runtime: preserve darwin/arm{,64} callee-save registers

CL 14603 attempted to preserve the callee-save registers for
the darwin/arm runtime initialization routine, but I believe it
wasn't sufficient and resulted in the crash reported in issue

Saving and restoring the registers on the stack the same way
linux/arm does seems more obvious and fixes #14778, so do that.

Even though #14778 is not reproducible on darwin/arm64, I applied
a similar change there, and to linux/arm64 which obeys the same
calling convention.

Finally, this CL is a candidate for a 1.6 minor release for the same
reason CL 14603 was in a 1.5 minor release (as CL 16968). It is
small and only touches the iOS platforms and gomobile on darwin/arm
is currently useless without it.

Fixes #14778
Fixes #12590 (again)

Change-Id: I7401daf0bbd7c579a7e84761384a7b763651752a
Reviewed-on: https://go-review.googlesource.com/20621
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Run-TryBot: Elias Naur <elias.naur@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-on: https://go-review.googlesource.com/22049

3 files changed, 61 insertions, 22 deletions

diff --git a/src/runtime/rt0_darwin_arm.s b/src/runtime/rt0_darwin_arm.s
index cbbea80ff8..59733d3ff6 100644
--- a/src/runtime/rt0_darwin_arm.s
+++ b/src/runtime/rt0_darwin_arm.s
@@ -16,12 +16,15 @@ TEXT _rt0_arm_darwin(SB),7,$-4
 //
 // Note that all currently shipping darwin/arm platforms require
 // cgo and do not support c-shared.
-TEXT _rt0_arm_darwin_lib(SB),NOSPLIT,$0
-	// R11 is REGTMP, reserved for liblink. It is used below to
-	// move R0/R1 into globals. However in the darwin ARMv7 calling
-	// convention, it is a callee-saved register. So we save it to a
-	// temporary register.
-	MOVW  R11, R2
+TEXT _rt0_arm_darwin_lib(SB),NOSPLIT,$32
+	// Preserve callee-save registers.
+	MOVW    R4, 12(R13)
+	MOVW    R5, 16(R13)
+	MOVW    R6, 20(R13)
+	MOVW    R7, 24(R13)
+	MOVW    R8, 28(R13)
+	MOVW    R11, 32(R13)
+
 	MOVW  R0, _rt0_arm_darwin_lib_argc<>(SB)
 	MOVW  R1, _rt0_arm_darwin_lib_argv<>(SB)
 
@@ -35,9 +38,8 @@ TEXT _rt0_arm_darwin_lib(SB),NOSPLIT,$0
 	B.EQ  nocgo
 	MOVW  $_rt0_arm_darwin_lib_go(SB), R0
 	MOVW  $0, R1
-	MOVW  R2, R11
 	BL    (R3)
-	RET
+	B rr
 nocgo:
 	MOVW  $0x400000, R0
 	MOVW  R0, (R13) // stacksize
@@ -46,10 +48,18 @@ nocgo:
 	MOVW  $0, R0
 	MOVW  R0, 8(R13) // fnarg
 	MOVW  $runtime·newosproc0(SB), R3
-	MOVW  R2, R11
 	BL    (R3)
+rr:
+	// Restore callee-save registers and return.
+	MOVW    12(R13), R4
+	MOVW    16(R13), R5
+	MOVW    20(R13), R6
+	MOVW    24(R13), R7
+	MOVW    28(R13), R8
+	MOVW    32(R13), R11
 	RET
 
+
 TEXT _rt0_arm_darwin_lib_go(SB),NOSPLIT,$0
 	MOVW  _rt0_arm_darwin_lib_argc<>(SB), R0
 	MOVW  _rt0_arm_darwin_lib_argv<>(SB), R1
diff --git a/src/runtime/rt0_darwin_arm64.s b/src/runtime/rt0_darwin_arm64.s
index fa676c0abe..0a1feb14a4 100644
--- a/src/runtime/rt0_darwin_arm64.s
+++ b/src/runtime/rt0_darwin_arm64.s
@@ -16,12 +16,17 @@ TEXT _rt0_arm64_darwin(SB),NOSPLIT,$-8
 //
 // Note that all currently shipping darwin/arm64 platforms require
 // cgo and do not support c-shared.
-TEXT _rt0_arm64_darwin_lib(SB),NOSPLIT,$0
-	// R27 is REGTMP, reserved for liblink. It is used below to
-	// move R0/R1 into globals. However in the standard ARM64 calling
-	// convention, it is a callee-saved register. So we save it to a
-	// temporary register.
-	MOVD  R27, R7
+TEXT _rt0_arm64_darwin_lib(SB),NOSPLIT,$88
+	// Preserve callee-save registers.
+	MOVD R19, 24(RSP)
+	MOVD R20, 32(RSP)
+	MOVD R21, 40(RSP)
+	MOVD R22, 48(RSP)
+	MOVD R23, 56(RSP)
+	MOVD R24, 64(RSP)
+	MOVD R25, 72(RSP)
+	MOVD R26, 80(RSP)
+	MOVD R27, 88(RSP)
 
 	MOVD  R0, _rt0_arm64_darwin_lib_argc<>(SB)
 	MOVD  R1, _rt0_arm64_darwin_lib_argv<>(SB)
@@ -36,7 +41,16 @@ TEXT _rt0_arm64_darwin_lib(SB),NOSPLIT,$0
 	MOVD  $0, R1
 	BL    (R4)
 
-	MOVD  R7, R27
+	// Restore callee-save registers.
+	MOVD 24(RSP), R19
+	MOVD 32(RSP), R20
+	MOVD 40(RSP), R21
+	MOVD 48(RSP), R22
+	MOVD 56(RSP), R23
+	MOVD 64(RSP), R24
+	MOVD 72(RSP), R25
+	MOVD 80(RSP), R26
+	MOVD 88(RSP), R27
 	RET
 
 TEXT _rt0_arm64_darwin_lib_go(SB),NOSPLIT,$0
diff --git a/src/runtime/rt0_linux_arm64.s b/src/runtime/rt0_linux_arm64.s
index 2b32e35209..31c2367bca 100644
--- a/src/runtime/rt0_linux_arm64.s
+++ b/src/runtime/rt0_linux_arm64.s
@@ -11,11 +11,17 @@ TEXT _rt0_arm64_linux(SB),NOSPLIT,$-8
 
 // When building with -buildmode=c-shared, this symbol is called when the shared
 // library is loaded.
-TEXT _rt0_arm64_linux_lib(SB),NOSPLIT,$40
-	// R27 is REGTMP, reserved for liblink. It is used below to
-	// move R0/R1 into globals. However in the standard ARM64 calling
-	// convention, it is a callee-saved register.
-	MOVD	R27, 24(RSP)
+TEXT _rt0_arm64_linux_lib(SB),NOSPLIT,$88
+	// Preserve callee-save registers.
+	MOVD R19, 24(RSP)
+	MOVD R20, 32(RSP)
+	MOVD R21, 40(RSP)
+	MOVD R22, 48(RSP)
+	MOVD R23, 56(RSP)
+	MOVD R24, 64(RSP)
+	MOVD R25, 72(RSP)
+	MOVD R26, 80(RSP)
+	MOVD R27, 88(RSP)
 
 	MOVD	R0, _rt0_arm64_linux_lib_argc<>(SB)
 	MOVD	R1, _rt0_arm64_linux_lib_argv<>(SB)
@@ -42,7 +48,16 @@ nocgo:
 	BL	(R4)
 
 restore:
-	MOVD	24(RSP), R27
+	// Restore callee-save registers.
+	MOVD 24(RSP), R19
+	MOVD 32(RSP), R20
+	MOVD 40(RSP), R21
+	MOVD 48(RSP), R22
+	MOVD 56(RSP), R23
+	MOVD 64(RSP), R24
+	MOVD 72(RSP), R25
+	MOVD 80(RSP), R26
+	MOVD 88(RSP), R27
 	RET
 
 TEXT _rt0_arm64_linux_lib_go(SB),NOSPLIT,$0