */
@@ -178,6 +179,7 @@
     12,	/* addi_f */
     4,	/* subr_f */
     12,	/* subi_f */
+    12,	/* rsbi_f */
     4,	/* mulr_f */
     12,	/* muli_f */
     4,	/* divr_f */
@@ -260,6 +262,7 @@
     12,	/* addi_d */
     4,	/* subr_d */
     12,	/* subi_d */
+    12,	/* rsbi_d */
     4,	/* mulr_d */
     12,	/* muli_d */
     4,	/* divr_d */
diff --git a/lib/jit_aarch64.c b/lib/jit_aarch64.c
index 6de141c2b..52d5c4f55 100644
--- a/lib/jit_aarch64.c
+++ b/lib/jit_aarch64.c
@@ -744,6 +744,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(subc,);
 		case_rrr(subx,);
 		case_rrw(subx,);
+		case_rrw(rsb,);
 		case_rrr(mul,);
 		case_rrw(mul,);
 		case_rrrr(qmul,);
@@ -913,6 +914,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add);
 		case_rrr(sub, _f);
 		case_rrf(sub);
+		case_rrf(rsb);
 		case_rrr(mul, _f);
 		case_rrf(mul);
 		case_rrr(div, _f);
@@ -995,6 +997,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrd(add);
 		case_rrr(sub, _d);
 		case_rrd(sub);
+		case_rrd(rsb);
 		case_rrr(mul, _d);
 		case_rrd(mul);
 		case_rrr(div, _d);
diff --git a/lib/jit_alpha-cpu.c b/lib/jit_alpha-cpu.c
index 8787b0ffd..03513fcd7 100644
--- a/lib/jit_alpha-cpu.c
+++ b/lib/jit_alpha-cpu.c
@@ -335,6 +335,8 @@ static void _subci(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 static void _subxr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define subxi(r0,r1,i0)		_subxi(_jit,r0,r1,i0)
 static void _subxi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
+#  define rsbi(r0, r1, i0)		_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t, jit_word_t);
 #  define mulr(r0,r1,r2)		MULQ(r1,r2,r0)
 #  define muli(r0,r1,i0)		_muli(_jit,r0,r1,i0)
 static void _muli(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
@@ -954,6 +956,13 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     jit_unget_reg(reg);
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    subi(r0, r1, i0);
+    negr(r0, r0);
+}
+
 static void
 _muli(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {
diff --git a/lib/jit_alpha-fpu.c b/lib/jit_alpha-fpu.c
index 026fd4127..e6f92cf92 100644
--- a/lib/jit_alpha-fpu.c
+++ b/lib/jit_alpha-fpu.c
@@ -340,6 +340,12 @@ static void _subi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
 static void _subr_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define subi_d(r0,r1,i0)		_subi_d(_jit,r0,r1,i0)
 static void _subi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
+#  define rsbr_f(r0, r1, r2)		subr_f(r0, r2, r1)
+#  define rsbi_f(r0, r1, i0)		_rsbi_f(_jit, r0, r1, i0)
+static void _rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
+#  define rsbr_d(r0, r1, r2)		subr_d(r0, r2, r1)
+#  define rsbi_d(r0, r1, i0)		_rsbi_d(_jit, r0, r1, i0)
+static void _rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
 #  define mulr_f(r0,r1,r2)		_mulr_f(_jit,r0,r1,r2)
 static void _mulr_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define muli_f(r0,r1,i0)		_muli_f(_jit,r0,r1,i0)
@@ -726,6 +732,7 @@ _subr_f(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
     TRAPB();
 }
 fopi(sub)
+fopi(rsb)
 
 static void
 _subr_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
@@ -734,6 +741,7 @@ _subr_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
     TRAPB();
 }
 dopi(sub)
+dopi(rsb)
 
 static void
 _mulr_f(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
diff --git a/lib/jit_alpha-sz.c b/lib/jit_alpha-sz.c
index a5c9ad227..5bcd24ed1 100644
--- a/lib/jit_alpha-sz.c
+++ b/lib/jit_alpha-sz.c
@@ -23,6 +23,7 @@
     40,	/* subci */
     28,	/* subxr */
     28,	/* subxi */
+    36,	/* rsbi */
     4,	/* mulr */
     32,	/* muli */
     44,	/* qmulr */
@@ -178,6 +179,7 @@
     32,	/* addi_f */
     8,	/* subr_f */
     32,	/* subi_f */
+    32,	/* rsbi_f */
     8,	/* mulr_f */
     32,	/* muli_f */
     8,	/* divr_f */
@@ -260,6 +262,7 @@
     28,	/* addi_d */
     8,	/* subr_d */
     28,	/* subi_d */
+    28,	/* rsbi_d */
     8,	/* mulr_d */
     28,	/* muli_d */
     8,	/* divr_d */
diff --git a/lib/jit_alpha.c b/lib/jit_alpha.c
index 6d78c70d4..6fbdf6d46 100644
--- a/lib/jit_alpha.c
+++ b/lib/jit_alpha.c
@@ -771,6 +771,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(subc,);
 		case_rrr(subx,);
 		case_rrw(subx,);
+		case_rrw(rsb,);
 		case_rrr(mul,);
 		case_rrw(mul,);
 		case_rrrr(qmul,);
@@ -940,6 +941,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _f, 32);
 		case_rrr(sub, _f);
 		case_rrf(sub, _f, 32);
+		case_rrf(rsb, _f, 32);
 		case_rrr(mul, _f);
 		case_rrf(mul, _f, 32);
 		case_rrr(div, _f);
@@ -1022,6 +1024,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _d, 64);
 		case_rrr(sub, _d);
 		case_rrf(sub, _d, 64);
+		case_rrf(rsb, _d, 64);
 		case_rrr(mul, _d);
 		case_rrf(mul, _d, 64);
 		case_rrr(div, _d);
diff --git a/lib/jit_arm-cpu.c b/lib/jit_arm-cpu.c
index 3f7db0559..552bce3d8 100644
--- a/lib/jit_arm-cpu.c
+++ b/lib/jit_arm-cpu.c
@@ -868,6 +868,8 @@ static void _subci(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 static void _subxr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define subxi(r0,r1,i0)		_subxi(_jit,r0,r1,i0)
 static void _subxi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
+#  define rsbi(r0, r1, i0)		_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #  define mulr(r0,r1,r2)		_mulr(_jit,r0,r1,r2)
 static void _mulr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define muli(r0,r1,i0)		_muli(_jit,r0,r1,i0)
@@ -1948,6 +1950,13 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     }
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    subi(r0, r1, i0);
+    negr(r0, r0);
+}
+
 static void
 _mulr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
 {
diff --git a/lib/jit_arm-swf.c b/lib/jit_arm-swf.c
index f07d65d1c..8afd9e54a 100644
--- a/lib/jit_arm-swf.c
+++ b/lib/jit_arm-swf.c
@@ -152,6 +152,12 @@ static void _swf_negr_d(jit_state_t*,jit_int32_t,jit_int32_t);
 #  define swf_subi_f(r0,r1,i0)		swf_fff_(__aeabi_fsub,r0,r1,i0)
 #  define swf_subr_d(r0,r1,r2)		swf_ddd(__aeabi_dsub,r0,r1,r2)
 #  define swf_subi_d(r0,r1,i0)		swf_ddd_(__aeabi_dsub,r0,r1,i0)
+#  define swf_rsbr_f(r0, r1, r2)	swf_subr_f(r0, r2, r1)
+#  define swf_rsbi_f(r0, r1, i0)	_swf_rsbi_f(_jit, r0, r1, i0)
+static void _swf_rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t);
+#  define swf_rsbr_d(r0, r1, r2)	swf_subr_d(r0, r2, r1)
+#  define swf_rsbi_d(r0, r1, i0)	_swf_rsbi_d(_jit, r0, r1, i0)
+static void _swf_rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t);
 #  define swf_mulr_f(r0,r1,r2)		swf_fff(__aeabi_fmul,r0,r1,r2)
 #  define swf_muli_f(r0,r1,i0)		swf_fff_(__aeabi_fmul,r0,r1,i0)
 #  define swf_mulr_d(r0,r1,r2)		swf_ddd(__aeabi_dmul,r0,r1,r2)
@@ -690,6 +696,28 @@ _swf_fff_(jit_state_t *_jit, float (*i0)(float, float),
     jit_unget_reg_args();
 }
 
+static void
+_swf_rsbi_f(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_float32_t i0)
+{
+    union {
+	jit_int32_t	i;
+	jit_float32_t	f;
+    } data;
+    jit_get_reg_args();
+    data.f = i0;
+    movi(_R0_REGNO, data.i);
+    if (jit_fpr_p(r1))
+	swf_ldrin(_R1_REGNO, _FP_REGNO, swf_off(r1) + 8);
+    else
+	movr(_R1_REGNO, r1);
+    swf_call(__aeabi_fsub, fallback, _R3_REGNO);
+    if (jit_fpr_p(r0))
+	swf_strin(_R0_REGNO, _FP_REGNO, swf_off(r0) + 8);
+    else
+	movr(r0, _R0_REGNO);
+    jit_unget_reg_args();
+}
+
 static void
 _swf_ddd_(jit_state_t *_jit, double (*i0)(double, double),
 	  jit_int32_t r0, jit_int32_t r1, jit_float64_t i1)
@@ -699,6 +727,7 @@ _swf_ddd_(jit_state_t *_jit, double (*i0)(double, double),
 	jit_float64_t	d;
     } data;
     jit_get_reg_args();
+
     data.d = i1;
     if (jit_fpr_p(r1)) {
 	if (!jit_thumb_p() && jit_armv5e_p())
@@ -730,6 +759,45 @@ _swf_ddd_(jit_state_t *_jit, double (*i0)(double, double),
     jit_unget_reg_args();
 }
 
+static void
+_swf_rsbi_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_float64_t i0)
+{
+    union {
+	jit_int32_t	i[2];
+	jit_float64_t	d;
+    } data;
+    jit_get_reg_args();
+    data.d = i0;
+    movi(_R0_REGNO, data.i[0]);
+    movi(_R1_REGNO, data.i[1]);
+    if (jit_fpr_p(r1)) {
+	if (!jit_thumb_p() && jit_armv5e_p())
+	    LDRDIN(_R2_REGNO, _FP_REGNO, swf_off(r1) + 8);
+	else {
+	    swf_ldrin(_R2_REGNO, _FP_REGNO, swf_off(r1) + 8);
+	    swf_ldrin(_R3_REGNO, _FP_REGNO, swf_off(r1) + 4);
+	}
+    }
+    else {
+	movr(_R2_REGNO, r1);
+	movr(_R3_REGNO, r1 + 1);
+    }
+    swf_call_with_get_reg(__aeabi_dsub, fallback);
+    if (jit_fpr_p(r0)) {
+	if (!jit_thumb_p() && jit_armv5e_p())
+	    STRDIN(_R0_REGNO, _FP_REGNO, swf_off(r0) + 8);
+	else {
+	    swf_strin(_R0_REGNO, _FP_REGNO, swf_off(r0) + 8);
+	    swf_strin(_R1_REGNO, _FP_REGNO, swf_off(r0) + 4);
+	}
+    }
+    else {
+	movr(r0, _R0_REGNO);
+	movr(r0 + 1, _R1_REGNO);
+    }
+    jit_unget_reg_args();
+}
+
 static void
 _swf_iff(jit_state_t *_jit, int (*i0)(float, float),
 	 jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
diff --git a/lib/jit_arm-sz.c b/lib/jit_arm-sz.c
index 988cc0a36..5d4502893 100644
--- a/lib/jit_arm-sz.c
+++ b/lib/jit_arm-sz.c
@@ -24,6 +24,7 @@
     8,	/* subci */
     4,	/* subxr */
     4,	/* subxi */
+    12,	/* rsbi */
     4,	/* mulr */
     8,	/* muli */
     4,	/* qmulr */
@@ -179,6 +180,7 @@
     8,	/* addi_f */
     4,	/* subr_f */
     8,	/* subi_f */
+    8,	/* rsbi_f */
     4,	/* mulr_f */
     8,	/* muli_f */
     4,	/* divr_f */
@@ -261,6 +263,7 @@
     20,	/* addi_d */
     4,	/* subr_d */
     20,	/* subi_d */
+    20,	/* rsbi_d */
     4,	/* mulr_d */
     20,	/* muli_d */
     4,	/* divr_d */
@@ -377,6 +380,7 @@
     8,	/* subci */
     4,	/* subxr */
     4,	/* subxi */
+    12,	/* rsbi */
     4,	/* mulr */
     8,	/* muli */
     4,	/* qmulr */
@@ -532,6 +536,7 @@
     38,	/* addi_f */
     38,	/* subr_f */
     38,	/* subi_f */
+    38,	/* rsbi_f */
     38,	/* mulr_f */
     38,	/* muli_f */
     38,	/* divr_f */
@@ -614,6 +619,7 @@
     52,	/* addi_d */
     50,	/* subr_d */
     52,	/* subi_d */
+    52,	/* rsbi_d */
     50,	/* mulr_d */
     52,	/* muli_d */
     50,	/* divr_d */
diff --git a/lib/jit_arm-vfp.c b/lib/jit_arm-vfp.c
index 9b61ddaa0..27e0ccf4a 100644
--- a/lib/jit_arm-vfp.c
+++ b/lib/jit_arm-vfp.c
@@ -504,6 +504,12 @@ static void _vfp_subi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t);
 #  define vfp_subr_d(r0,r1,r2)		VSUB_F64(r0,r1,r2)
 #  define vfp_subi_d(r0,r1,i0)		_vfp_subi_d(_jit,r0,r1,i0)
 static void _vfp_subi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t);
+#  define vfp_rsbr_f(r0,r1,r2)		vfp_subr_f(r0,r2,r1)
+#  define vfp_rsbi_f(r0,r1,i0)		_vfp_rsbi_f(_jit,r0,r1,i0)
+static void _vfp_rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t);
+#  define vfp_rsbr_d(r0,r1,r2)		vfp_subr_d(r0,r2,r1)
+#  define vfp_rsbi_d(r0,r1,i0)		_vfp_rsbi_d(_jit,r0,r1,i0)
+static void _vfp_rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t);
 #  define vfp_mulr_f(r0,r1,r2)		VMUL_F32(r0,r1,r2)
 #  define vfp_muli_f(r0,r1,i0)		_vfp_muli_f(_jit,r0,r1,i0)
 static void _vfp_muli_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t);
@@ -1455,6 +1461,8 @@ _vfp_b##name##i_d(jit_state_t *_jit,					\
 fopi(add)
 dopi(add)
 fopi(sub)
+fopi(rsb)
+dopi(rsb)
 dopi(sub)
 fopi(mul)
 dopi(mul)
diff --git a/lib/jit_arm.c b/lib/jit_arm.c
index 179bf68f6..ed086dd5f 100644
--- a/lib/jit_arm.c
+++ b/lib/jit_arm.c
@@ -1074,6 +1074,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(subc,);
 		case_rrr(subx,);
 		case_rrw(subx,);
+		case_rrw(rsb,);
 		case_rrr(mul,);
 		case_rrw(mul,);
 		case_rrrr(qmul,);
@@ -1227,6 +1228,7 @@ _emit_code(jit_state_t *_jit)
 		case_vvf(add);
 		case_vvv(sub, _f);
 		case_vvf(sub);
+		case_vvf(rsb);
 		case_vvv(mul, _f);
 		case_vvf(mul);
 		case_vvv(div, _f);
@@ -1312,6 +1314,7 @@ _emit_code(jit_state_t *_jit)
 		case_vvd(add);
 		case_vvv(sub, _d);
 		case_vvd(sub);
+		case_vvd(rsb);
 		case_vvv(mul, _d);
 		case_vvd(mul);
 		case_vvv(div, _d);
diff --git a/lib/jit_hppa-cpu.c b/lib/jit_hppa-cpu.c
index 3fec62310..59008cbe1 100644
--- a/lib/jit_hppa-cpu.c
+++ b/lib/jit_hppa-cpu.c
@@ -677,6 +677,8 @@ static void _subci(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #define subxr(r0,r1,r2)		SUB_B(r1,r2,r0)
 #define subxi(r0,r1,i0)		_subxi(_jit,r0,r1,i0)
 static void _subxi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
+#define rsbi(r0, r1, i0)	_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #define mulr(r0,r1,r2)		_mulr(_jit,r0,r1,r2)
 static void _mulr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #define muli(r0,r1,i0)		_muli(_jit,r0,r1,i0)
@@ -1629,7 +1631,7 @@ static void
 _addi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {
     jit_int32_t		reg;
-    if (i0 >= -2048 && i0 <= 2047)
+    if (i0 >= -1024 && i0 <= 1023)
 	ADDI(i0, r1, r0);
     else {
 	reg = jit_get_reg(jit_class_gpr);
@@ -1689,6 +1691,20 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     jit_unget_reg(reg);
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    jit_int32_t		reg;
+    if (i0 >= -1024 && i0 <= 1023)
+	SUBI(i0, r1, r0);
+    else {
+	reg = jit_get_reg(jit_class_gpr);
+	movi(rn(reg), i0);
+	subr(r0, rn(reg), r1);
+	jit_unget_reg(reg);
+    }
+}
+
 static void
 _mulr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
 {
diff --git a/lib/jit_hppa-fpu.c b/lib/jit_hppa-fpu.c
index d1b6efec0..6b6b8bfe5 100644
--- a/lib/jit_hppa-fpu.c
+++ b/lib/jit_hppa-fpu.c
@@ -301,6 +301,12 @@ static void _subi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
 #define subr_d(r0,r1,r2)		FSUB_D(r1,r2,r0)
 #define subi_d(r0,r1,i0)		_subi_d(_jit,r0,r1,i0)
 static void _subi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
+#define rsbr_f(r0,r1,r2)		subr_f(r0,r2,r1)
+#define rsbi_f(r0,r1,i0)		_rsbi_f(_jit,r0,r1,i0)
+static void _rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
+#define rsbr_d(r0,r1,r2)		subr_d(r0,r2,r1)
+#define rsbi_d(r0,r1,i0)		_rsbi_d(_jit,r0,r1,i0)
+static void _rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
 #define mulr_f(r0,r1,r2)		FMPY_S(r1,r2,r0)
 #define muli_f(r0,r1,i0)		_muli_f(_jit,r0,r1,i0)
 static void _muli_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
@@ -741,6 +747,8 @@ fopi(add)
 dopi(add)
 fopi(sub)
 dopi(sub)
+fopi(rsb)
+dopi(rsb)
 fopi(mul)
 dopi(mul)
 fopi(div)
diff --git a/lib/jit_hppa-sz.c b/lib/jit_hppa-sz.c
index 2c6916f8a..663162b88 100644
--- a/lib/jit_hppa-sz.c
+++ b/lib/jit_hppa-sz.c
@@ -23,6 +23,7 @@
     12,	/* subci */
     4,	/* subxr */
     8,	/* subxi */
+    16,	/* rsbi */
     28,	/* mulr */
     36,	/* muli */
     40,	/* qmulr */
@@ -178,6 +179,7 @@
     16,	/* addi_f */
     4,	/* subr_f */
     16,	/* subi_f */
+    16,	/* rsbi_f */
     4,	/* mulr_f */
     16,	/* muli_f */
     4,	/* divr_f */
@@ -260,6 +262,7 @@
     24,	/* addi_d */
     4,	/* subr_d */
     24,	/* subi_d */
+    24,	/* rsbi_d */
     4,	/* mulr_d */
     24,	/* muli_d */
     4,	/* divr_d */
diff --git a/lib/jit_hppa.c b/lib/jit_hppa.c
index 636f36403..bc277bcb3 100644
--- a/lib/jit_hppa.c
+++ b/lib/jit_hppa.c
@@ -749,6 +749,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(subc,);
 		case_rrr(subx,);
 		case_rrw(subx,);
+		case_rrw(rsb,);
 		case_rrr(mul,);
 		case_rrw(mul,);
 		case_rrrr(qmul,);
@@ -924,8 +925,10 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _d, 64);
 		case_rrr(sub, _f);
 		case_rrf(sub, _f, 32);
+		case_rrf(rsb, _f, 32);
 		case_rrr(sub, _d);
 		case_rrf(sub, _d, 64);
+		case_rrf(rsb, _d, 64);
 		case_rrr(mul, _f);
 		case_rrf(mul, _f, 32);
 		case_rrr(mul, _d);
diff --git a/lib/jit_ia64-cpu.c b/lib/jit_ia64-cpu.c
index bf701418d..9baeafab7 100644
--- a/lib/jit_ia64-cpu.c
+++ b/lib/jit_ia64-cpu.c
@@ -1197,6 +1197,8 @@ static void _subci(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 static void _subxr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define subxi(r0, r1, i0)		_subxi(_jit, r0, r1, i0)
 static void _subxi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
+#  define rsbi(r0, r1, i0)		_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #define mulr(r0,r1,r2)			_mulr(_jit,r0,r1,r2)
 static void _mulr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #define muli(r0,r1,i0)			_muli(_jit,r0,r1,i0)
@@ -3621,6 +3623,22 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     jit_unget_reg(reg);
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    jit_int32_t		reg;
+    if (i0 >= -128 && i0 <= 127)
+	SUBI(r0, i0, r1);
+    else if (!(r1 & ~3) && i0 >= -2097151 && i0 <= 2097152)
+	ADDL(r1, -i0, r0);
+    else {
+	reg = jit_get_reg(jit_class_gpr);
+	movi(rn(reg), i0);
+	subr(r0, rn(reg), r1);
+	jit_unget_reg(reg);
+    }
+}
+
 static void
 _mulr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
 {
diff --git a/lib/jit_ia64-fpu.c b/lib/jit_ia64-fpu.c
index 5557d7b5d..1c5057ca4 100644
--- a/lib/jit_ia64-fpu.c
+++ b/lib/jit_ia64-fpu.c
@@ -474,6 +474,12 @@ static void _subi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
 #define subr_d(r0,r1,r2)		FSUB_D(r0,r1,r2)
 #define subi_d(r0,r1,i0)		_subi_d(_jit,r0,r1,i0)
 static void _subi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
+#define rsbr_f(r0,r1,r2)		subr_f(r0,r2,r1)
+#define rsbi_f(r0,r1,i0)		_rsbi_f(_jit,r0,r1,i0)
+static void _rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
+#define rsbr_d(r0,r1,r2)		subr_d(r0,r2,r1)
+#define rsbi_d(r0,r1,i0)		_rsbi_d(_jit,r0,r1,i0)
+static void _rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
 #define mulr_f(r0,r1,r2)		FMPY_S(r0,r1,r2)
 #define muli_f(r0,r1,i0)		_muli_f(_jit,r0,r1,i0)
 static void _muli_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
@@ -1109,10 +1115,12 @@ _b##name##i_##type(jit_state_t *_jit,					\
 
 fopi(add)
 fopi(sub)
+fopi(rsb)
 fopi(mul)
 fopi(div)
 dopi(add)
 dopi(sub)
+dopi(rsb)
 dopi(mul)
 dopi(div)
 
diff --git a/lib/jit_ia64-sz.c b/lib/jit_ia64-sz.c
index 4711f6160..8059d8cbe 100644
--- a/lib/jit_ia64-sz.c
+++ b/lib/jit_ia64-sz.c
@@ -23,6 +23,7 @@
     48,	/* subci */
     64,	/* subxr */
     64,	/* subxi */
+    48,	/* rsbi */
     32,	/* mulr */
     48,	/* muli */
     96,	/* qmulr */
@@ -178,6 +179,7 @@
     32,	/* addi_f */
     16,	/* subr_f */
     16,	/* subi_f */
+    16,	/* rsbi_f */
     0,	/* mulr_f */
     16,	/* muli_f */
     144,	/* divr_f */
@@ -260,6 +262,7 @@
     32,	/* addi_d */
     16,	/* subr_d */
     16,	/* subi_d */
+    16,	/* rsbi_d */
     0,	/* mulr_d */
     16,	/* muli_d */
     128,	/* divr_d */
diff --git a/lib/jit_ia64.c b/lib/jit_ia64.c
index e8eee8a03..f35f396dc 100644
--- a/lib/jit_ia64.c
+++ b/lib/jit_ia64.c
@@ -884,6 +884,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(sub,);
 		case_rrr(subx,);
 		case_rrw(subx,);
+		case_rrw(rsb,);
 		case_rrr(subc,);
 		case_rrw(subc,);
 		case_rrr(mul,);
@@ -1055,6 +1056,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _f, 32);
 		case_rrr(sub, _f);
 		case_rrf(sub, _f, 32);
+		case_rrf(rsb, _f, 32);
 		case_rrr(mul, _f);
 		case_rrf(mul, _f, 32);
 		case_rrr(div, _f);
@@ -1137,6 +1139,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _d, 64);
 		case_rrr(sub, _d);
 		case_rrf(sub, _d, 64);
+		case_rrf(rsb, _d, 64);
 		case_rrr(mul, _d);
 		case_rrf(mul, _d, 64);
 		case_rrr(div, _d);
diff --git a/lib/jit_mips-cpu.c b/lib/jit_mips-cpu.c
index 9e9778fc7..1f48ecf08 100644
--- a/lib/jit_mips-cpu.c
+++ b/lib/jit_mips-cpu.c
@@ -419,6 +419,8 @@ static void _subci(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 static void _subxr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define subxi(r0,r1,i0)		_subxi(_jit,r0,r1,i0)
 static void _subxi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
+#  define rsbi(r0, r1, i0)		_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #  define mulr(r0,r1,r2)		_mulr(_jit,r0,r1,r2)
 static void _mulr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define muli(r0,r1,i0)		_muli(_jit,r0,r1,i0)
@@ -959,6 +961,13 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     jit_unget_reg(t0);
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    subi(r0, r1, i0);
+    negr(r0, r0);
+}
+
 static void
 _mulr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
 {
diff --git a/lib/jit_mips-fpu.c b/lib/jit_mips-fpu.c
index bd83103b8..ff31609b1 100644
--- a/lib/jit_mips-fpu.c
+++ b/lib/jit_mips-fpu.c
@@ -193,6 +193,12 @@ static void _subi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
 #  define subr_d(r0,r1,r2)		SUB_D(r0,r1,r2)
 #  define subi_d(r0,r1,i0)		_subi_d(_jit,r0,r1,i0)
 static void _subi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
+#  define rsbr_f(r0,r1,r2)		subr_f(r0,r2,r1)
+#  define rsbi_f(r0,r1,i0)		_rsbi_f(_jit,r0,r1,i0)
+static void _rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
+#  define rsbr_d(r0,r1,r2)		subr_d(r0,r2,r1)
+#  define rsbi_d(r0,r1,i0)		_rsbi_d(_jit,r0,r1,i0)
+static void _rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
 #  define mulr_f(r0,r1,r2)		MUL_S(r0,r1,r2)
 #  define muli_f(r0,r1,i0)		_muli_f(_jit,r0,r1,i0)
 static void _muli_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
@@ -585,6 +591,7 @@ _b##name##i_##type(jit_state_t *_jit,					\
 
 fopi(add)
 fopi(sub)
+fopi(rsb)
 fopi(mul)
 fopi(div)
 
@@ -746,6 +753,7 @@ _movi_f(jit_state_t *_jit, jit_int32_t r0, jit_float32_t *i0)
 
 dopi(add)
 dopi(sub)
+dopi(rsb)
 dopi(mul)
 dopi(div)
 
diff --git a/lib/jit_mips-sz.c b/lib/jit_mips-sz.c
index 2efe8f822..517bedc28 100644
--- a/lib/jit_mips-sz.c
+++ b/lib/jit_mips-sz.c
@@ -24,6 +24,7 @@
     20,	/* subci */
     28,	/* subxr */
     28,	/* subxi */
+    16,	/* rsbi */
     8,	/* mulr */
     16,	/* muli */
     12,	/* qmulr */
@@ -179,6 +180,7 @@
     16,	/* addi_f */
     4,	/* subr_f */
     16,	/* subi_f */
+    16,	/* rsbi_f */
     4,	/* mulr_f */
     16,	/* muli_f */
     4,	/* divr_f */
@@ -261,6 +263,7 @@
     16,	/* addi_d */
     4,	/* subr_d */
     16,	/* subi_d */
+    16,	/* rsbi_d */
     4,	/* mulr_d */
     16,	/* muli_d */
     4,	/* divr_d */
@@ -377,6 +380,7 @@
     20,	/* subci */
     28,	/* subxr */
     28,	/* subxi */
+    16,	/* rsbi */
     8,	/* mulr */
     16,	/* muli */
     12,	/* qmulr */
@@ -532,6 +536,7 @@
     16,	/* addi_f */
     4,	/* subr_f */
     16,	/* subi_f */
+    16,	/* rsbi_f */
     4,	/* mulr_f */
     16,	/* muli_f */
     4,	/* divr_f */
@@ -614,6 +619,7 @@
     20,	/* addi_d */
     4,	/* subr_d */
     20,	/* subi_d */
+    20,	/* rsbi_d */
     4,	/* mulr_d */
     20,	/* muli_d */
     4,	/* divr_d */
@@ -729,6 +735,7 @@
     36,	/* subci */
     28,	/* subxr */
     28,	/* subxi */
+    32,	/* rsbi */
     8,	/* mulr */
     32,	/* muli */
     12,	/* qmulr */
@@ -884,6 +891,7 @@
     16,	/* addi_f */
     4,	/* subr_f */
     16,	/* subi_f */
+    16,	/* rsbi_f */
     4,	/* mulr_f */
     16,	/* muli_f */
     4,	/* divr_f */
@@ -966,6 +974,7 @@
     16,	/* addi_d */
     4,	/* subr_d */
     16,	/* subi_d */
+    16,	/* rsbi_d */
     4,	/* mulr_d */
     16,	/* muli_d */
     4,	/* divr_d */
diff --git a/lib/jit_mips.c b/lib/jit_mips.c
index c31d43bfa..aed730c8c 100644
--- a/lib/jit_mips.c
+++ b/lib/jit_mips.c
@@ -1005,6 +1005,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(subc,);
 		case_rrr(subx,);
 		case_rrw(subx,);
+		case_rrw(rsb,);
 		case_rrr(mul,);
 		case_rrw(mul,);
 		case_rrrr(qmul,);
@@ -1186,6 +1187,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _f, 32);
 		case_rrr(sub, _f);
 		case_rrf(sub, _f, 32);
+		case_rrf(rsb, _f, 32);
 		case_rrr(mul, _f);
 		case_rrf(mul, _f, 32);
 		case_rrr(div, _f);
@@ -1268,6 +1270,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _d, 64);
 		case_rrr(sub, _d);
 		case_rrf(sub, _d, 64);
+		case_rrf(rsb, _d, 64);
 		case_rrr(mul, _d);
 		case_rrf(mul, _d, 64);
 		case_rrr(div, _d);
diff --git a/lib/jit_names.c b/lib/jit_names.c
index 171fbfcd5..6af443383 100644
--- a/lib/jit_names.c
+++ b/lib/jit_names.c
@@ -31,6 +31,7 @@ static char *code_name[] = {
     "subr",		"subi",
     "subcr",		"subci",
     "subxr",		"subxi",
+    "rsbi",
     "mulr",		"muli",
     "qmulr",		"qmuli",
     "qmulr_u",		"qmuli_u",
diff --git a/lib/jit_ppc-cpu.c b/lib/jit_ppc-cpu.c
index 844c8b132..2f65d36b5 100644
--- a/lib/jit_ppc-cpu.c
+++ b/lib/jit_ppc-cpu.c
@@ -447,10 +447,10 @@ static void _FXS(jit_state_t*,int,int,int,int,int,int,int);
 #  define TWI(t,a,s)			FDs(3,t,a,s)
 #  define TWGTI(a,s)			TWI(8,a,s)
 #  define TWLLEI(a,s)			TWI(6,a,s)
-#  define XOR(d,a,b)			FXO(31,a,d,b,0,316)
-#  define XOR_(d,a,b)			FXO_(31,a,d,b,0,316)
-#  define XORI(s,a,u)			FDu(26,s,a,u)
-#  define XORIS(s,a,u)			FDu(27,s,a,u)
+#  define XOR(d,a,b)			FX(31,a,d,b,316)
+#  define XOR_(d,a,b)			FX_(31,a,d,b,316)
+#  define XORI(s,a,u)			FDu(26,a,s,u)
+#  define XORIS(s,a,u)			FDu(27,a,s,u)
 #  define nop(c)			_nop(_jit,c)
 static void _nop(jit_state_t*,jit_int32_t);
 #  define movr(r0,r1)			_movr(_jit,r0,r1)
@@ -492,6 +492,8 @@ static void _subci(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #  define subxr(r0,r1,r2)		SUBFE(r0,r2,r1)
 #  define subxi(r0,r1,i0)		_subxi(_jit,r0,r1,i0)
 static void _subxi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
+#  define rsbi(r0, r1, i0)		_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #  if __WORDSIZE == 32
 #    define mulr(r0,r1,r2)		MULLW(r0,r1,r2)
 #    define mullr(r0,r1,r2)		MULLW(r0,r1,r2)
@@ -1119,6 +1121,13 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     jit_unget_reg(reg);
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    subi(r0, r1, i0);
+    negr(r0, r0);
+}
+
 static void
 _muli(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {
diff --git a/lib/jit_ppc-fpu.c b/lib/jit_ppc-fpu.c
index fcfca82eb..4bef6c729 100644
--- a/lib/jit_ppc-fpu.c
+++ b/lib/jit_ppc-fpu.c
@@ -24,7 +24,6 @@ static void _FA(jit_state_t*,int,int,int,int,int,int,int);
 #define FXFL(o,m,b,x)			_FXFL(_jit,o,m,b,x,0)
 #define FXFL_(o,m,b,x)			_FXFL(_jit,o,m,b,x,1)
 static void _FXFL(jit_state_t*,int,int,int,int,int) maybe_unused;
-
 #  define FABS(d,b)			FX(63,d,0,b,264)
 #  define FABS_(d,b)			FX_(63,d,0,b,264)
 #  define FADD(d,a,b)			FA(63,d,a,b,0,21)
@@ -116,7 +115,6 @@ static void _FXFL(jit_state_t*,int,int,int,int,int) maybe_unused;
 #  define STFSU(s,a,d)			FDs(53,s,a,d)
 #  define STFSUX(s,a,b)			FX(31,s,a,b,695)
 #  define STFSX(s,a,b)			FX(31,s,a,b,663)
-
 #  define movr_f(r0,r1)			movr_d(r0,r1)
 #  define movr_d(r0,r1)			_movr_d(_jit,r0,r1)
 static void _movr_d(jit_state_t*,jit_int32_t,jit_int32_t);
@@ -127,7 +125,6 @@ static void _movi_d(jit_state_t*,jit_int32_t,jit_float64_t*);
 #  define extr_f(r0,r1)			extr_d(r0,r1)
 #  define extr_d(r0,r1)			_extr_d(_jit,r0,r1)
 static void _extr_d(jit_state_t*,jit_int32_t,jit_int32_t);
-
 #  define truncr_f(r0,r1)		truncr_d(r0,r1)
 #  define truncr_f_i(r0,r1)		truncr_d_i(r0,r1)
 #  define truncr_d_i(r0,r1)		_truncr_d_i(_jit,r0,r1)
@@ -140,17 +137,14 @@ static void _truncr_d_i(jit_state_t*,jit_int32_t,jit_int32_t);
 #    define truncr_d_l(r0,r1)		_truncr_d_l(_jit,r0,r1)
 static void _truncr_d_l(jit_state_t*,jit_int32_t,jit_int32_t);
 #  endif
-
 #  define extr_d_f(r0,r1)		FRSP(r0,r1)
 #  define extr_f_d(r0,r1)		movr_d(r0,r1)
-
 #  define absr_f(r0,r1)			absr_d(r0,r1)
 #  define absr_d(r0,r1)			FABS(r0,r1)
 #  define negr_f(r0,r1)			negr_d(r0,r1)
 #  define negr_d(r0,r1)			FNEG(r0,r1)
 #  define sqrtr_f(r0,r1)		FSQRTS(r0,r1)
 #  define sqrtr_d(r0,r1)		FSQRT(r0,r1)
-
 #  define addr_f(r0,r1,r2)		FADDS(r0,r1,r2)
 #  define addr_d(r0,r1,r2)		FADD(r0,r1,r2)
 #  define addi_f(r0,r1,i0)		_addi_f(_jit,r0,r1,i0)
@@ -163,6 +157,12 @@ static void _subi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
 #  define subr_d(r0,r1,r2)		FSUB(r0,r1,r2)
 #  define subi_d(r0,r1,i0)		_subi_d(_jit,r0,r1,i0)
 static void _subi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
+#  define rsbr_f(r0,r1,r2)		subr_f(r0,r2,r1)
+#  define rsbi_f(r0,r1,i0)		_rsbi_f(_jit,r0,r1,i0)
+static void _rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
+#  define rsbr_d(r0,r1,r2)		subr_d(r0,r2,r1)
+#  define rsbi_d(r0,r1,i0)		_rsbi_d(_jit,r0,r1,i0)
+static void _rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
 #  define mulr_f(r0,r1,r2)		FMULS(r0,r1,r2)
 #  define muli_f(r0,r1,i0)		_muli_f(_jit,r0,r1,i0)
 static void _muli_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
@@ -175,7 +175,6 @@ static void _divi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
 #  define divr_d(r0,r1,r2)		FDIV(r0,r1,r2)
 #  define divi_d(r0,r1,i0)		_divi_d(_jit,r0,r1,i0)
 static void _divi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
-
 #  define ltr_f(r0,r1,r2)		ltr_d(r0,r1,r2)
 #  define ltr_d(r0,r1,r2)		_ltr_d(_jit,r0,r1,r2)
 static void _ltr_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
@@ -372,7 +371,6 @@ static jit_word_t _bunordr_d(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
 static jit_word_t _bunordi_f(jit_state_t*,jit_word_t,jit_int32_t,jit_float32_t*);
 #  define bunordi_d(i0,r0,i1)		_bunordi_d(_jit,i0,r0,i1)
 static jit_word_t _bunordi_d(jit_state_t*,jit_word_t,jit_int32_t,jit_float64_t*);
-
 #  define ldr_f(r0,r1)			LFSX(r0, _R0_REGNO, r1)
 #  define ldi_f(r0,i0)			_ldi_f(_jit,r0,i0)
 static void _ldi_f(jit_state_t*,jit_int32_t,jit_word_t);
@@ -565,6 +563,8 @@ fopi(add)
 dopi(add)
 fopi(sub)
 dopi(sub)
+fopi(rsb)
+dopi(rsb)
 fopi(mul)
 dopi(mul)
 fopi(div)
diff --git a/lib/jit_ppc-sz.c b/lib/jit_ppc-sz.c
index 7fae39436..ba7881306 100644
--- a/lib/jit_ppc-sz.c
+++ b/lib/jit_ppc-sz.c
@@ -24,6 +24,7 @@
     12,	/* subci */
     4,	/* subxr */
     8,	/* subxi */
+    16,	/* rsbi */
     4,	/* mulr */
     12,	/* muli */
     12,	/* qmulr */
@@ -179,6 +180,7 @@
     12,	/* addi_f */
     4,	/* subr_f */
     12,	/* subi_f */
+    12,	/* rsbi_f */
     4,	/* mulr_f */
     12,	/* muli_f */
     4,	/* divr_f */
@@ -261,6 +263,7 @@
     12,	/* addi_d */
     4,	/* subr_d */
     12,	/* subi_d */
+    12,	/* rsbi_d */
     4,	/* mulr_d */
     12,	/* muli_d */
     4,	/* divr_d */
@@ -377,6 +380,7 @@
     12,	/* subci */
     4,	/* subxr */
     8,	/* subxi */
+    16,	/* rsbi */
     4,	/* mulr */
     12,	/* muli */
     12,	/* qmulr */
@@ -532,6 +536,7 @@
     16,	/* addi_f */
     4,	/* subr_f */
     16,	/* subi_f */
+    16,	/* rsbi_f */
     4,	/* mulr_f */
     16,	/* muli_f */
     4,	/* divr_f */
@@ -614,6 +619,7 @@
     24,	/* addi_d */
     4,	/* subr_d */
     24,	/* subi_d */
+    24,	/* rsbi_d */
     4,	/* mulr_d */
     24,	/* muli_d */
     4,	/* divr_d */
@@ -730,6 +736,7 @@
     28,	/* subci */
     4,	/* subxr */
     8,	/* subxi */
+    32,	/* rsbi */
     4,	/* mulr */
     28,	/* muli */
     12,	/* qmulr */
@@ -885,6 +892,7 @@
     28,	/* addi_f */
     4,	/* subr_f */
     28,	/* subi_f */
+    28,	/* rsbi_f */
     4,	/* mulr_f */
     28,	/* muli_f */
     4,	/* divr_f */
@@ -967,6 +975,7 @@
     28,	/* addi_d */
     4,	/* subr_d */
     28,	/* subi_d */
+    28,	/* rsbi_d */
     4,	/* mulr_d */
     28,	/* muli_d */
     4,	/* divr_d */
diff --git a/lib/jit_ppc.c b/lib/jit_ppc.c
index 88488f091..cbc5ddaab 100644
--- a/lib/jit_ppc.c
+++ b/lib/jit_ppc.c
@@ -902,6 +902,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(subc,);
 		case_rrr(subx,);
 		case_rrw(subx,);
+		case_rrw(rsb,);
 		case_rrr(mul,);
 		case_rrw(mul,);
 		case_rrrr(qmul,);
@@ -1089,6 +1090,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _f, 32);
 		case_rrr(sub, _f);
 		case_rrf(sub, _f, 32);
+		case_rrf(rsb, _f, 32);
 		case_rrr(mul, _f);
 		case_rrf(mul, _f, 32);
 		case_rrr(div, _f);
@@ -1171,6 +1173,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _d, 64);
 		case_rrr(sub, _d);
 		case_rrf(sub, _d, 64);
+		case_rrf(rsb, _d, 64);
 		case_rrr(mul, _d);
 		case_rrf(mul, _d, 64);
 		case_rrr(div, _d);
diff --git a/lib/jit_s390x-cpu.c b/lib/jit_s390x-cpu.c
index dfa4e56e5..0030bac02 100644
--- a/lib/jit_s390x-cpu.c
+++ b/lib/jit_s390x-cpu.c
@@ -934,6 +934,8 @@ static void _subci(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 static void _subxr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define subxi(r0,r1,i0)		_subxi(_jit,r0,r1,i0)
 static void _subxi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
+#  define rsbi(r0, r1, i0)		_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #  define mulr(r0,r1,r2)		_mulr(_jit,r0,r1,r2)
 static void _mulr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define muli(r0,r1,i0)		_muli(_jit,r0,r1,i0)
@@ -2446,6 +2448,13 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     jit_unget_reg(reg);
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    subi(r0, r1, i0);
+    negr(r0, r0);
+}
+
 static void
 _mulr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
 {
diff --git a/lib/jit_s390x-fpu.c b/lib/jit_s390x-fpu.c
index 25c2b6ff5..d36f0fbfc 100644
--- a/lib/jit_s390x-fpu.c
+++ b/lib/jit_s390x-fpu.c
@@ -372,6 +372,10 @@ static void _subr_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define subr_d(r0,r1,r2)		_subr_d(_jit,r0,r1,r2)
 static void _subr_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define subi_d(r0,r1,i0)		dp(sub,r0,r1,i0)
+#  define rsbr_f(r0,r1,r2)		subr_f(r0,r2,r1)
+#  define rsbi_f(r0,r1,i0)		fp(rsb,r0,r1,i0)
+#  define rsbr_d(r0,r1,r2)		subr_d(r0,r2,r1)
+#  define rsbi_d(r0,r1,i0)		dp(rsb,r0,r1,i0)
 #  define mulr_f(r0,r1,r2)		_mulr_f(_jit,r0,r1,r2)
 static void _mulr_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define muli_f(r0,r1,i0)		fp(mul,r0,r1,i0)
@@ -597,6 +601,7 @@ _fp(jit_state_t *_jit, jit_code_t code,
     switch (code) {
 	case jit_code_addi_f:	addr_f(r0, r1, rn(reg));	break;
 	case jit_code_subi_f:	subr_f(r0, r1, rn(reg));	break;
+	case jit_code_rsbi_f:	rsbr_f(r0, r1, rn(reg));	break;
 	case jit_code_muli_f:	mulr_f(r0, r1, rn(reg));	break;
 	case jit_code_divi_f:	divr_f(r0, r1, rn(reg));	break;
 	case jit_code_uneqi_f:	uneqr_f(r0, r1, rn(reg));	break;
@@ -616,6 +621,7 @@ _dp(jit_state_t *_jit, jit_code_t code,
     switch (code) {
 	case jit_code_addi_d:	addr_d(r0, r1, rn(reg));	break;
 	case jit_code_subi_d:	subr_d(r0, r1, rn(reg));	break;
+	case jit_code_rsbi_d:	rsbr_d(r0, r1, rn(reg));	break;
 	case jit_code_muli_d:	mulr_d(r0, r1, rn(reg));	break;
 	case jit_code_divi_d:	divr_d(r0, r1, rn(reg));	break;
 	case jit_code_uneqi_d:	uneqr_d(r0, r1, rn(reg));	break;
diff --git a/lib/jit_s390x-sz.c b/lib/jit_s390x-sz.c
index a80b97346..a933e1553 100644
--- a/lib/jit_s390x-sz.c
+++ b/lib/jit_s390x-sz.c
@@ -23,6 +23,7 @@
     20,	/* subci */
     12,	/* subxr */
     12,	/* subxi */
+    28,	/* rsbi */
     8,	/* mulr */
     24,	/* muli */
     60,	/* qmulr */
@@ -178,6 +179,7 @@
     26,	/* addi_f */
     8,	/* subr_f */
     26,	/* subi_f */
+    26,	/* rsbi_f */
     6,	/* mulr_f */
     26,	/* muli_f */
     8,	/* divr_f */
@@ -260,6 +262,7 @@
     26,	/* addi_d */
     8,	/* subr_d */
     26,	/* subi_d */
+    26,	/* rsbi_d */
     6,	/* mulr_d */
     26,	/* muli_d */
     8,	/* divr_d */
diff --git a/lib/jit_s390x.c b/lib/jit_s390x.c
index 4c45db0b0..460d793c5 100644
--- a/lib/jit_s390x.c
+++ b/lib/jit_s390x.c
@@ -727,6 +727,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(subc,);
 		case_rrr(subx,);
 		case_rrw(subx,);
+		case_rrw(rsb,);
 		case_rrr(mul,);
 		case_rrw(mul,);
 		case_rrrr(qmul,);
@@ -896,6 +897,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add);
 		case_rrr(sub, _f);
 		case_rrf(sub);
+		case_rrf(rsb);
 		case_rrr(mul, _f);
 		case_rrf(mul);
 		case_rrr(div, _f);
@@ -978,6 +980,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrd(add);
 		case_rrr(sub, _d);
 		case_rrd(sub);
+		case_rrd(rsb);
 		case_rrr(mul, _d);
 		case_rrd(mul);
 		case_rrr(div, _d);
diff --git a/lib/jit_sparc-cpu.c b/lib/jit_sparc-cpu.c
index 095c36466..96150d3f9 100644
--- a/lib/jit_sparc-cpu.c
+++ b/lib/jit_sparc-cpu.c
@@ -452,6 +452,8 @@ static void _subci(jit_state_t*, jit_int32_t, jit_int32_t, jit_word_t);
 #  define subxr(r0, r1, r2)		SUBXcc(r1, r2, r0)
 #  define subxi(r0, r1, i0)		_subxi(_jit, r0, r1, i0)
 static void _subxi(jit_state_t*, jit_int32_t, jit_int32_t, jit_word_t);
+#  define rsbi(r0, r1, i0)		_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #  define mulr(r0, r1, r2)		UMUL(r1, r2, r0)
 #  define muli(r0, r1, i0)		_muli(_jit, r0, r1, i0)
 static void _muli(jit_state_t*, jit_int32_t, jit_int32_t, jit_word_t);
@@ -927,6 +929,13 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     }
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    subi(r0, r1, i0);
+    negr(r0, r0);
+}
+
 static void
 _muli(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {
@@ -1325,7 +1334,7 @@ _ldxi_s(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     else {
 	reg = jit_get_reg(jit_class_gpr);
 	movi(rn(reg), i0);
-	ldxr_c(r0, r1, rn(reg));
+	ldxr_s(r0, r1, rn(reg));
 	jit_unget_reg(reg);
     }
 }
diff --git a/lib/jit_sparc-fpu.c b/lib/jit_sparc-fpu.c
index 52d760d9f..dc4d574de 100644
--- a/lib/jit_sparc-fpu.c
+++ b/lib/jit_sparc-fpu.c
@@ -143,7 +143,6 @@ _f3f(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t, jit_int32_t,jit_int32_t);
 #  define FCMPEQ(rs1, rs2)		FPop2(0, rs1, SPARC_FCMPEQ, rs2)
 #  define CPop1(rd, rs1, opc, rs2)	f3f(rd, 54, rs1, opf, rs2)
 #  define CPop2(rd, rs1, opc, rs2)	f3f(rd, 55, rs1, opf, rs2)
-
 #  define extr_f(r0, r1)		_extr_f(_jit, r0, r1)
 static void _extr_f(jit_state_t*, jit_int32_t, jit_int32_t);
 #  define truncr_f(r0, r1)		truncr_f_i(r0, r1)
@@ -156,7 +155,6 @@ static void _movi_f(jit_state_t*, jit_int32_t, jit_float32_t*);
 #  define negr_f(r0, r1)		FNEGS(r1, r0)
 #  define absr_f(r0, r1)		FABSS(r1, r0)
 #  define sqrtr_f(r0, r1)		FSQRTS(r1, r0)
-
 #  define extr_d(r0, r1)		_extr_d(_jit, r0, r1)
 static void _extr_d(jit_state_t*, jit_int32_t, jit_int32_t);
 #  define truncr_d(r0, r1)		truncr_d_i(r0, r1)
@@ -172,32 +170,40 @@ static void _negr_d(jit_state_t*, jit_int32_t, jit_int32_t);
 #  define absr_d(r0, r1)		_absr_d(_jit, r0, r1)
 static void _absr_d(jit_state_t*, jit_int32_t, jit_int32_t);
 #  define sqrtr_d(r0, r1)		FSQRTD(r1, r0)
-
 #  define fop1f(op, r0, r1, i0)		_fop1f(_jit, op, r0, r1, i0)
 static void _fop1f(jit_state_t*,jit_int32_t,
 		   jit_int32_t,jit_int32_t,jit_float32_t*);
+#  define rfop1f(op, r0, r1, i0)	_rfop1f(_jit, op, r0, r1, i0)
+static void _rfop1f(jit_state_t*,jit_int32_t,
+		    jit_int32_t,jit_int32_t,jit_float32_t*);
 #  define fop1d(op, r0, r1, i0)		_fop1d(_jit, op, r0, r1, i0)
 static void _fop1d(jit_state_t*,jit_int32_t,
 		   jit_int32_t,jit_int32_t,jit_float64_t*);
-
+#  define rfop1d(op, r0, r1, i0)	_rfop1d(_jit, op, r0, r1, i0)
+static void _rfop1d(jit_state_t*,jit_int32_t,
+		    jit_int32_t,jit_int32_t,jit_float64_t*);
 #  define addr_f(r0, r1, r2)		FADDS(r1, r2, r0)
 #  define addi_f(r0, r1, i0)		fop1f(SPARC_FADDS, r0, r1, i0)
 #  define subr_f(r0, r1, r2)		FSUBS(r1, r2, r0)
 #  define subi_f(r0, r1, i0)		fop1f(SPARC_FSUBS, r0, r1, i0)
+#  define rsbr_f(r0, r1, r2)		subr_f(r0, r2, r1)
+#  define rsbi_f(r0, r1, i0)		rfop1f(SPARC_FSUBS, r0, r1, i0)
+#  define rsbr_d(r0, r1, r2)		subr_d(r0, r2, r1)
+#  define rsbi_d(r0, r1, i0)		rfop1d(SPARC_FSUBD, r0, r1, i0)
 #  define mulr_f(r0, r1, r2)		FMULS(r1, r2, r0)
 #  define muli_f(r0, r1, i0)		fop1f(SPARC_FMULS, r0, r1, i0)
 #  define divr_f(r0, r1, r2)		FDIVS(r1, r2, r0)
 #  define divi_f(r0, r1, i0)		fop1f(SPARC_FDIVS, r0, r1, i0)
-
 #  define addr_d(r0, r1, r2)		FADDD(r1, r2, r0)
 #  define addi_d(r0, r1, i0)		fop1d(SPARC_FADDD, r0, r1, i0)
 #  define subr_d(r0, r1, r2)		FSUBD(r1, r2, r0)
 #  define subi_d(r0, r1, i0)		fop1d(SPARC_FSUBD, r0, r1, i0)
+#  define rsbr_d(r0, r1, r2)		subr_d(r0, r2, r1)
+#  define rsbi_d(r0, r1, i0)		rfop1d(SPARC_FSUBD, r0, r1, i0)
 #  define mulr_d(r0, r1, r2)		FMULD(r1, r2, r0)
 #  define muli_d(r0, r1, i0)		fop1d(SPARC_FMULD, r0, r1, i0)
 #  define divr_d(r0, r1, r2)		FDIVD(r1, r2, r0)
 #  define divi_d(r0, r1, i0)		fop1d(SPARC_FDIVD, r0, r1, i0)
-
 #define fcr(cc, r0, r1, r2)		_fcr(_jit, cc, r0, r1, r2)
 static void _fcr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t,jit_int32_t);
 #define fcw(cc, r0, r1, i0)		_fcw(_jit, cc, r0, r1, i0)
@@ -231,7 +237,6 @@ _fcw(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t,jit_float32_t*);
 #  define ordi_f(r0, r1, i0)		fcw(SPARC_FBO, r0, r1, i0)
 #  define unordr_f(r0, r1, r2)		fcr(SPARC_FBU, r0, r1, r2)
 #  define unordi_f(r0, r1, i0)		fcw(SPARC_FBU, r0, r1, i0)
-
 #define dcr(cc, r0, r1, r2)		_dcr(_jit, cc, r0, r1, r2)
 static void _dcr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t,jit_int32_t);
 #define dcw(cc, r0, r1, i0)		_dcw(_jit, cc, r0, r1, i0)
@@ -265,7 +270,6 @@ _dcw(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t,jit_float64_t*);
 #  define ordi_d(r0, r1, i0)		dcw(SPARC_FBO, r0, r1, i0)
 #  define unordr_d(r0, r1, r2)		dcr(SPARC_FBU, r0, r1, r2)
 #  define unordi_d(r0, r1, i0)		dcw(SPARC_FBU, r0, r1, i0)
-
 #  define ldr_f(r0, r1)			LDF(r1, 0, r0)
 #  define ldi_f(r0, i0)			_ldi_f(_jit, r0, i0)
 static void _ldi_f(jit_state_t*,jit_int32_t,jit_word_t);
@@ -278,7 +282,6 @@ static void _sti_f(jit_state_t*,jit_int32_t,jit_word_t);
 #  define stxr_f(r0, r1, r2)		STF(r2, r1, r0)
 #  define stxi_f(r0, r1, i0)		_stxi_f(_jit, r0, r1, i0)
 static void _stxi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
-
 #  define ldr_d(r0, r1)			LDDF(r1, 0, r0)
 #  define ldi_d(r0, i0)			_ldi_d(_jit, r0, i0)
 static void _ldi_d(jit_state_t*,jit_int32_t,jit_word_t);
@@ -291,7 +294,6 @@ static void _sti_d(jit_state_t*,jit_int32_t,jit_word_t);
 #  define stxr_d(r0, r1, r2)		STDF(r2, r1, r0)
 #  define stxi_d(r0, r1, i0)		_stxi_d(_jit, r0, r1, i0)
 static void _stxi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
-
 #  define fbr(cc, i0, r0, r1)		_fbr(_jit, cc, i0, r0, r1)
 static jit_word_t
 _fbr(jit_state_t*,jit_int32_t,jit_word_t,jit_int32_t,jit_int32_t);
@@ -360,7 +362,6 @@ _dbw(jit_state_t*,jit_int32_t,jit_word_t,jit_int32_t,jit_float64_t*);
 #  define bordi_d(i0, r0, i1)		dbw(SPARC_FBO, i0, r0, i1)
 #  define bunordr_d(i0, r0, r1)		dbr(SPARC_FBU, i0, r0, r1)
 #  define bunordi_d(i0, r0, i1)		dbw(SPARC_FBU, i0, r0, i1)
-
 #endif
 
 #if CODE
@@ -469,6 +470,17 @@ _fop1f(jit_state_t *_jit, jit_int32_t op,
     jit_unget_reg(reg);
 }
 
+static void
+_rfop1f(jit_state_t *_jit, jit_int32_t op,
+	jit_int32_t r0, jit_int32_t r1, jit_float32_t *i0)
+{
+    jit_int32_t		reg;
+    reg = jit_get_reg(jit_class_fpr);
+    movi_f(rn(reg), i0);
+    FPop1(r0, rn(reg), op, r1);
+    jit_unget_reg(reg);
+}
+
 static void
 _fop1d(jit_state_t *_jit, jit_int32_t op,
        jit_int32_t r0, jit_int32_t r1, jit_float64_t *i0)
@@ -480,6 +492,17 @@ _fop1d(jit_state_t *_jit, jit_int32_t op,
     jit_unget_reg(reg);
 }
 
+static void
+_rfop1d(jit_state_t *_jit, jit_int32_t op,
+	jit_int32_t r0, jit_int32_t r1, jit_float64_t *i0)
+{
+    jit_int32_t		reg;
+    reg = jit_get_reg(jit_class_fpr);
+    movi_d(rn(reg), i0);
+    FPop1(r0, rn(reg), op, r1);
+    jit_unget_reg(reg);
+}
+
 static void
 _extr_f(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
diff --git a/lib/jit_sparc-sz.c b/lib/jit_sparc-sz.c
index df52ca357..6acfd0989 100644
--- a/lib/jit_sparc-sz.c
+++ b/lib/jit_sparc-sz.c
@@ -23,6 +23,7 @@
     12,	/* subci */
     4,	/* subxr */
     4,	/* subxi */
+    16,	/* rsbi */
     4,	/* mulr */
     12,	/* muli */
     8,	/* qmulr */
@@ -178,6 +179,7 @@
     16,	/* addi_f */
     4,	/* subr_f */
     16,	/* subi_f */
+    16,	/* rsbi_f */
     4,	/* mulr_f */
     16,	/* muli_f */
     4,	/* divr_f */
@@ -260,6 +262,7 @@
     24,	/* addi_d */
     4,	/* subr_d */
     24,	/* subi_d */
+    24,	/* rsbi_d */
     4,	/* mulr_d */
     24,	/* muli_d */
     4,	/* divr_d */
diff --git a/lib/jit_sparc.c b/lib/jit_sparc.c
index a8488e2c5..dc1557cc4 100644
--- a/lib/jit_sparc.c
+++ b/lib/jit_sparc.c
@@ -745,6 +745,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(subc,);
 		case_rrr(subx,);
 		case_rrw(subx,);
+		case_rrw(rsb,);
 		case_rrr(mul,);
 		case_rrw(mul,);
 		case_rrrr(qmul,);
@@ -898,6 +899,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _f, 32);
 		case_rrr(sub, _f);
 		case_rrf(sub, _f, 32);
+		case_rrf(rsb, _f, 32);
 		case_rrr(mul, _f);
 		case_rrf(mul, _f, 32);
 		case_rrr(div, _f);
@@ -980,6 +982,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrf(add, _d, 64);
 		case_rrr(sub, _d);
 		case_rrf(sub, _d, 64);
+		case_rrf(rsb, _d, 64);
 		case_rrr(mul, _d);
 		case_rrf(mul, _d, 64);
 		case_rrr(div, _d);
diff --git a/lib/jit_x86-cpu.c b/lib/jit_x86-cpu.c
index 0a5d3302a..2e284f917 100644
--- a/lib/jit_x86-cpu.c
+++ b/lib/jit_x86-cpu.c
@@ -202,6 +202,8 @@ static void _subxr(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define isubxi(r0, i0)		alui(X86_SBB, r0, i0)
 #  define subxi(r0, r1, i0)		_subxi(_jit, r0, r1, i0)
 static void _subxi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
+#  define rsbi(r0, r1, i0)		_rsbi(_jit, r0, r1, i0)
+static void _rsbi(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
 #  define imulr(r0, r1)			_imulr(_jit, r0, r1)
 static void _imulr(jit_state_t*, jit_int32_t, jit_int32_t);
 #  define imuli(r0, r1, i0)		_imuli(_jit, r0, r1, i0)
@@ -1109,6 +1111,13 @@ _subxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
     }
 }
 
+static void
+_rsbi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
+{
+    subi(r0, r1, i0);
+    negr(r0, r0);
+}
+
 static void
 _imulr(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1)
 {
diff --git a/lib/jit_x86-sse.c b/lib/jit_x86-sse.c
index d3bd7aeea..6a3d996e6 100644
--- a/lib/jit_x86-sse.c
+++ b/lib/jit_x86-sse.c
@@ -131,6 +131,12 @@ static void _sse_subi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
 static void _sse_subr_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define sse_subi_d(r0, r1, i0)	_sse_subi_d(_jit, r0, r1, i0)
 static void _sse_subi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
+#  define sse_rsbr_f(r0, r1, r2)	sse_subr_f(r0, r2, r1)
+#  define sse_rsbi_f(r0, r1, i0)	_sse_rsbi_f(_jit, r0, r1, i0)
+static void _sse_rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
+#  define sse_rsbr_d(r0, r1, r2)	sse_subr_d(r0, r2, r1)
+#  define sse_rsbi_d(r0, r1, i0)	_sse_rsbi_d(_jit, r0, r1, i0)
+static void _sse_rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
 #  define sse_mulr_f(r0, r1, r2)	_sse_mulr_f(_jit, r0, r1, r2)
 static void _sse_mulr_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
 #  define sse_muli_f(r0, r1, i0)	_sse_muli_f(_jit, r0, r1, i0)
@@ -592,6 +598,10 @@ _sse_subr_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
 
 dopi(sub)
 
+fopi(rsb)
+
+dopi(rsb)
+
 static void
 _sse_mulr_f(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
 {
diff --git a/lib/jit_x86-sz.c b/lib/jit_x86-sz.c
index 70f8a03ff..1c73ea3e3 100644
--- a/lib/jit_x86-sz.c
+++ b/lib/jit_x86-sz.c
@@ -23,6 +23,7 @@
     6,	/* subci */
     6,	/* subxr */
     5,	/* subxi */
+    10,	/* rsbi */
     5,	/* mulr */
     7,	/* muli */
     20,	/* qmulr */
@@ -178,6 +179,7 @@
     19,	/* addi_f */
     12,	/* subr_f */
     19,	/* subi_f */
+    19,	/* rsbi_f */
     8,	/* mulr_f */
     19,	/* muli_f */
     12,	/* divr_f */
@@ -260,6 +262,7 @@
     26,	/* addi_d */
     12,	/* subr_d */
     26,	/* subi_d */
+    26,	/* rsbi_d */
     8,	/* mulr_d */
     26,	/* muli_d */
     12,	/* divr_d */
@@ -375,6 +378,7 @@
     13,	/* subci */
     9,	/* subxr */
     7,	/* subxi */
+    19,	/* rsbi */
     7,	/* mulr */
     14,	/* muli */
     20,	/* qmulr */
@@ -530,6 +534,7 @@
     21,	/* addi_f */
     15,	/* subr_f */
     21,	/* subi_f */
+    21,	/* rsbi_f */
     10,	/* mulr_f */
     21,	/* muli_f */
     15,	/* divr_f */
@@ -612,6 +617,7 @@
     25,	/* addi_d */
     15,	/* subr_d */
     25,	/* subi_d */
+    25,	/* rsbi_d */
     10,	/* mulr_d */
     25,	/* muli_d */
     15,	/* divr_d */
@@ -725,6 +731,7 @@
     13,	/* subci */
     9,	/* subxr */
     7,	/* subxi */
+    19,	/* rsbi */
     7,	/* mulr */
     14,	/* muli */
     20,	/* qmulr */
@@ -880,6 +887,7 @@
     20,	/* addi_f */
     15,	/* subr_f */
     20,	/* subi_f */
+    20,	/* rsbi_f */
     10,	/* mulr_f */
     20,	/* muli_f */
     15,	/* divr_f */
@@ -962,6 +970,7 @@
     25,	/* addi_d */
     15,	/* subr_d */
     25,	/* subi_d */
+    25,	/* rsbi_d */
     10,	/* mulr_d */
     25,	/* muli_d */
     15,	/* divr_d */
diff --git a/lib/jit_x86-x87.c b/lib/jit_x86-x87.c
index 8faa1401c..2677d4349 100644
--- a/lib/jit_x86-x87.c
+++ b/lib/jit_x86-x87.c
@@ -83,6 +83,12 @@ static void _x87_subi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
 static void _x87_subr_d(jit_state_t*, jit_int32_t, jit_int32_t, jit_int32_t);
 #  define x87_subi_d(r0, r1, i0)	_x87_subi_d(_jit, r0, r1, i0)
 static void _x87_subi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
+#  define x87_rsbr_f(r0, r1, r2)	x87_subr_f(r0, r2, r1)
+#  define x87_rsbi_f(r0, r1, i0)	_x87_rsbi_f(_jit, r0, r1, i0)
+static void _x87_rsbi_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
+#  define x87_rsbr_d(r0, r1, r2)	x87_subr_d(r0, r2, r1)
+#  define x87_rsbi_d(r0, r1, i0)	_x87_rsbi_d(_jit, r0, r1, i0)
+static void _x87_rsbi_d(jit_state_t*,jit_int32_t,jit_int32_t,jit_float64_t*);
 #  define x87_mulr_f(r0, r1, r2)	_x87_mulr_d(_jit, r0, r1, r2)
 #  define x87_muli_f(r0, r1, i0)	_x87_muli_f(_jit, r0, r1, i0)
 static void _x87_muli_f(jit_state_t*,jit_int32_t,jit_int32_t,jit_float32_t*);
@@ -446,6 +452,7 @@ _x87rri(jit_state_t *_jit, jit_int32_t code, jit_int32_t r0, jit_int32_t r1)
 
 fopi(add)
 fopi(sub)
+fopi(rsb)
 fopi(mul)
 fopi(div)
 
@@ -517,6 +524,8 @@ _x87_subr_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
 
 dopi(sub)
 
+dopi(rsb)
+
 static void
 _x87_mulr_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_int32_t r2)
 {
diff --git a/lib/jit_x86.c b/lib/jit_x86.c
index 20dcc211d..6f55ebd1d 100644
--- a/lib/jit_x86.c
+++ b/lib/jit_x86.c
@@ -1297,6 +1297,7 @@ _emit_code(jit_state_t *_jit)
 		case_rrw(subx,);
 		case_rrr(subc,);
 		case_rrw(subc,);
+		case_rrw(rsb,);
 		case_rrr(mul,);
 		case_rrw(mul,);
 		case_rrrr(qmul,);
@@ -1478,6 +1479,7 @@ _emit_code(jit_state_t *_jit)
 		case_ffw(add, _f, 32);
 		case_fff(sub, _f);
 		case_ffw(sub, _f, 32);
+		case_ffw(rsb, _f, 32);
 		case_fff(mul, _f);
 		case_ffw(mul, _f, 32);
 		case_fff(div, _f);
@@ -1576,6 +1578,7 @@ _emit_code(jit_state_t *_jit)
 		case_ffw(add, _d, 64);
 		case_fff(sub, _d);
 		case_ffw(sub, _d, 64);
+		case_ffw(rsb, _d, 64);
 		case_fff(mul, _d);
 		case_ffw(mul, _d, 64);
 		case_fff(div, _d);
diff --git a/lib/lightning.c b/lib/lightning.c
index 9ee261d79..20aa1bd74 100644
--- a/lib/lightning.c
+++ b/lib/lightning.c
@@ -1237,6 +1237,7 @@ _jit_classify(jit_state_t *_jit, jit_code_t code)
 	    break;
 	case jit_code_addi:	case jit_code_addxi:	case jit_code_addci:
 	case jit_code_subi:	case jit_code_subxi:	case jit_code_subci:
+	case jit_code_rsbi:
 	case jit_code_muli:	case jit_code_divi:	case jit_code_divi_u:
 	case jit_code_remi:	case jit_code_remi_u:	case jit_code_andi:
 	case jit_code_ori:	case jit_code_xori:	case jit_code_lshi:
@@ -1254,20 +1255,22 @@ _jit_classify(jit_state_t *_jit, jit_code_t code)
 	    mask = jit_cc_a0_reg|jit_cc_a0_rlh|jit_cc_a0_chg|
 		   jit_cc_a1_reg|jit_cc_a2_int;
 	    break;
-	case jit_code_addi_f:	case jit_code_subi_f:	case jit_code_muli_f:
-	case jit_code_divi_f:	case jit_code_lti_f:	case jit_code_lei_f:
-	case jit_code_eqi_f:	case jit_code_gei_f:	case jit_code_gti_f:
-	case jit_code_nei_f:	case jit_code_unlti_f:	case jit_code_unlei_f:
-	case jit_code_uneqi_f:	case jit_code_ungei_f:	case jit_code_ungti_f:
-	case jit_code_ltgti_f:	case jit_code_ordi_f:	case jit_code_unordi_f:
+	case jit_code_addi_f:	case jit_code_subi_f:	case jit_code_rsbi_f:
+	case jit_code_muli_f:	case jit_code_divi_f:	case jit_code_lti_f:
+	case jit_code_lei_f:	case jit_code_eqi_f:	case jit_code_gei_f:
+	case jit_code_gti_f:	case jit_code_nei_f:	case jit_code_unlti_f:
+	case jit_code_unlei_f:	case jit_code_uneqi_f:	case jit_code_ungei_f:
+	case jit_code_ungti_f:	case jit_code_ltgti_f:	case jit_code_ordi_f:
+	case jit_code_unordi_f:
 	    mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg|jit_cc_a2_flt;
 	    break;
-	case jit_code_addi_d:	case jit_code_subi_d:	case jit_code_muli_d:
-	case jit_code_divi_d:	case jit_code_lti_d:	case jit_code_lei_d:
-	case jit_code_eqi_d:	case jit_code_gei_d:	case jit_code_gti_d:
-	case jit_code_nei_d:	case jit_code_unlti_d:	case jit_code_unlei_d:
-	case jit_code_uneqi_d:	case jit_code_ungei_d:	case jit_code_ungti_d:
-	case jit_code_ltgti_d:	case jit_code_ordi_d:	case jit_code_unordi_d:
+	case jit_code_addi_d:	case jit_code_subi_d:	case jit_code_rsbi_d:
+	case jit_code_muli_d:	case jit_code_divi_d:	case jit_code_lti_d:
+	case jit_code_lei_d:	case jit_code_eqi_d:	case jit_code_gei_d:
+	case jit_code_gti_d:	case jit_code_nei_d:	case jit_code_unlti_d:
+	case jit_code_unlei_d:	case jit_code_uneqi_d:	case jit_code_ungei_d:
+	case jit_code_ungti_d:	case jit_code_ltgti_d:	case jit_code_ordi_d:
+	case jit_code_unordi_d:
 	    mask = jit_cc_a0_reg|jit_cc_a0_chg|jit_cc_a1_reg|jit_cc_a2_dbl;
 	    break;
 	case jit_code_addr:	case jit_code_addxr:	case jit_code_addcr: