runtime, runtime/cgo: save callee-saved FP registers on arm64

For #14876.

Change-Id: I0992859264cbaf9c9b691fad53345bbb01b4cf3b
Reviewed-on: https://go-review.googlesource.com/21085Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

src/runtime/cgo/asm_arm64.s

@@ -16,7 +16,7 @@ TEXT crosscall2(SB),NOSPLIT,$-8
 	 *  first arg.
 	 * TODO(minux): use LDP/STP here if it matters.
 	 */
-	SUB	$128, RSP
+	SUB	$(8*24), RSP
 	MOVD	R1, (8*1)(RSP)
 	MOVD	R2, (8*2)(RSP)
 	MOVD	R19, (8*3)(RSP)
@@ -31,6 +31,14 @@ TEXT crosscall2(SB),NOSPLIT,$-8
 	MOVD	g, (8*12)(RSP)
 	MOVD	R29, (8*13)(RSP)
 	MOVD	R30, (8*14)(RSP)
+	FMOVD	F8, (8*15)(RSP)
+	FMOVD	F9, (8*16)(RSP)
+	FMOVD	F10, (8*17)(RSP)
+	FMOVD	F11, (8*18)(RSP)
+	FMOVD	F12, (8*19)(RSP)
+	FMOVD	F13, (8*20)(RSP)
+	FMOVD	F14, (8*21)(RSP)
+	FMOVD	F15, (8*22)(RSP)
 
 	MOVD	R0, R19
 
@@ -53,5 +61,13 @@ TEXT crosscall2(SB),NOSPLIT,$-8
 	MOVD	(8*12)(RSP), g
 	MOVD	(8*13)(RSP), R29
 	MOVD	(8*14)(RSP), R30
-	ADD	$128, RSP
+	FMOVD	(8*15)(RSP), F8
+	FMOVD	(8*16)(RSP), F9
+	FMOVD	(8*17)(RSP), F10
+	FMOVD	(8*18)(RSP), F11
+	FMOVD	(8*19)(RSP), F12
+	FMOVD	(8*20)(RSP), F13
+	FMOVD	(8*21)(RSP), F14
+	FMOVD	(8*22)(RSP), F15
+	ADD	$(8*24), RSP
 	RET

src/runtime/rt0_darwin_arm64.s

@@ -16,7 +16,7 @@ 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,$88
+TEXT _rt0_arm64_darwin_lib(SB),NOSPLIT,$168
 	// Preserve callee-save registers.
 	MOVD R19, 24(RSP)
 	MOVD R20, 32(RSP)
@@ -27,6 +27,14 @@ TEXT _rt0_arm64_darwin_lib(SB),NOSPLIT,$88
 	MOVD R25, 72(RSP)
 	MOVD R26, 80(RSP)
 	MOVD R27, 88(RSP)
+	FMOVD F8, 96(RSP)
+	FMOVD F9, 104(RSP)
+	FMOVD F10, 112(RSP)
+	FMOVD F11, 120(RSP)
+	FMOVD F12, 128(RSP)
+	FMOVD F13, 136(RSP)
+	FMOVD F14, 144(RSP)
+	FMOVD F15, 152(RSP)
 
 	MOVD  R0, _rt0_arm64_darwin_lib_argc<>(SB)
 	MOVD  R1, _rt0_arm64_darwin_lib_argv<>(SB)
@@ -51,6 +59,14 @@ TEXT _rt0_arm64_darwin_lib(SB),NOSPLIT,$88
 	MOVD 72(RSP), R25
 	MOVD 80(RSP), R26
 	MOVD 88(RSP), R27
+	FMOVD 96(RSP), F8
+	FMOVD 104(RSP), F9
+	FMOVD 112(RSP), F10
+	FMOVD 120(RSP), F11
+	FMOVD 128(RSP), F12
+	FMOVD 136(RSP), F13
+	FMOVD 144(RSP), F14
+	FMOVD 152(RSP), F15
 	RET
 
 TEXT _rt0_arm64_darwin_lib_go(SB),NOSPLIT,$0

src/runtime/rt0_linux_arm64.s

@@ -11,7 +11,7 @@ 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,$88
+TEXT _rt0_arm64_linux_lib(SB),NOSPLIT,$168
 	// Preserve callee-save registers.
 	MOVD R19, 24(RSP)
 	MOVD R20, 32(RSP)
@@ -22,6 +22,14 @@ TEXT _rt0_arm64_linux_lib(SB),NOSPLIT,$88
 	MOVD R25, 72(RSP)
 	MOVD R26, 80(RSP)
 	MOVD R27, 88(RSP)
+	FMOVD F8, 96(RSP)
+	FMOVD F9, 104(RSP)
+	FMOVD F10, 112(RSP)
+	FMOVD F11, 120(RSP)
+	FMOVD F12, 128(RSP)
+	FMOVD F13, 136(RSP)
+	FMOVD F14, 144(RSP)
+	FMOVD F15, 152(RSP)
 
 	MOVD	R0, _rt0_arm64_linux_lib_argc<>(SB)
 	MOVD	R1, _rt0_arm64_linux_lib_argv<>(SB)
@@ -58,6 +66,14 @@ restore:
 	MOVD 72(RSP), R25
 	MOVD 80(RSP), R26
 	MOVD 88(RSP), R27
+	FMOVD 96(RSP), F8
+	FMOVD 104(RSP), F9
+	FMOVD 112(RSP), F10
+	FMOVD 120(RSP), F11
+	FMOVD 128(RSP), F12
+	FMOVD 136(RSP), F13
+	FMOVD 144(RSP), F14
+	FMOVD 152(RSP), F15
 	RET
 
 TEXT _rt0_arm64_linux_lib_go(SB),NOSPLIT,$0