diff --git a/internal/fakecgo/callbacks.go b/internal/fakecgo/callbacks.go
index 27d4c98c..069a55e2 100644
--- a/internal/fakecgo/callbacks.go
+++ b/internal/fakecgo/callbacks.go
@@ -12,35 +12,6 @@ import (
 
 // TODO: decide if we need _runtime_cgo_panic_internal
 
-//go:linkname x_cgo_init_trampoline x_cgo_init_trampoline
-//go:linkname _cgo_init _cgo_init
-var x_cgo_init_trampoline byte
-var _cgo_init = &x_cgo_init_trampoline
-
-// Creates a new system thread without updating any Go state.
-//
-// This method is invoked during shared library loading to create a new OS
-// thread to perform the runtime initialization. This method is similar to
-// _cgo_sys_thread_start except that it doesn't update any Go state.
-
-//go:linkname x_cgo_thread_start_trampoline x_cgo_thread_start_trampoline
-//go:linkname _cgo_thread_start _cgo_thread_start
-var x_cgo_thread_start_trampoline byte
-var _cgo_thread_start = &x_cgo_thread_start_trampoline
-
-// Notifies that the runtime has been initialized.
-//
-// We currently block at every CGO entry point (via _cgo_wait_runtime_init_done)
-// to ensure that the runtime has been initialized before the CGO call is
-// executed. This is necessary for shared libraries where we kickoff runtime
-// initialization in a separate thread and return without waiting for this
-// thread to complete the init.
-
-//go:linkname x_cgo_notify_runtime_init_done_trampoline x_cgo_notify_runtime_init_done_trampoline
-//go:linkname _cgo_notify_runtime_init_done _cgo_notify_runtime_init_done
-var x_cgo_notify_runtime_init_done_trampoline byte
-var _cgo_notify_runtime_init_done = &x_cgo_notify_runtime_init_done_trampoline
-
 // Indicates whether a dummy thread key has been created or not.
 //
 // When calling go exported function from C, we register a destructor
@@ -60,34 +31,5 @@ func set_crosscall2() {
 //go:linkname _set_crosscall2 runtime.set_crosscall2
 var _set_crosscall2 = set_crosscall2
 
-// Store the g into the thread-specific value.
-// So that pthread_key_destructor will dropm when the thread is exiting.
-
-//go:linkname x_cgo_bindm_trampoline x_cgo_bindm_trampoline
-//go:linkname _cgo_bindm _cgo_bindm
-var x_cgo_bindm_trampoline byte
-var _cgo_bindm = &x_cgo_bindm_trampoline
-
 // TODO: decide if we need x_cgo_set_context_function
 // TODO: decide if we need _cgo_yield
-
-var (
-	// In Go 1.20 the race detector was rewritten to pure Go
-	// on darwin. This means that when CGO_ENABLED=0 is set
-	// fakecgo is built with race detector code. This is not
-	// good since this code is pretending to be C. The go:norace
-	// pragma is not enough, since it only applies to the native
-	// ABIInternal function. The ABIO wrapper (which is necessary,
-	// since all references to text symbols from assembly will use it)
-	// does not inherit the go:norace pragma, so it will still be
-	// instrumented by the race detector.
-	//
-	// To circumvent this issue, using closure calls in the
-	// assembly, which forces the compiler to use the ABIInternal
-	// native implementation (which has go:norace) instead.
-	threadentry_call        = threadentry
-	x_cgo_init_call         = x_cgo_init
-	x_cgo_setenv_call       = x_cgo_setenv
-	x_cgo_unsetenv_call     = x_cgo_unsetenv
-	x_cgo_thread_start_call = x_cgo_thread_start
-)
diff --git a/internal/fakecgo/gen.go b/internal/fakecgo/gen.go
index 59f1f44d..428ac01d 100644
--- a/internal/fakecgo/gen.go
+++ b/internal/fakecgo/gen.go
@@ -17,6 +17,75 @@ import (
 	"text/template"
 )
 
+var templateArchTrampolines = template.Must(template.New("arch_trampolines").Funcs(funcs).Parse(
+	`// Code generated by 'go generate' with gen.go. DO NOT EDIT.
+
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: 2022 The Ebitengine Authors
+
+//go:build {{.Tag}}
+
+#include "textflag.h"
+#include "go_asm.h"
+
+// these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
+{{ range .Symbols }}
+{{- if $.Arch.StackBased }}
+{{- if eq .ArgsCount 0 }}
+TEXT _x_cgo_{{.Name}}_trampoline(SB), NOSPLIT, $0-0
+	CALL ·x_cgo_{{.Name}}(SB)
+	RET
+{{- else }}
+TEXT _x_cgo_{{.Name}}_trampoline(SB), NOSPLIT, ${{mul .ArgsCount $.Arch.WordSize}}-0
+{{- if eq $.Arch.Goarch "arm" }}
+{{- range $i := .ArgsCount }}
+	{{- $dstOffset := add 4 (mul $i $.Arch.WordSize) }}
+	{{$.Arch.MOV}} R{{$i}}, {{$dstOffset}}(R13)
+{{- end }}
+	{{$.Arch.MOV}} ·_cgo_{{.Name}}_call(SB), {{$.Arch.VolatileReg}}
+	{{$.Arch.MOV}} ({{$.Arch.VolatileReg}}), {{$.Arch.VolatileReg}}
+{{- if $.Arch.CallNeedsParens }}
+	CALL ({{$.Arch.VolatileReg}})
+{{- else }}
+	CALL {{$.Arch.VolatileReg}}
+{{- end }}
+	RET
+{{- else }}
+{{- $frameSize := mul .ArgsCount $.Arch.WordSize }}
+{{- $retAddrSize := $.Arch.WordSize }}
+{{- range $i := .ArgsCount }}
+	{{- $srcOffset := add $frameSize $retAddrSize }}
+	{{- $srcOffset = add $srcOffset (mul $i $.Arch.WordSize) }}
+	{{- $dstOffset := mul $i $.Arch.WordSize }}
+	{{$.Arch.MOV}} {{$srcOffset}}(SP), AX
+	{{$.Arch.MOV}} AX, {{$dstOffset}}(SP)
+{{- end }}
+	{{$.Arch.MOV}} ·_cgo_{{.Name}}_call(SB), {{$.Arch.VolatileReg}}
+	{{$.Arch.MOV}} ({{$.Arch.VolatileReg}}), {{$.Arch.VolatileReg}}
+	CALL {{$.Arch.VolatileReg}}
+	RET
+{{- end }}
+{{- end }}
+{{- else }}
+TEXT _x_cgo_{{.Name}}_trampoline(SB), NOSPLIT, ${{mul .ArgsCount $.Arch.WordSize}}
+{{- range $i := .ArgsCount }}
+	{{- $src := index $.Arch.C.IntRegArgs $i }}
+	{{- $dst := index $.Arch.GoABI0.IntRegArgs $i }}
+	{{- if ne $src $dst }}
+	{{$.Arch.MOV}} {{$src}}, {{$dst}}
+	{{- end }}
+{{- end }}
+	{{$.Arch.MOV}} ·_cgo_{{.Name}}_call(SB), {{$.Arch.VolatileReg}}
+	{{$.Arch.MOV}} ({{$.Arch.VolatileReg}}), {{$.Arch.VolatileReg}}
+{{- if $.Arch.CallNeedsParens }}
+	CALL ({{$.Arch.VolatileReg}})
+{{- else }}
+	CALL {{$.Arch.VolatileReg}}
+{{- end }}
+	RET
+{{- end }}
+{{ end }}`))
+
 var templateSymbols = template.Must(template.New("symbols").Funcs(funcs).Parse(
 	`// Code generated by 'go generate' with gen.go. DO NOT EDIT.
 
@@ -99,6 +168,35 @@ TEXT _{{.Name}}(SB), NOSPLIT|NOFRAME, $0-0
 {{ end -}}
 `))
 
+var templateCallbacks = template.Must(template.New("callbacks").Funcs(funcs).Parse(
+	`// Code generated by 'go generate' with gen.go. DO NOT EDIT.
+
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: 2022 The Ebitengine Authors
+
+//go:build !cgo && (darwin || freebsd || linux || netbsd)
+
+package fakecgo
+
+import (
+	_ "unsafe"
+)
+
+{{ range .Symbols }}
+//go:linkname _x_cgo_{{.Name}}_trampoline _x_cgo_{{.Name}}_trampoline
+//go:linkname _cgo_{{.Name}} {{if .Package}}{{.Package}}.{{end}}_cgo_{{.Name}}
+var _x_cgo_{{.Name}}_trampoline byte
+var _cgo_{{.Name}} = &_x_cgo_{{.Name}}_trampoline
+{{- end }}
+
+var (
+	threadentry_call = threadentry
+{{- range .Symbols }}
+	_cgo_{{.Name}}_call = x_cgo_{{.Name}}
+{{- end }}
+)
+`))
+
 type Arg struct {
 	Name string
 	Type string
@@ -142,6 +240,8 @@ var (
 var funcs = map[string]any{
 	"hasPrefix": strings.HasPrefix,
 	"imports":   imports,
+	"mul":       func(a, b int) int { return a * b },
+	"add":       func(a, b int) int { return a + b },
 }
 
 var GOOSes = []string{"darwin", "freebsd", "linux", "netbsd"}
@@ -285,9 +385,133 @@ func run() error {
 			}
 		}
 	}
+	for _, arch := range archs {
+		if err := writeArchTrampolines(arch); err != nil {
+			return err
+		}
+	}
+	if err := execute(templateCallbacks, "zcallbacks.go", struct{ Symbols []AsmGoSymbol }{Symbols: asmGoSymbols}); err != nil {
+		return err
+	}
 	return nil
 }
 
+type Arch struct {
+	Goarch          string
+	GoABI0          ABI
+	C               ABI
+	WordSize        int    // 4 on 32-bit systems, 8 on 64-bit systems
+	MOV             string // MOV instruction, e.g., "MOVL" or "MOVQ"
+	VolatileReg     string // Scratch register for intermediate values
+	CallNeedsParens bool
+	StackBased      bool // true if using stack-based calling convention (e.g., 386)
+}
+
+type ABI struct {
+	IntRegArgs [5]string // Name of integer argument registers in order. If empty, stack-based calling convention is used.
+	OutRegArg  string    // Name of the register for the single return value. If empty, stack-based calling convention is used.
+}
+
+// AsmGoSymbols are symbols that called from Go Assembly.
+type AsmGoSymbol struct {
+	Name      string
+	ArgsCount int
+	Package   string
+}
+
+var (
+	archs = []Arch{
+		{
+			Goarch:      "386",
+			WordSize:    4,
+			GoABI0:      ABI{}, // stack-based calling convention
+			C:           ABI{}, // stack-based calling convention
+			MOV:         "MOVL",
+			VolatileReg: "CX",
+			StackBased:  true,
+		},
+		{
+			Goarch:      "amd64",
+			WordSize:    8,
+			GoABI0:      ABI{IntRegArgs: [5]string{"AX", "BX", "CX", "DX", "SI"}, OutRegArg: "AX"},
+			C:           ABI{IntRegArgs: [5]string{"DI", "SI", "DX", "CX", "R8"}, OutRegArg: "AX"},
+			MOV:         "MOVQ",
+			VolatileReg: "R11",
+		},
+		{
+			Goarch:          "arm",
+			WordSize:        4,
+			GoABI0:          ABI{}, // stack-based calling convention
+			C:               ABI{}, // stack-based calling convention
+			MOV:             "MOVW",
+			VolatileReg:     "R12",
+			StackBased:      true,
+			CallNeedsParens: true,
+		},
+		{
+			Goarch:      "arm64",
+			WordSize:    8,
+			GoABI0:      ABI{IntRegArgs: [5]string{"R0", "R1", "R2", "R3", "R4"}, OutRegArg: "R0"},
+			C:           ABI{IntRegArgs: [5]string{"R0", "R1", "R2", "R3", "R4"}, OutRegArg: "R0"},
+			MOV:         "MOVD",
+			VolatileReg: "R9",
+		},
+		{
+			Goarch:          "loong64",
+			WordSize:        8,
+			GoABI0:          ABI{IntRegArgs: [5]string{"R4", "R5", "R6", "R7", "R8"}, OutRegArg: "R4"},
+			C:               ABI{IntRegArgs: [5]string{"R4", "R5", "R6", "R7", "R8"}, OutRegArg: "R4"},
+			MOV:             "MOVV",
+			VolatileReg:     "R23",
+			CallNeedsParens: true,
+		},
+		{
+			Goarch:      "riscv64",
+			WordSize:    8,
+			GoABI0:      ABI{IntRegArgs: [5]string{"X10", "X11", "X12", "X13", "X14"}, OutRegArg: "X10"},
+			C:           ABI{IntRegArgs: [5]string{"X10", "X11", "X12", "X13", "X14"}, OutRegArg: "X10"},
+			MOV:         "MOV ",
+			VolatileReg: "X5",
+		},
+	}
+	asmGoSymbols = []AsmGoSymbol{
+		{"init", 2, ""},
+		{"thread_start", 1, ""},
+		{"setenv", 1, "runtime"},
+		{"unsetenv", 1, "runtime"},
+		{"notify_runtime_init_done", 0, ""},
+		{"bindm", 0, ""},
+	}
+)
+
+func writeArchTrampolines(arch Arch) error {
+	var gooses []string
+	switch arch.Goarch {
+	case "amd64", "arm64":
+		gooses = []string{"darwin", "freebsd", "linux"}
+	case "loong64", "riscv64":
+		gooses = []string{"linux"}
+	case "386":
+		gooses = []string{"freebsd", "linux"}
+	default:
+		gooses = GOOSes
+	}
+	tag := "!cgo && (" + strings.Join(gooses, " || ") + ")"
+	if len(gooses) == 1 {
+		tag = "!cgo && " + gooses[0]
+	}
+	data := struct {
+		Tag     string
+		Symbols []AsmGoSymbol
+		Arch    Arch
+	}{
+		Tag:     tag,
+		Symbols: asmGoSymbols,
+		Arch:    arch,
+	}
+	return execute(templateArchTrampolines, fmt.Sprintf("ztrampolines_%s.s", arch.Goarch), data)
+}
+
 func main() {
 	if err := run(); err != nil {
 		log.Fatal(err)
diff --git a/internal/fakecgo/setenv.go b/internal/fakecgo/setenv.go
deleted file mode 100644
index 82308b8c..00000000
--- a/internal/fakecgo/setenv.go
+++ /dev/null
@@ -1,19 +0,0 @@
-// Copyright 2011 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.
-
-//go:build !cgo && (darwin || freebsd || linux || netbsd)
-
-package fakecgo
-
-import _ "unsafe" // for go:linkname
-
-//go:linkname x_cgo_setenv_trampoline x_cgo_setenv_trampoline
-//go:linkname _cgo_setenv runtime._cgo_setenv
-var x_cgo_setenv_trampoline byte
-var _cgo_setenv = &x_cgo_setenv_trampoline
-
-//go:linkname x_cgo_unsetenv_trampoline x_cgo_unsetenv_trampoline
-//go:linkname _cgo_unsetenv runtime._cgo_unsetenv
-var x_cgo_unsetenv_trampoline byte
-var _cgo_unsetenv = &x_cgo_unsetenv_trampoline
diff --git a/internal/fakecgo/trampolines_386.s b/internal/fakecgo/trampolines_386.s
index cd3492ea..4683d429 100644
--- a/internal/fakecgo/trampolines_386.s
+++ b/internal/fakecgo/trampolines_386.s
@@ -24,51 +24,6 @@
 // Go ABI0 on 386 expects arguments starting at 0(FP) which equals N+4(SP)
 // after the prologue (where N is the local frame size).
 
-TEXT x_cgo_init_trampoline(SB), NOSPLIT, $8-0
-	// C args at 12(SP) and 16(SP) after frame setup (8 bytes local + 4 bytes ret addr)
-	// Go function expects args at 0(SP) and 4(SP) in local frame
-	MOVL 12(SP), AX               // first C arg
-	MOVL 16(SP), BX               // second C arg
-	MOVL AX, 0(SP)                // Go arg 1
-	MOVL BX, 4(SP)                // Go arg 2
-	MOVL ·x_cgo_init_call(SB), CX
-	MOVL (CX), CX
-	CALL CX
-	RET
-
-TEXT x_cgo_thread_start_trampoline(SB), NOSPLIT, $4-0
-	// C args at 8(SP) after frame setup (4 bytes local + 4 bytes ret addr)
-	MOVL 8(SP), AX                        // first C arg
-	MOVL AX, 0(SP)                        // Go arg 1
-	MOVL ·x_cgo_thread_start_call(SB), CX
-	MOVL (CX), CX
-	CALL CX
-	RET
-
-TEXT x_cgo_setenv_trampoline(SB), NOSPLIT, $4-0
-	MOVL 8(SP), AX                  // first C arg
-	MOVL AX, 0(SP)                  // Go arg 1
-	MOVL ·x_cgo_setenv_call(SB), CX
-	MOVL (CX), CX
-	CALL CX
-	RET
-
-TEXT x_cgo_unsetenv_trampoline(SB), NOSPLIT, $4-0
-	MOVL 8(SP), AX                    // first C arg
-	MOVL AX, 0(SP)                    // Go arg 1
-	MOVL ·x_cgo_unsetenv_call(SB), CX
-	MOVL (CX), CX
-	CALL CX
-	RET
-
-TEXT x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0-0
-	CALL ·x_cgo_notify_runtime_init_done(SB)
-	RET
-
-TEXT x_cgo_bindm_trampoline(SB), NOSPLIT, $0
-	CALL ·x_cgo_bindm(SB)
-	RET
-
 // func setg_trampoline(setg uintptr, g uintptr)
 // This is called from Go, so args are at normal FP positions
 TEXT ·setg_trampoline(SB), NOSPLIT, $4-8
diff --git a/internal/fakecgo/trampolines_amd64.s b/internal/fakecgo/trampolines_amd64.s
index e4e4c75a..71f59cd8 100644
--- a/internal/fakecgo/trampolines_amd64.s
+++ b/internal/fakecgo/trampolines_amd64.s
@@ -24,41 +24,6 @@ temporary register is R11
 
 // these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
 
-TEXT x_cgo_init_trampoline(SB), NOSPLIT, $16
-	MOVQ DI, AX
-	MOVQ SI, BX
-	MOVQ ·x_cgo_init_call(SB), R11
-	MOVQ (R11), R11
-	CALL R11
-	RET
-
-TEXT x_cgo_thread_start_trampoline(SB), NOSPLIT, $8
-	MOVQ DI, AX
-	MOVQ ·x_cgo_thread_start_call(SB), R11
-	MOVQ (R11), R11
-	CALL R11
-	RET
-
-TEXT x_cgo_setenv_trampoline(SB), NOSPLIT, $8
-	MOVQ DI, AX
-	MOVQ ·x_cgo_setenv_call(SB), R11
-	MOVQ (R11), R11
-	CALL R11
-	RET
-
-TEXT x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8
-	MOVQ DI, AX
-	MOVQ ·x_cgo_unsetenv_call(SB), R11
-	MOVQ (R11), R11
-	CALL R11
-	RET
-
-TEXT x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0
-	JMP ·x_cgo_notify_runtime_init_done(SB)
-
-TEXT x_cgo_bindm_trampoline(SB), NOSPLIT, $0
-	JMP ·x_cgo_bindm(SB)
-
 // func setg_trampoline(setg uintptr, g uintptr)
 TEXT ·setg_trampoline(SB), NOSPLIT, $0-16
 	MOVQ G+8(FP), DI
diff --git a/internal/fakecgo/trampolines_arm.s b/internal/fakecgo/trampolines_arm.s
index 00b3177e..3035d7b8 100644
--- a/internal/fakecgo/trampolines_arm.s
+++ b/internal/fakecgo/trampolines_arm.s
@@ -10,43 +10,6 @@
 // On ARM32, Go ABI0 uses stack-based calling convention.
 // Arguments are placed on the stack starting at 4(SP) after the prologue.
 
-TEXT x_cgo_init_trampoline(SB), NOSPLIT, $8-0
-	MOVW R0, 4(R13)
-	MOVW R1, 8(R13)
-	MOVW ·x_cgo_init_call(SB), R12
-	MOVW (R12), R12
-	CALL (R12)
-	RET
-
-TEXT x_cgo_thread_start_trampoline(SB), NOSPLIT, $8-0
-	MOVW R0, 4(R13)
-	MOVW ·x_cgo_thread_start_call(SB), R12
-	MOVW (R12), R12
-	CALL (R12)
-	RET
-
-TEXT x_cgo_setenv_trampoline(SB), NOSPLIT, $8-0
-	MOVW R0, 4(R13)
-	MOVW ·x_cgo_setenv_call(SB), R12
-	MOVW (R12), R12
-	CALL (R12)
-	RET
-
-TEXT x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8-0
-	MOVW R0, 4(R13)
-	MOVW ·x_cgo_unsetenv_call(SB), R12
-	MOVW (R12), R12
-	CALL (R12)
-	RET
-
-TEXT x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0-0
-	CALL ·x_cgo_notify_runtime_init_done(SB)
-	RET
-
-TEXT x_cgo_bindm_trampoline(SB), NOSPLIT, $0
-	CALL ·x_cgo_bindm(SB)
-	RET
-
 // func setg_trampoline(setg uintptr, g uintptr)
 TEXT ·setg_trampoline(SB), NOSPLIT, $0-8
 	MOVW G+4(FP), R0
diff --git a/internal/fakecgo/trampolines_arm64.s b/internal/fakecgo/trampolines_arm64.s
index dceb1cac..cc6f1eca 100644
--- a/internal/fakecgo/trampolines_arm64.s
+++ b/internal/fakecgo/trampolines_arm64.s
@@ -11,38 +11,6 @@
 // Note that C arguments are passed in R0-R7, which matches Go ABIInternal for the first eight arguments.
 // R9 is used as a temporary register.
 
-TEXT x_cgo_init_trampoline(SB), NOSPLIT, $0-0
-	MOVD ·x_cgo_init_call(SB), R9
-	MOVD (R9), R9
-	CALL R9
-	RET
-
-TEXT x_cgo_thread_start_trampoline(SB), NOSPLIT, $0-0
-	MOVD ·x_cgo_thread_start_call(SB), R9
-	MOVD (R9), R9
-	CALL R9
-	RET
-
-TEXT x_cgo_setenv_trampoline(SB), NOSPLIT, $0-0
-	MOVD ·x_cgo_setenv_call(SB), R9
-	MOVD (R9), R9
-	CALL R9
-	RET
-
-TEXT x_cgo_unsetenv_trampoline(SB), NOSPLIT, $0-0
-	MOVD ·x_cgo_unsetenv_call(SB), R9
-	MOVD (R9), R9
-	CALL R9
-	RET
-
-TEXT x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0-0
-	CALL ·x_cgo_notify_runtime_init_done(SB)
-	RET
-
-TEXT x_cgo_bindm_trampoline(SB), NOSPLIT, $0
-	CALL ·x_cgo_bindm(SB)
-	RET
-
 // func setg_trampoline(setg uintptr, g uintptr)
 TEXT ·setg_trampoline(SB), NOSPLIT, $0-16
 	MOVD G+8(FP), R0
diff --git a/internal/fakecgo/trampolines_loong64.s b/internal/fakecgo/trampolines_loong64.s
index 7596f0da..889b5887 100644
--- a/internal/fakecgo/trampolines_loong64.s
+++ b/internal/fakecgo/trampolines_loong64.s
@@ -10,43 +10,6 @@
 // these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
 // R23 is used as temporary register.
 
-TEXT x_cgo_init_trampoline(SB), NOSPLIT, $16
-	MOVV R4, 8(R3)
-	MOVV R5, 16(R3)
-	MOVV ·x_cgo_init_call(SB), R23
-	MOVV (R23), R23
-	CALL (R23)
-	RET
-
-TEXT x_cgo_thread_start_trampoline(SB), NOSPLIT, $8
-	MOVV R4, 8(R3)
-	MOVV ·x_cgo_thread_start_call(SB), R23
-	MOVV (R23), R23
-	CALL (R23)
-	RET
-
-TEXT x_cgo_setenv_trampoline(SB), NOSPLIT, $8
-	MOVV R4, 8(R3)
-	MOVV ·x_cgo_setenv_call(SB), R23
-	MOVV (R23), R23
-	CALL (R23)
-	RET
-
-TEXT x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8
-	MOVV R4, 8(R3)
-	MOVV ·x_cgo_unsetenv_call(SB), R23
-	MOVV (R23), R23
-	CALL (R23)
-	RET
-
-TEXT x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0
-	CALL ·x_cgo_notify_runtime_init_done(SB)
-	RET
-
-TEXT x_cgo_bindm_trampoline(SB), NOSPLIT, $0
-	CALL ·x_cgo_bindm(SB)
-	RET
-
 // func setg_trampoline(setg uintptr, g uintptr)
 TEXT ·setg_trampoline(SB), NOSPLIT, $0
 	MOVV G+8(FP), R4
diff --git a/internal/fakecgo/trampolines_riscv64.s b/internal/fakecgo/trampolines_riscv64.s
index 9298f8c7..004549b6 100644
--- a/internal/fakecgo/trampolines_riscv64.s
+++ b/internal/fakecgo/trampolines_riscv64.s
@@ -9,43 +9,6 @@
 // these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
 // X5 is used as temporary register.
 
-TEXT x_cgo_init_trampoline(SB), NOSPLIT, $16
-	MOV  X10, 8(SP)
-	MOV  X11, 16(SP)
-	MOV  ·x_cgo_init_call(SB), X5
-	MOV  (X5), X5
-	CALL X5
-	RET
-
-TEXT x_cgo_thread_start_trampoline(SB), NOSPLIT, $8
-	MOV  X10, 8(SP)
-	MOV  ·x_cgo_thread_start_call(SB), X5
-	MOV  (X5), X5
-	CALL X5
-	RET
-
-TEXT x_cgo_setenv_trampoline(SB), NOSPLIT, $8
-	MOV  X10, 8(SP)
-	MOV  ·x_cgo_setenv_call(SB), X5
-	MOV  (X5), X5
-	CALL X5
-	RET
-
-TEXT x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8
-	MOV  X10, 8(SP)
-	MOV  ·x_cgo_unsetenv_call(SB), X5
-	MOV  (X5), X5
-	CALL X5
-	RET
-
-TEXT x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0
-	CALL ·x_cgo_notify_runtime_init_done(SB)
-	RET
-
-TEXT x_cgo_bindm_trampoline(SB), NOSPLIT, $0
-	CALL ·x_cgo_bindm(SB)
-	RET
-
 // func setg_trampoline(setg uintptr, g uintptr)
 TEXT ·setg_trampoline(SB), NOSPLIT, $0
 	MOV  gp+8(FP), X10
diff --git a/internal/fakecgo/zcallbacks.go b/internal/fakecgo/zcallbacks.go
new file mode 100644
index 00000000..94dc189c
--- /dev/null
+++ b/internal/fakecgo/zcallbacks.go
@@ -0,0 +1,52 @@
+// Code generated by 'go generate' with gen.go. DO NOT EDIT.
+
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: 2022 The Ebitengine Authors
+
+//go:build !cgo && (darwin || freebsd || linux || netbsd)
+
+package fakecgo
+
+import (
+	_ "unsafe"
+)
+
+//go:linkname _x_cgo_init_trampoline _x_cgo_init_trampoline
+//go:linkname _cgo_init _cgo_init
+var _x_cgo_init_trampoline byte
+var _cgo_init = &_x_cgo_init_trampoline
+
+//go:linkname _x_cgo_thread_start_trampoline _x_cgo_thread_start_trampoline
+//go:linkname _cgo_thread_start _cgo_thread_start
+var _x_cgo_thread_start_trampoline byte
+var _cgo_thread_start = &_x_cgo_thread_start_trampoline
+
+//go:linkname _x_cgo_setenv_trampoline _x_cgo_setenv_trampoline
+//go:linkname _cgo_setenv runtime._cgo_setenv
+var _x_cgo_setenv_trampoline byte
+var _cgo_setenv = &_x_cgo_setenv_trampoline
+
+//go:linkname _x_cgo_unsetenv_trampoline _x_cgo_unsetenv_trampoline
+//go:linkname _cgo_unsetenv runtime._cgo_unsetenv
+var _x_cgo_unsetenv_trampoline byte
+var _cgo_unsetenv = &_x_cgo_unsetenv_trampoline
+
+//go:linkname _x_cgo_notify_runtime_init_done_trampoline _x_cgo_notify_runtime_init_done_trampoline
+//go:linkname _cgo_notify_runtime_init_done _cgo_notify_runtime_init_done
+var _x_cgo_notify_runtime_init_done_trampoline byte
+var _cgo_notify_runtime_init_done = &_x_cgo_notify_runtime_init_done_trampoline
+
+//go:linkname _x_cgo_bindm_trampoline _x_cgo_bindm_trampoline
+//go:linkname _cgo_bindm _cgo_bindm
+var _x_cgo_bindm_trampoline byte
+var _cgo_bindm = &_x_cgo_bindm_trampoline
+
+var (
+	threadentry_call                   = threadentry
+	_cgo_init_call                     = x_cgo_init
+	_cgo_thread_start_call             = x_cgo_thread_start
+	_cgo_setenv_call                   = x_cgo_setenv
+	_cgo_unsetenv_call                 = x_cgo_unsetenv
+	_cgo_notify_runtime_init_done_call = x_cgo_notify_runtime_init_done
+	_cgo_bindm_call                    = x_cgo_bindm
+)
diff --git a/internal/fakecgo/ztrampolines_386.s b/internal/fakecgo/ztrampolines_386.s
new file mode 100644
index 00000000..070d48d3
--- /dev/null
+++ b/internal/fakecgo/ztrampolines_386.s
@@ -0,0 +1,53 @@
+// Code generated by 'go generate' with gen.go. DO NOT EDIT.
+
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: 2022 The Ebitengine Authors
+
+//go:build !cgo && (freebsd || linux)
+
+#include "textflag.h"
+#include "go_asm.h"
+
+// these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
+
+TEXT _x_cgo_init_trampoline(SB), NOSPLIT, $8-0
+	MOVL 12(SP), AX
+	MOVL AX, 0(SP)
+	MOVL 16(SP), AX
+	MOVL AX, 4(SP)
+	MOVL ·_cgo_init_call(SB), CX
+	MOVL (CX), CX
+	CALL CX
+	RET
+
+TEXT _x_cgo_thread_start_trampoline(SB), NOSPLIT, $4-0
+	MOVL 8(SP), AX
+	MOVL AX, 0(SP)
+	MOVL ·_cgo_thread_start_call(SB), CX
+	MOVL (CX), CX
+	CALL CX
+	RET
+
+TEXT _x_cgo_setenv_trampoline(SB), NOSPLIT, $4-0
+	MOVL 8(SP), AX
+	MOVL AX, 0(SP)
+	MOVL ·_cgo_setenv_call(SB), CX
+	MOVL (CX), CX
+	CALL CX
+	RET
+
+TEXT _x_cgo_unsetenv_trampoline(SB), NOSPLIT, $4-0
+	MOVL 8(SP), AX
+	MOVL AX, 0(SP)
+	MOVL ·_cgo_unsetenv_call(SB), CX
+	MOVL (CX), CX
+	CALL CX
+	RET
+
+TEXT _x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0-0
+	CALL ·x_cgo_notify_runtime_init_done(SB)
+	RET
+
+TEXT _x_cgo_bindm_trampoline(SB), NOSPLIT, $0-0
+	CALL ·x_cgo_bindm(SB)
+	RET
diff --git a/internal/fakecgo/ztrampolines_amd64.s b/internal/fakecgo/ztrampolines_amd64.s
new file mode 100644
index 00000000..9775fd22
--- /dev/null
+++ b/internal/fakecgo/ztrampolines_amd64.s
@@ -0,0 +1,52 @@
+// Code generated by 'go generate' with gen.go. DO NOT EDIT.
+
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: 2022 The Ebitengine Authors
+
+//go:build !cgo && (darwin || freebsd || linux)
+
+#include "textflag.h"
+#include "go_asm.h"
+
+// these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
+
+TEXT _x_cgo_init_trampoline(SB), NOSPLIT, $16
+	MOVQ DI, AX
+	MOVQ SI, BX
+	MOVQ ·_cgo_init_call(SB), R11
+	MOVQ (R11), R11
+	CALL R11
+	RET
+
+TEXT _x_cgo_thread_start_trampoline(SB), NOSPLIT, $8
+	MOVQ DI, AX
+	MOVQ ·_cgo_thread_start_call(SB), R11
+	MOVQ (R11), R11
+	CALL R11
+	RET
+
+TEXT _x_cgo_setenv_trampoline(SB), NOSPLIT, $8
+	MOVQ DI, AX
+	MOVQ ·_cgo_setenv_call(SB), R11
+	MOVQ (R11), R11
+	CALL R11
+	RET
+
+TEXT _x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8
+	MOVQ DI, AX
+	MOVQ ·_cgo_unsetenv_call(SB), R11
+	MOVQ (R11), R11
+	CALL R11
+	RET
+
+TEXT _x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0
+	MOVQ ·_cgo_notify_runtime_init_done_call(SB), R11
+	MOVQ (R11), R11
+	CALL R11
+	RET
+
+TEXT _x_cgo_bindm_trampoline(SB), NOSPLIT, $0
+	MOVQ ·_cgo_bindm_call(SB), R11
+	MOVQ (R11), R11
+	CALL R11
+	RET
diff --git a/internal/fakecgo/ztrampolines_arm.s b/internal/fakecgo/ztrampolines_arm.s
new file mode 100644
index 00000000..4e860582
--- /dev/null
+++ b/internal/fakecgo/ztrampolines_arm.s
@@ -0,0 +1,48 @@
+// Code generated by 'go generate' with gen.go. DO NOT EDIT.
+
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: 2022 The Ebitengine Authors
+
+//go:build !cgo && (darwin || freebsd || linux || netbsd)
+
+#include "textflag.h"
+#include "go_asm.h"
+
+// these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
+
+TEXT _x_cgo_init_trampoline(SB), NOSPLIT, $8-0
+	MOVW R0, 4(R13)
+	MOVW R1, 8(R13)
+	MOVW ·_cgo_init_call(SB), R12
+	MOVW (R12), R12
+	CALL (R12)
+	RET
+
+TEXT _x_cgo_thread_start_trampoline(SB), NOSPLIT, $4-0
+	MOVW R0, 4(R13)
+	MOVW ·_cgo_thread_start_call(SB), R12
+	MOVW (R12), R12
+	CALL (R12)
+	RET
+
+TEXT _x_cgo_setenv_trampoline(SB), NOSPLIT, $4-0
+	MOVW R0, 4(R13)
+	MOVW ·_cgo_setenv_call(SB), R12
+	MOVW (R12), R12
+	CALL (R12)
+	RET
+
+TEXT _x_cgo_unsetenv_trampoline(SB), NOSPLIT, $4-0
+	MOVW R0, 4(R13)
+	MOVW ·_cgo_unsetenv_call(SB), R12
+	MOVW (R12), R12
+	CALL (R12)
+	RET
+
+TEXT _x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0-0
+	CALL ·x_cgo_notify_runtime_init_done(SB)
+	RET
+
+TEXT _x_cgo_bindm_trampoline(SB), NOSPLIT, $0-0
+	CALL ·x_cgo_bindm(SB)
+	RET
diff --git a/internal/fakecgo/ztrampolines_arm64.s b/internal/fakecgo/ztrampolines_arm64.s
new file mode 100644
index 00000000..e829903e
--- /dev/null
+++ b/internal/fakecgo/ztrampolines_arm64.s
@@ -0,0 +1,47 @@
+// Code generated by 'go generate' with gen.go. DO NOT EDIT.
+
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: 2022 The Ebitengine Authors
+
+//go:build !cgo && (darwin || freebsd || linux)
+
+#include "textflag.h"
+#include "go_asm.h"
+
+// these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
+
+TEXT _x_cgo_init_trampoline(SB), NOSPLIT, $16
+	MOVD ·_cgo_init_call(SB), R9
+	MOVD (R9), R9
+	CALL R9
+	RET
+
+TEXT _x_cgo_thread_start_trampoline(SB), NOSPLIT, $8
+	MOVD ·_cgo_thread_start_call(SB), R9
+	MOVD (R9), R9
+	CALL R9
+	RET
+
+TEXT _x_cgo_setenv_trampoline(SB), NOSPLIT, $8
+	MOVD ·_cgo_setenv_call(SB), R9
+	MOVD (R9), R9
+	CALL R9
+	RET
+
+TEXT _x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8
+	MOVD ·_cgo_unsetenv_call(SB), R9
+	MOVD (R9), R9
+	CALL R9
+	RET
+
+TEXT _x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0
+	MOVD ·_cgo_notify_runtime_init_done_call(SB), R9
+	MOVD (R9), R9
+	CALL R9
+	RET
+
+TEXT _x_cgo_bindm_trampoline(SB), NOSPLIT, $0
+	MOVD ·_cgo_bindm_call(SB), R9
+	MOVD (R9), R9
+	CALL R9
+	RET
diff --git a/internal/fakecgo/ztrampolines_loong64.s b/internal/fakecgo/ztrampolines_loong64.s
new file mode 100644
index 00000000..82b375ad
--- /dev/null
+++ b/internal/fakecgo/ztrampolines_loong64.s
@@ -0,0 +1,47 @@
+// Code generated by 'go generate' with gen.go. DO NOT EDIT.
+
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: 2022 The Ebitengine Authors
+
+//go:build !cgo && linux
+
+#include "textflag.h"
+#include "go_asm.h"
+
+// these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
+
+TEXT _x_cgo_init_trampoline(SB), NOSPLIT, $16
+	MOVV ·_cgo_init_call(SB), R23
+	MOVV (R23), R23
+	CALL (R23)
+	RET
+
+TEXT _x_cgo_thread_start_trampoline(SB), NOSPLIT, $8
+	MOVV ·_cgo_thread_start_call(SB), R23
+	MOVV (R23), R23
+	CALL (R23)
+	RET
+
+TEXT _x_cgo_setenv_trampoline(SB), NOSPLIT, $8
+	MOVV ·_cgo_setenv_call(SB), R23
+	MOVV (R23), R23
+	CALL (R23)
+	RET
+
+TEXT _x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8
+	MOVV ·_cgo_unsetenv_call(SB), R23
+	MOVV (R23), R23
+	CALL (R23)
+	RET
+
+TEXT _x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0
+	MOVV ·_cgo_notify_runtime_init_done_call(SB), R23
+	MOVV (R23), R23
+	CALL (R23)
+	RET
+
+TEXT _x_cgo_bindm_trampoline(SB), NOSPLIT, $0
+	MOVV ·_cgo_bindm_call(SB), R23
+	MOVV (R23), R23
+	CALL (R23)
+	RET
diff --git a/internal/fakecgo/ztrampolines_riscv64.s b/internal/fakecgo/ztrampolines_riscv64.s
new file mode 100644
index 00000000..f062b89a
--- /dev/null
+++ b/internal/fakecgo/ztrampolines_riscv64.s
@@ -0,0 +1,47 @@
+// Code generated by 'go generate' with gen.go. DO NOT EDIT.
+
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: 2022 The Ebitengine Authors
+
+//go:build !cgo && linux
+
+#include "textflag.h"
+#include "go_asm.h"
+
+// these trampolines map the gcc ABI to Go ABI and then calls into the Go equivalent functions.
+
+TEXT _x_cgo_init_trampoline(SB), NOSPLIT, $16
+	MOV  ·_cgo_init_call(SB), X5
+	MOV  (X5), X5
+	CALL X5
+	RET
+
+TEXT _x_cgo_thread_start_trampoline(SB), NOSPLIT, $8
+	MOV  ·_cgo_thread_start_call(SB), X5
+	MOV  (X5), X5
+	CALL X5
+	RET
+
+TEXT _x_cgo_setenv_trampoline(SB), NOSPLIT, $8
+	MOV  ·_cgo_setenv_call(SB), X5
+	MOV  (X5), X5
+	CALL X5
+	RET
+
+TEXT _x_cgo_unsetenv_trampoline(SB), NOSPLIT, $8
+	MOV  ·_cgo_unsetenv_call(SB), X5
+	MOV  (X5), X5
+	CALL X5
+	RET
+
+TEXT _x_cgo_notify_runtime_init_done_trampoline(SB), NOSPLIT, $0
+	MOV  ·_cgo_notify_runtime_init_done_call(SB), X5
+	MOV  (X5), X5
+	CALL X5
+	RET
+
+TEXT _x_cgo_bindm_trampoline(SB), NOSPLIT, $0
+	MOV  ·_cgo_bindm_call(SB), X5
+	MOV  (X5), X5
+	CALL X5
+	RET