cmd/compile: mark R16, R17 clobbered for non-standard calls on ARM64

On ARM64, (external) linker generated trampoline may clobber R16
and R17. In CL 183842 we change Duff's devices not to use those
registers. However, this is not enough. The register allocator
also needs to know that these registers may be clobbered in any
calls that don't follow the standard Go calling convention. This
include Duff's devices and the write barrier.

Fixes #32773, second attempt.

Change-Id: Ia52a891d9bbb8515c927617dd53aee5af5bd9aa4
Reviewed-on: https://go-review.googlesource.com/c/go/+/184437
Run-TryBot: Cherry Zhang <[email protected]>
TryBot-Result: Go Bot <[email protected]>
Reviewed-by: Meng Zhuo <[email protected]>
Reviewed-by: Keith Randall <[email protected]>
Trust: Meng Zhuo <[email protected]>

Author: cherrymui

src/cmd/compile/internal/ssa/gen/ARM64Ops.go

@@ -512,13 +512,14 @@ func init() {
 		// auxint = offset into duffzero code to start executing
 		// returns mem
 		// R20 changed as side effect
+		// R16 and R17 may be clobbered by linker trampoline.
 		{
 			name:      "DUFFZERO",
 			aux:       "Int64",
 			argLength: 2,
 			reg: regInfo{
 				inputs:   []regMask{buildReg("R20")},
-				clobbers: buildReg("R20 R30"),
+				clobbers: buildReg("R16 R17 R20 R30"),
 			},
 			faultOnNilArg0: true,
 			unsafePoint:    true, // FP maintenance around DUFFZERO can be clobbered by interrupts
@@ -552,13 +553,14 @@ func init() {
 		// auxint = offset into duffcopy code to start executing
 		// returns mem
 		// R20, R21 changed as side effect
+		// R16 and R17 may be clobbered by linker trampoline.
 		{
 			name:      "DUFFCOPY",
 			aux:       "Int64",
 			argLength: 3,
 			reg: regInfo{
 				inputs:   []regMask{buildReg("R21"), buildReg("R20")},
-				clobbers: buildReg("R20 R21 R26 R30"),
+				clobbers: buildReg("R16 R17 R20 R21 R26 R30"),
 			},
 			faultOnNilArg0: true,
 			faultOnNilArg1: true,
@@ -717,7 +719,8 @@ func init() {
 		// LoweredWB invokes runtime.gcWriteBarrier. arg0=destptr, arg1=srcptr, arg2=mem, aux=runtime.gcWriteBarrier
 		// It saves all GP registers if necessary,
 		// but clobbers R30 (LR) because it's a call.
-		{name: "LoweredWB", argLength: 3, reg: regInfo{inputs: []regMask{buildReg("R2"), buildReg("R3")}, clobbers: (callerSave &^ gpg) | buildReg("R30")}, clobberFlags: true, aux: "Sym", symEffect: "None"},
+		// R16 and R17 may be clobbered by linker trampoline.
+		{name: "LoweredWB", argLength: 3, reg: regInfo{inputs: []regMask{buildReg("R2"), buildReg("R3")}, clobbers: (callerSave &^ gpg) | buildReg("R16 R17 R30")}, clobberFlags: true, aux: "Sym", symEffect: "None"},
 
 		// There are three of these functions so that they can have three different register inputs.
 		// When we check 0 <= c <= cap (A), then 0 <= b <= c (B), then 0 <= a <= b (C), we want the

src/cmd/compile/internal/ssa/opGen.go

@@ -1158,10 +1158,10 @@ TEXT ·checkASM(SB),NOSPLIT,$0-1
 // It does not clobber any general-purpose registers,
 // but may clobber others (e.g., floating point registers)
 // The act of CALLing gcWriteBarrier will clobber R30 (LR).
-TEXT runtime·gcWriteBarrier(SB),NOSPLIT,$216
+TEXT runtime·gcWriteBarrier(SB),NOSPLIT,$200
 	// Save the registers clobbered by the fast path.
-	MOVD	R0, 200(RSP)
-	MOVD	R1, 208(RSP)
+	MOVD	R0, 184(RSP)
+	MOVD	R1, 192(RSP)
 	MOVD	g_m(g), R0
 	MOVD	m_p(R0), R0
 	MOVD	(p_wbBuf+wbBuf_next)(R0), R1
@@ -1177,8 +1177,8 @@ TEXT runtime·gcWriteBarrier(SB),NOSPLIT,$216
 	// Is the buffer full? (flags set in CMP above)
 	BEQ	flush
 ret:
-	MOVD	200(RSP), R0
-	MOVD	208(RSP), R1
+	MOVD	184(RSP), R0
+	MOVD	192(RSP), R1
 	// Do the write.
 	MOVD	R3, (R2)
 	RET
@@ -1202,17 +1202,16 @@ flush:
 	MOVD	R13, 96(RSP)
 	MOVD	R14, 104(RSP)
 	MOVD	R15, 112(RSP)
-	MOVD	R16, 120(RSP)
-	MOVD	R17, 128(RSP)
+	// R16, R17 may be clobbered by linker trampoline
 	// R18 is unused.
-	MOVD	R19, 136(RSP)
-	MOVD	R20, 144(RSP)
-	MOVD	R21, 152(RSP)
-	MOVD	R22, 160(RSP)
-	MOVD	R23, 168(RSP)
-	MOVD	R24, 176(RSP)
-	MOVD	R25, 184(RSP)
-	MOVD	R26, 192(RSP)
+	MOVD	R19, 120(RSP)
+	MOVD	R20, 128(RSP)
+	MOVD	R21, 136(RSP)
+	MOVD	R22, 144(RSP)
+	MOVD	R23, 152(RSP)
+	MOVD	R24, 160(RSP)
+	MOVD	R25, 168(RSP)
+	MOVD	R26, 176(RSP)
 	// R27 is temp register.
 	// R28 is g.
 	// R29 is frame pointer (unused).
@@ -1236,16 +1235,14 @@ flush:
 	MOVD	96(RSP), R13
 	MOVD	104(RSP), R14
 	MOVD	112(RSP), R15
-	MOVD	120(RSP), R16
-	MOVD	128(RSP), R17
-	MOVD	136(RSP), R19
-	MOVD	144(RSP), R20
-	MOVD	152(RSP), R21
-	MOVD	160(RSP), R22
-	MOVD	168(RSP), R23
-	MOVD	176(RSP), R24
-	MOVD	184(RSP), R25
-	MOVD	192(RSP), R26
+	MOVD	120(RSP), R19
+	MOVD	128(RSP), R20
+	MOVD	136(RSP), R21
+	MOVD	144(RSP), R22
+	MOVD	152(RSP), R23
+	MOVD	160(RSP), R24
+	MOVD	168(RSP), R25
+	MOVD	176(RSP), R26
 	JMP	ret
 
 // Note: these functions use a special calling convention to save generated code space.