runtime: nacl/arm support.

LGTM=rsc
R=rsc, iant, dave
CC=golang-codereviews
https://codereview.appspot.com/103680046

12 files changed, 470 insertions, 29 deletions

diff --git a/src/pkg/runtime/arch_arm.h b/src/pkg/runtime/arch_arm.h
index b9711289f..2a1077e2f 100644
--- a/src/pkg/runtime/arch_arm.h
+++ b/src/pkg/runtime/arch_arm.h
@@ -6,7 +6,11 @@ enum {
 	thechar = '5',
 	BigEndian = 0,
 	CacheLineSize = 32,
-	RuntimeGogoBytes = 80,
+	RuntimeGogoBytes = 84,
+#ifdef GOOS_nacl
+	PhysPageSize = 65536,
+#else
 	PhysPageSize = 4096,
+#endif
 	PCQuantum = 4
 };
diff --git a/src/pkg/runtime/asm_arm.s b/src/pkg/runtime/asm_arm.s
index 6711d5105..36b2577f3 100644
--- a/src/pkg/runtime/asm_arm.s
+++ b/src/pkg/runtime/asm_arm.s
@@ -36,6 +36,7 @@ TEXT _rt0_go(SB),NOSPLIT,$-4
 	MOVW	R13, g_stackbase(g)
 	BL	runtime·emptyfunc(SB)	// fault if stack check is wrong
 
+#ifndef GOOS_nacl
 	// if there is an _cgo_init, call it.
 	MOVW	_cgo_init(SB), R4
 	CMP	$0, R4
@@ -46,6 +47,7 @@ TEXT _rt0_go(SB),NOSPLIT,$-4
 	MOVW	$setg_gcc<>(SB), R1 	// arg 1: setg
 	MOVW	g, R0 			// arg 0: G
 	BL	(R4) // will clobber R0-R3
+#endif
 
 nocgo:
 	// update stackguard after _cgo_init
@@ -90,7 +92,11 @@ GLOBL	runtime·main·f(SB),RODATA,$4
 TEXT runtime·breakpoint(SB),NOSPLIT,$0-0
 	// gdb won't skip this breakpoint instruction automatically,
 	// so you must manually "set $pc+=4" to skip it and continue.
+#ifdef GOOS_nacl
+	WORD	$0xe125be7f	// BKPT 0x5bef, NACL_INSTR_ARM_BREAKPOINT
+#else
 	WORD	$0xe1200071	// BKPT 0x0001
+#endif
 	RET
 
 TEXT runtime·asminit(SB),NOSPLIT,$0-0
@@ -139,7 +145,8 @@ TEXT runtime·gogo(SB), NOSPLIT, $-4-4
 	MOVW	R11, gobuf_lr(R1)
 	MOVW	R11, gobuf_ctxt(R1)
 	CMP	R11, R11 // set condition codes for == test, needed by stack split
-	MOVW	gobuf_pc(R1), PC
+	MOVW	gobuf_pc(R1), R11
+	B	(R11)
 
 // void mcall(void (*fn)(G*))
 // Switch to m->g0's stack, call fn(g).
diff --git a/src/pkg/runtime/defs_nacl_arm.h b/src/pkg/runtime/defs_nacl_arm.h
new file mode 100644
index 000000000..9ce07ccb2
--- /dev/null
+++ b/src/pkg/runtime/defs_nacl_arm.h
@@ -0,0 +1,70 @@
+// Copyright 2014 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Created by hand, not machine generated.
+
+enum
+{
+	// These values are referred to in the source code
+	// but really don't matter. Even so, use the standard numbers.
+	SIGSEGV = 11,
+	SIGPROF = 27,
+};
+
+typedef struct Siginfo Siginfo;
+
+// native_client/src/trusted/service_runtime/include/machine/_types.h
+typedef struct Timespec Timespec;
+
+struct Timespec
+{
+	int64 tv_sec;
+	int32 tv_nsec;
+};
+
+// native_client/src/trusted/service_runtime/nacl_exception.h
+// native_client/src/include/nacl/nacl_exception.h
+
+typedef struct ExcContext ExcContext;
+typedef struct ExcPortable ExcPortable;
+typedef struct ExcRegsARM ExcRegsARM;
+
+struct ExcRegsARM
+{
+	uint32	r0;
+	uint32	r1;
+	uint32	r2;
+	uint32	r3;
+	uint32	r4;
+	uint32	r5;
+	uint32	r6;
+	uint32	r7;
+	uint32	r8;
+	uint32	r9;	// the value reported here is undefined.
+	uint32	r10;
+	uint32	r11;
+	uint32	r12;
+	uint32	sp;	/* r13 */
+	uint32	lr;	/* r14 */
+	uint32	pc;	/* r15 */
+	uint32	cpsr;
+};
+
+struct ExcContext
+{
+	uint32	size;
+	uint32	portable_context_offset;
+	uint32	portable_context_size;
+	uint32	arch;
+	uint32	regs_size;
+	uint32	reserved[11];
+	ExcRegsARM	regs;
+};
+
+struct ExcPortableContext
+{
+	uint32	pc;
+	uint32	sp;
+	uint32	fp;
+};
diff --git a/src/pkg/runtime/os_nacl_arm.c b/src/pkg/runtime/os_nacl_arm.c
new file mode 100644
index 000000000..c94b2c78a
--- /dev/null
+++ b/src/pkg/runtime/os_nacl_arm.c
@@ -0,0 +1,24 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include "runtime.h"
+#include "defs_GOOS_GOARCH.h"
+#include "os_GOOS.h"
+#include "../../cmd/ld/textflag.h"
+
+void
+runtime·checkgoarm(void)
+{
+	return; // NaCl/ARM only supports ARMv7
+}
+
+#pragma textflag NOSPLIT
+int64
+runtime·cputicks(void)
+{
+	// Currently cputicks() is used in blocking profiler and to seed runtime·fastrand1().
+	// runtime·nanotime() is a poor approximation of CPU ticks that is enough for the profiler.
+	// TODO: need more entropy to better seed fastrand1.
+	return runtime·nanotime();
+}
diff --git a/src/pkg/runtime/rt0_nacl_arm.s b/src/pkg/runtime/rt0_nacl_arm.s
new file mode 100644
index 000000000..df84d5d02
--- /dev/null
+++ b/src/pkg/runtime/rt0_nacl_arm.s
@@ -0,0 +1,20 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include "../../cmd/ld/textflag.h"
+
+// NaCl entry has:
+//	0(FP) - 0
+//	4(FP) - cleanup function pointer, always 0
+//	8(FP) - envc
+//	12(FP) - argc
+//	16(FP) - argv, then 0, then envv, then 0, then auxv
+TEXT _rt0_arm_nacl(SB),NOSPLIT,$-4
+	MOVW	8(R13), R0
+	MOVW	$12(R13), R1
+	MOVM.DB.W [R0-R1], (R13)
+	B	main(SB)
+
+TEXT main(SB),NOSPLIT,$0
+	B	_rt0_go(SB)
diff --git a/src/pkg/runtime/runtime.c b/src/pkg/runtime/runtime.c
index 26dbbbd3e..ec0df1d17 100644
--- a/src/pkg/runtime/runtime.c
+++ b/src/pkg/runtime/runtime.c
@@ -370,11 +370,6 @@ runtime·timediv(int64 v, int32 div, int32 *rem)
 {
 	int32 res, bit;
 
-	if(v >= (int64)div*0x7fffffffLL) {
-		if(rem != nil)
-			*rem = 0;
-		return 0x7fffffff;
-	}
 	res = 0;
 	for(bit = 30; bit >= 0; bit--) {
 		if(v >= ((int64)div<<bit)) {
@@ -382,6 +377,11 @@ runtime·timediv(int64 v, int32 div, int32 *rem)
 			res += 1<<bit;
 		}
 	}
+	if(v >= (int64)div) {
+		if(rem != nil)
+			*rem = 0;
+		return 0x7fffffff;
+	}
 	if(rem != nil)
 		*rem = v;
 	return res;
diff --git a/src/pkg/runtime/signal_arm.c b/src/pkg/runtime/signal_arm.c
index 1e8636846..1f9a2325d 100644
--- a/src/pkg/runtime/signal_arm.c
+++ b/src/pkg/runtime/signal_arm.c
@@ -2,7 +2,7 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// +build darwin dragonfly freebsd linux netbsd openbsd
+// +build darwin dragonfly freebsd linux nacl netbsd openbsd
 
 #include "runtime.h"
 #include "defs_GOOS_GOARCH.h"
diff --git a/src/pkg/runtime/signal_nacl_arm.h b/src/pkg/runtime/signal_nacl_arm.h
new file mode 100644
index 000000000..e5bbb211d
--- /dev/null
+++ b/src/pkg/runtime/signal_nacl_arm.h
@@ -0,0 +1,28 @@
+// Copyright 2014 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#define SIG_REGS(ctxt) (((ExcContext*)(ctxt))->regs)
+
+#define SIG_R0(info, ctxt) (SIG_REGS(ctxt).r0)
+#define SIG_R1(info, ctxt) (SIG_REGS(ctxt).r1)
+#define SIG_R2(info, ctxt) (SIG_REGS(ctxt).r2)
+#define SIG_R3(info, ctxt) (SIG_REGS(ctxt).r3)
+#define SIG_R4(info, ctxt) (SIG_REGS(ctxt).r4)
+#define SIG_R5(info, ctxt) (SIG_REGS(ctxt).r5)
+#define SIG_R6(info, ctxt) (SIG_REGS(ctxt).r6)
+#define SIG_R7(info, ctxt) (SIG_REGS(ctxt).r7)
+#define SIG_R8(info, ctxt) (SIG_REGS(ctxt).r8)
+#define SIG_R9(info, ctxt) (SIG_REGS(ctxt).r9)
+#define SIG_R10(info, ctxt) (SIG_REGS(ctxt).r10)
+#define SIG_FP(info, ctxt) (SIG_REGS(ctxt).r11)
+#define SIG_IP(info, ctxt) (SIG_REGS(ctxt).r12)
+#define SIG_SP(info, ctxt) (SIG_REGS(ctxt).sp)
+#define SIG_LR(info, ctxt) (SIG_REGS(ctxt).lr)
+#define SIG_PC(info, ctxt) (SIG_REGS(ctxt).pc)
+#define SIG_CPSR(info, ctxt) (SIG_REGS(ctxt).cpsr)
+#define SIG_FAULT(info, ctxt) (~0)
+#define SIG_TRAP(info, ctxt) (~0)
+#define SIG_ERROR(info, ctxt) (~0)
+#define SIG_OLDMASK(info, ctxt) (~0)
+#define SIG_CODE0(info, ctxt) (~0)
diff --git a/src/pkg/runtime/sys_nacl_arm.s b/src/pkg/runtime/sys_nacl_arm.s
new file mode 100644
index 000000000..6a2236850
--- /dev/null
+++ b/src/pkg/runtime/sys_nacl_arm.s
@@ -0,0 +1,278 @@
+// Copyright 2014 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include "zasm_GOOS_GOARCH.h"
+#include "../../cmd/ld/textflag.h"
+#include "syscall_nacl.h"
+
+#define NACL_SYSCALL(code) \
+	MOVW	$(0x10000 + ((code)<<5)), R8; BL (R8)
+
+#define NACL_SYSJMP(code) \
+	MOVW	$(0x10000 + ((code)<<5)), R8; B (R8)
+
+TEXT runtime·exit(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_exit)
+
+TEXT runtime·exit1(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_thread_exit)
+
+TEXT runtime·open(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+0(FP), R1
+	MOVW	arg3+0(FP), R2
+	NACL_SYSJMP(SYS_open)
+
+TEXT runtime·close(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_close)
+
+TEXT runtime·read(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	MOVW	arg3+8(FP), R2
+	NACL_SYSJMP(SYS_read)
+
+// func naclWrite(fd int, b []byte) int
+TEXT syscall·naclWrite(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	MOVW	arg3+8(FP), R2
+	NACL_SYSCALL(SYS_write)
+	MOVW	R0, ret+16(FP)
+	RET
+
+TEXT runtime·write(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	MOVW	arg3+8(FP), R2
+	NACL_SYSJMP(SYS_write)
+
+TEXT runtime·nacl_exception_stack(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	NACL_SYSJMP(SYS_exception_stack)
+
+TEXT runtime·nacl_exception_handler(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	NACL_SYSJMP(SYS_exception_handler)
+
+TEXT runtime·nacl_sem_create(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_sem_create)
+
+TEXT runtime·nacl_sem_wait(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_sem_wait)
+
+TEXT runtime·nacl_sem_post(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_sem_post)
+
+TEXT runtime·nacl_mutex_create(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_mutex_create)
+
+TEXT runtime·nacl_mutex_lock(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_mutex_lock)
+
+TEXT runtime·nacl_mutex_trylock(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_mutex_trylock)
+
+TEXT runtime·nacl_mutex_unlock(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_mutex_unlock)
+
+TEXT runtime·nacl_cond_create(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_cond_create)
+
+TEXT runtime·nacl_cond_wait(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	NACL_SYSJMP(SYS_cond_wait)
+
+TEXT runtime·nacl_cond_signal(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_cond_signal)
+
+TEXT runtime·nacl_cond_broadcast(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	NACL_SYSJMP(SYS_cond_broadcast)
+
+TEXT runtime·nacl_cond_timed_wait_abs(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	MOVW	arg3+8(FP), R2
+	NACL_SYSJMP(SYS_cond_timed_wait_abs)
+
+TEXT runtime·nacl_thread_create(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	MOVW	arg3+8(FP), R2
+	MOVW	arg4+12(FP), R3
+	NACL_SYSJMP(SYS_thread_create)
+
+TEXT runtime·mstart_nacl(SB),NOSPLIT,$0
+	MOVW	0(R9), R0 // TLS
+	MOVW	-8(R0), R1 // g
+	MOVW	-4(R0), R2 // m
+	MOVW	R2, g_m(R1)
+	MOVW	R1, g
+	B runtime·mstart(SB)
+
+TEXT runtime·nacl_nanosleep(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	NACL_SYSJMP(SYS_nanosleep)
+
+TEXT runtime·osyield(SB),NOSPLIT,$0
+	NACL_SYSJMP(SYS_sched_yield)
+
+TEXT runtime·mmap(SB),NOSPLIT,$8
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	MOVW	arg3+8(FP), R2
+	MOVW	arg4+12(FP), R3
+	MOVW	arg5+16(FP), R4
+	// arg6:offset should be passed as a pointer (to int64)
+	MOVW	arg6+20(FP), R5
+	MOVW	R5, 4(R13)
+	MOVW	$0, R6
+	MOVW	R6, 8(R13)
+	MOVW	$4(R13), R5
+	MOVM.DB.W [R4,R5], (R13) // arg5 and arg6 are passed on stack
+	NACL_SYSCALL(SYS_mmap)
+	MOVM.IA.W (R13), [R4, R5]
+	CMP	$-4095, R0
+	RSB.HI	$0, R0
+	RET
+
+TEXT time·now(SB),NOSPLIT,$16
+	MOVW	$0, R0 // real time clock
+	MOVW	$4(R13), R1
+	NACL_SYSCALL(SYS_clock_gettime)
+	MOVW	4(R13), R0 // low 32-bit sec
+	MOVW	8(R13), R1 // high 32-bit sec
+	MOVW	12(R13), R2 // nsec
+	MOVW	R0, sec+0(FP)
+	MOVW	R1, sec+4(FP)
+	MOVW	R2, sec+8(FP)
+	RET
+
+TEXT syscall·now(SB),NOSPLIT,$0
+	B time·now(SB)
+
+TEXT runtime·nacl_clock_gettime(SB),NOSPLIT,$0
+	MOVW	arg1+0(FP), R0
+	MOVW	arg2+4(FP), R1
+	NACL_SYSJMP(SYS_clock_gettime)
+
+// int64 nanotime(void) so really
+// void nanotime(int64 *nsec)
+TEXT runtime·nanotime(SB),NOSPLIT,$16
+	MOVW	$0, R0 // real time clock
+	MOVW	$4(R13), R1
+	NACL_SYSCALL(SYS_clock_gettime)
+	MOVW	4(R13), R0 // low 32-bit sec
+	MOVW	8(R13), R1 // high 32-bit sec (ignored for now)
+	MOVW	12(R13), R2 // nsec
+	MOVW	$1000000000, R3
+	MULLU	R0, R3, (R1, R0)
+	MOVW	$0, R4
+	ADD.S	R2, R0
+	ADC	R4, R1
+	MOVW	0(FP), R2
+	MOVW	R0, 0(R2)
+	MOVW	R1, 4(R2)
+	RET
+
+TEXT runtime·sigtramp(SB),NOSPLIT,$80
+	// load g from thread context
+	MOVW	$ctxt+-4(FP), R0
+	MOVW	(16*4+10*4)(R0), g
+
+	// check that g exists
+	CMP	$0, g
+	BNE 	4(PC)
+	MOVW  	$runtime·badsignal2(SB), R11
+	BL	(R11)
+	RET
+
+	// save g
+	MOVW	g, R3
+	MOVW	g, 20(R13)
+
+	// g = m->gsignal
+	MOVW	g_m(g), R8
+	MOVW	m_gsignal(R8), g
+
+	// copy arguments for call to sighandler
+	MOVW	$11, R0
+	MOVW	R0, 4(R13) // signal
+	MOVW	$0, R0
+	MOVW	R0, 8(R13) // siginfo
+	MOVW	$ctxt+-4(FP), R0
+	MOVW	R0, 12(R13) // context
+	MOVW	R3, 16(R13) // g
+
+	BL	runtime·sighandler(SB)
+
+	// restore g
+	MOVW	20(R13), g
+
+sigtramp_ret:
+	// Enable exceptions again.
+	NACL_SYSCALL(SYS_exception_clear_flag)
+
+	// Restore registers as best we can. Impossible to do perfectly.
+	// See comment in sys_nacl_386.s for extended rationale.
+	MOVW	$ctxt+-4(FP), R1
+	ADD	$64, R1
+	MOVW	(0*4)(R1), R0
+	MOVW	(2*4)(R1), R2
+	MOVW	(3*4)(R1), R3
+	MOVW	(4*4)(R1), R4
+	MOVW	(5*4)(R1), R5
+	MOVW	(6*4)(R1), R6
+	MOVW	(7*4)(R1), R7
+	MOVW	(8*4)(R1), R8
+	// cannot write to R9
+	MOVW	(10*4)(R1), g
+	MOVW	(11*4)(R1), R11
+	MOVW	(12*4)(R1), R12
+	MOVW	(13*4)(R1), R13
+	MOVW	(14*4)(R1), R14
+	MOVW	(15*4)(R1), R1
+	B	(R1)
+
+nog:
+	MOVW	$0, R0
+	RET
+
+TEXT runtime·nacl_sysinfo(SB),NOSPLIT,$16
+	RET
+
+TEXT runtime·casp(SB),NOSPLIT,$0
+	B	runtime·cas(SB)
+
+// This is only valid for ARMv6+, however, NaCl/ARM is only defined
+// for ARMv7A anyway.
+// bool armcas(int32 *val, int32 old, int32 new)
+// AtomiBLy:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	}else
+//		return 0;
+TEXT runtime·cas(SB),NOSPLIT,$0
+	B runtime·armcas(SB)
+
+TEXT runtime·read_tls_fallback(SB),NOSPLIT,$-4
+	WORD $0xe7fedef0 // NACL_INSTR_ARM_ABORT_NOW (UDF #0xEDE0)
diff --git a/src/pkg/runtime/tls_arm.s b/src/pkg/runtime/tls_arm.s
index 040ce7d85..37edfa968 100644
--- a/src/pkg/runtime/tls_arm.s
+++ b/src/pkg/runtime/tls_arm.s
@@ -23,6 +23,10 @@
 // NOTE: runtime.gogo assumes that R1 is preserved by this function.
 //       runtime.mcall assumes this function only clobbers R0 and R11.
 TEXT runtime·save_g(SB),NOSPLIT,$0
+#ifdef GOOS_nacl
+	// nothing to do as nacl/arm does not use TLS at all.
+	RET
+#endif
 	MRC		15, 0, R0, C13, C0, 3 // fetch TLS base pointer
 	// $runtime.tlsg(SB) is a special linker symbol.
 	// It is the offset from the TLS base pointer to our
@@ -40,6 +44,10 @@ TEXT runtime·save_g(SB),NOSPLIT,$0
 // thread-local memory, for use after calling externally compiled
 // ARM code that overwrote those registers.
 TEXT runtime·load_g(SB),NOSPLIT,$0
+#ifdef GOOS_nacl
+	// nothing to do as nacl/arm does not use TLS at all.
+	RET
+#endif
 	MRC		15, 0, R0, C13, C0, 3 // fetch TLS base pointer
 	// $runtime.tlsg(SB) is a special linker symbol.
 	// It is the offset from the TLS base pointer to our
diff --git a/src/pkg/runtime/vlop_arm.s b/src/pkg/runtime/vlop_arm.s
index c33640695..02bab3107 100644
--- a/src/pkg/runtime/vlop_arm.s
+++ b/src/pkg/runtime/vlop_arm.s
@@ -108,7 +108,8 @@ TEXT udiv<>(SB),NOSPLIT,$-4
 	CLZ 	R(q), R(s) // find normalizing shift
 	MOVW.S	R(q)<<R(s), R(a)
 	MOVW	$fast_udiv_tab<>-64(SB), R(M)
-	MOVBU.NE	R(a)>>25(R(M)), R(a) // index by most significant 7 bits of divisor
+	ADD.NE	R(a)>>25, R(M), R(a) // index by most significant 7 bits of divisor
+	MOVBU.NE	(R(a)), R(a)
 
 	SUB.S	$7, R(s)
 	RSB 	$0, R(q), R(M) // M = -q
@@ -181,26 +182,26 @@ udiv_by_0:
 	MOVW	R1, 0(R13) // expected here for traceback
 	B 	runtime·panicdivide(SB)
 
-TEXT fast_udiv_tab<>(SB),NOSPLIT,$-4
-	// var tab [64]byte
-	// tab[0] = 255; for i := 1; i <= 63; i++ { tab[i] = (1<<14)/(64+i) }
-	// laid out here as little-endian uint32s
-	WORD $0xf4f8fcff
-	WORD $0xe6eaedf0
-	WORD $0xdadde0e3
-	WORD $0xcfd2d4d7
-	WORD $0xc5c7cacc
-	WORD $0xbcbec0c3
-	WORD $0xb4b6b8ba
-	WORD $0xacaeb0b2
-	WORD $0xa5a7a8aa
-	WORD $0x9fa0a2a3
-	WORD $0x999a9c9d
-	WORD $0x93949697
-	WORD $0x8e8f9092
-	WORD $0x898a8c8d
-	WORD $0x85868788
-	WORD $0x81828384
+// var tab [64]byte
+// tab[0] = 255; for i := 1; i <= 63; i++ { tab[i] = (1<<14)/(64+i) }
+// laid out here as little-endian uint32s
+DATA fast_udiv_tab<>+0x00(SB)/4, $0xf4f8fcff
+DATA fast_udiv_tab<>+0x04(SB)/4, $0xe6eaedf0
+DATA fast_udiv_tab<>+0x08(SB)/4, $0xdadde0e3
+DATA fast_udiv_tab<>+0x0c(SB)/4, $0xcfd2d4d7
+DATA fast_udiv_tab<>+0x10(SB)/4, $0xc5c7cacc
+DATA fast_udiv_tab<>+0x14(SB)/4, $0xbcbec0c3
+DATA fast_udiv_tab<>+0x18(SB)/4, $0xb4b6b8ba
+DATA fast_udiv_tab<>+0x1c(SB)/4, $0xacaeb0b2
+DATA fast_udiv_tab<>+0x20(SB)/4, $0xa5a7a8aa
+DATA fast_udiv_tab<>+0x24(SB)/4, $0x9fa0a2a3
+DATA fast_udiv_tab<>+0x28(SB)/4, $0x999a9c9d
+DATA fast_udiv_tab<>+0x2c(SB)/4, $0x93949697
+DATA fast_udiv_tab<>+0x30(SB)/4, $0x8e8f9092
+DATA fast_udiv_tab<>+0x34(SB)/4, $0x898a8c8d
+DATA fast_udiv_tab<>+0x38(SB)/4, $0x85868788
+DATA fast_udiv_tab<>+0x3c(SB)/4, $0x81828384
+GLOBL fast_udiv_tab<>(SB), RODATA, $64
 
 // The linker will pass numerator in R(TMP), and it also
 // expects the result in R(TMP)
diff --git a/src/pkg/runtime/vlrt_arm.c b/src/pkg/runtime/vlrt_arm.c
index 7dd71b40e..016fd7a35 100644
--- a/src/pkg/runtime/vlrt_arm.c
+++ b/src/pkg/runtime/vlrt_arm.c
@@ -506,6 +506,7 @@ _mmv(Vlong *l, Vlong *r)
 	l->lo = r->lo;
 }
 
+#pragma textflag NOSPLIT
 void
 _vasop(Vlong *ret, void *lv, void fn(Vlong*, Vlong, Vlong), int type, Vlong rv)
 {