internal/fakecgo: use C ABI in amd64 trampolines (#394)

Go might expect (in the future) that x_cgo_init_trampoline trampolines and
friends honor the C calling convention, i.e. callee-saved (aka non-volatile)
registers shouldn't be clobbered by the called function.

It is more future proof, safer, and more standard, to use R11 as a temporary
register.

commit: 4f3cc2789f2e2d70a6fe3cda1d183244c90804bd
author: Quim Muntal · 9mo ago · 2026-01-23 04:02 UTC
committer: GitHub · 9mo ago · 2026-01-23 04:02 UTC
parent: a280c0d
signature: GPG UnverifiedB5690EEEBB952194 The key is unknown

1 file · +22 -29

M internal/fakecgo/trampolines_amd64.s +22 -29


@@ -7,12 +7,6 @@
7	7		trampoline for emulating required C functions for cgo in go (see cgo.go)
8	8		(we convert cdecl calling convention to go and vice-versa)
9	9
10		-	Since we're called from go and call into C we can cheat a bit with the calling conventions:
11		-	- in go all the registers are caller saved
12		-	- in C we have a couple of callee saved registers
13		-
14		-	=> we can use BX, R12, R13, R14, R15 instead of the stack
15		-
16	10		C Calling convention cdecl used here (we only need integer args):
17	11		1. arg: DI
18	12		2. arg: SI

@@ -22,6 +16,7 @@ C Calling convention cdecl used here (we only need integer args):
22	16		6. arg: R9
23	17		We don't need floats with these functions -> AX=0
24	18		return value will be in AX
	19	+	temporary register is R11
25	20		*/
26	21		#include "textflag.h"
27	22		#include "go_asm.h"

@@ -32,46 +27,44 @@ return value will be in AX
32	27		TEXT x_cgo_init_trampoline(SB), NOSPLIT, $16
33	28		MOVQ DI, AX
34	29		MOVQ SI, BX
35		-	MOVQ ·x_cgo_init_call(SB), DX
36		-	MOVQ (DX), CX
37		-	CALL CX
	30	+	MOVQ ·x_cgo_init_call(SB), R11
	31	+	MOVQ (R11), R11
	32	+	CALL R11
38	33		RET
39	34
40	35		TEXT x_cgo_thread_start_trampoline(SB), NOSPLIT, $8
41	36		MOVQ DI, AX
42		-	MOVQ ·x_cgo_thread_start_call(SB), DX
43		-	MOVQ (DX), CX
44		-	CALL CX
	37	+	MOVQ ·x_cgo_thread_start_call(SB), R11
	38	+	MOVQ (R11), R11
	39	+	CALL R11
45	40		RET
46	41
47	42		TEXT x_cgo_setenv_trampoline(SB), NOSPLIT, $8
48	43		MOVQ DI, AX
49		-	MOVQ ·x_cgo_setenv_call(SB), DX
50		-	MOVQ (DX), CX
51		-	CALL CX
	44	+	MOVQ ·x_cgo_setenv_call(SB), R11
	45	+	MOVQ (R11), R11
	46	+	CALL R11
52	47		RET
53	48
54	49		TEXT x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8
55	50		MOVQ DI, AX
56		-	MOVQ ·x_cgo_unsetenv_call(SB), DX
57		-	MOVQ (DX), CX
58		-	CALL CX
	51	+	MOVQ ·x_cgo_unsetenv_call(SB), R11
	52	+	MOVQ (R11), R11
	53	+	CALL R11
59	54		RET
60	55
61	56		TEXT x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0
62		-	CALL ·x_cgo_notify_runtime_init_done(SB)
63		-	RET
	57	+	JMP ·x_cgo_notify_runtime_init_done(SB)
64	58
65	59		TEXT x_cgo_bindm_trampoline(SB), NOSPLIT, $0
66		-	CALL ·x_cgo_bindm(SB)
67		-	RET
	60	+	JMP ·x_cgo_bindm(SB)
68	61
69	62		// func setg_trampoline(setg uintptr, g uintptr)
70	63		TEXT ·setg_trampoline(SB), NOSPLIT, $0-16
71	64		MOVQ G+8(FP), DI
72		-	MOVQ setg+0(FP), BX
	65	+	MOVQ setg+0(FP), R11
73	66		XORL AX, AX
74		-	CALL BX
	67	+	CALL R11
75	68		RET
76	69
77	70		TEXT threadentry_trampoline(SB), NOSPLIT, $0

@@ -83,15 +76,15 @@ TEXT threadentry_trampoline(SB), NOSPLIT, $0
83	76		PXOR X15, X15
84	77
85	78		MOVQ DI, AX
86		-	MOVQ ·threadentry_call(SB), DX
87		-	MOVQ (DX), CX
88		-	CALL CX
	79	+	MOVQ ·threadentry_call(SB), R11
	80	+	MOVQ (R11), R11
	81	+	CALL R11
89	82
90	83		POP_REGS_HOST_TO_ABI0()
91	84		RET
92	85
93	86		TEXT ·call5(SB), NOSPLIT, $0-56
94		-	MOVQ fn+0(FP), BX
	87	+	MOVQ fn+0(FP), R11
95	88		MOVQ a1+8(FP), DI
96	89		MOVQ a2+16(FP), SI
97	90		MOVQ a3+24(FP), DX

@@ -105,7 +98,7 @@ TEXT ·call5(SB), NOSPLIT, $0-56
105	98		SUBQ $16, SP // allocate space for alignment
106	99		ANDQ $-16, SP // align on 16 bytes for SSE
107	100
108		-	CALL BX
	101	+	CALL R11
109	102
110	103		MOVQ BP, SP // get SP back
111	104		POPQ BP // restore BP

 trampoline for emulating required C functions for cgo in go (see cgo.go)
 (we convert cdecl calling convention to go and vice-versa)
-Since we're called from go and call into C we can cheat a bit with the calling conventions:
- - in go all the registers are caller saved
- - in C we have a couple of callee saved registers
-=> we can use BX, R12, R13, R14, R15 instead of the stack
 C Calling convention cdecl used here (we only need integer args):
 1. arg: DI
 2. arg: SI
 6. arg: R9
 We don't need floats with these functions -> AX=0
 return value will be in AX
+temporary register is R11
 */
 #include "textflag.h"
 #include "go_asm.h"
 TEXT x_cgo_init_trampoline(SB), NOSPLIT, $16
 	MOVQ DI, AX
 	MOVQ SI, BX
-	MOVQ ·x_cgo_init_call(SB), DX
+	MOVQ ·x_cgo_init_call(SB), R11
-	MOVQ (DX), CX
+	MOVQ (R11), R11
-	CALL CX
+	CALL R11
 	RET
 TEXT x_cgo_thread_start_trampoline(SB), NOSPLIT, $8
 	MOVQ DI, AX
-	MOVQ ·x_cgo_thread_start_call(SB), DX
+	MOVQ ·x_cgo_thread_start_call(SB), R11
-	MOVQ (DX), CX
+	MOVQ (R11), R11
-	CALL CX
+	CALL R11
 	RET
 TEXT x_cgo_setenv_trampoline(SB), NOSPLIT, $8
 	MOVQ DI, AX
-	MOVQ ·x_cgo_setenv_call(SB), DX
+	MOVQ ·x_cgo_setenv_call(SB), R11
-	MOVQ (DX), CX
+	MOVQ (R11), R11
-	CALL CX
+	CALL R11
 	RET
 TEXT x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8
 	MOVQ DI, AX
-	MOVQ ·x_cgo_unsetenv_call(SB), DX
+	MOVQ ·x_cgo_unsetenv_call(SB), R11
-	MOVQ (DX), CX
+	MOVQ (R11), R11
-	CALL CX
+	CALL R11
 	RET
 TEXT x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0
-	CALL ·x_cgo_notify_runtime_init_done(SB)
+	JMP ·x_cgo_notify_runtime_init_done(SB)
-	RET
 TEXT x_cgo_bindm_trampoline(SB), NOSPLIT, $0
-	CALL ·x_cgo_bindm(SB)
+	JMP ·x_cgo_bindm(SB)
-	RET
 // func setg_trampoline(setg uintptr, g uintptr)
 TEXT ·setg_trampoline(SB), NOSPLIT, $0-16
 	MOVQ G+8(FP), DI
-	MOVQ setg+0(FP), BX
+	MOVQ setg+0(FP), R11
 	XORL AX, AX
-	CALL BX
+	CALL R11
 	RET
 TEXT threadentry_trampoline(SB), NOSPLIT, $0
 	PXOR X15, X15
 	MOVQ DI, AX
-	MOVQ ·threadentry_call(SB), DX
+	MOVQ ·threadentry_call(SB), R11
-	MOVQ (DX), CX
+	MOVQ (R11), R11
-	CALL CX
+	CALL R11
 	POP_REGS_HOST_TO_ABI0()
 	RET
 TEXT ·call5(SB), NOSPLIT, $0-56
-	MOVQ fn+0(FP), BX
+	MOVQ fn+0(FP), R11
 	MOVQ a1+8(FP), DI
 	MOVQ a2+16(FP), SI
 	MOVQ a3+24(FP), DX
 	SUBQ  $16, SP  // allocate space for alignment
 	ANDQ  $-16, SP // align on 16 bytes for SSE
-	CALL BX
+	CALL R11
 	MOVQ BP, SP // get SP back
 	POPQ BP     // restore BP