Fix ABI on 32-bit non-Windows x86: go back to trampoline size 10

The trampoline size is part of the ABI, so it cannot change.  Move the
logic from the stdcall and thiscall trampolines to the functions they
call, to reduce them both to 10 bytes.

This drops the previously added support for raw THISCALL closures on
non-Windows.  (Non-raw THISCALL closures still work.)

src/x86/win32.S

@@ -33,8 +33,13 @@
 #include <fficonfig.h>
 #include <ffi.h>
 
+#define CIF_ABI_OFFSET 0
+#define CIF_BYTES_OFFSET 16
+
 #ifdef _MSC_VER
 
+#define CLOSURE_CIF_OFFSET ((FFI_TRAMPOLINE_SIZE + 3) AND NOT 3)
+
 .386
 .MODEL FLAT, C
 
@@ -188,12 +193,12 @@ ca_epilogue:
         ret
 ffi_call_win32 ENDP
 
-ffi_closure_THISCALL PROC NEAR FORCEFRAME
-	sub	esp, 40
-	lea	edx, [ebp -24]
-	mov	[ebp - 12], edx	/* resp */
-	lea	edx, [ebp + 12]  /* account for stub return address on stack */
-	jmp	stub
+ffi_closure_THISCALL PROC NEAR
+	;; Insert the register argument on the stack as the first argument
+	xchg	DWORD PTR [esp+4], ecx
+	xchg	DWORD PTR [esp], ecx
+	push	ecx
+	jmp	ffi_closure_STDCALL
 ffi_closure_THISCALL ENDP
 
 ffi_closure_SYSV PROC NEAR FORCEFRAME
@@ -464,8 +469,18 @@ cd_retlongdouble:
         jmp   cd_epilogue
 
 cd_epilogue:
-        ;; Epilogue code is autogenerated.
-        ret
+        mov   esp, ebp
+        pop   ebp
+        pop   ecx
+        mov   ecx, DWORD PTR [ecx + (CLOSURE_CIF_OFFSET-10)]
+        cmp   DWORD PTR [ecx + CIF_ABI_OFFSET], 3
+        mov   ecx, DWORD PTR [ecx + CIF_BYTES_OFFSET]
+        jne   cd_not_thiscall
+        add   ecx, 4
+cd_not_thiscall:
+        pop   edx
+        add   esp, ecx
+        jmp   edx
 ffi_closure_STDCALL ENDP
 
 _TEXT ENDS
@@ -473,6 +488,8 @@ END
 
 #else
 
+#define CLOSURE_CIF_OFFSET ((FFI_TRAMPOLINE_SIZE + 3) & ~3)
+
 #if defined(SYMBOL_UNDERSCORE)
 #define USCORE_SYMBOL(x) _##x
 #else
@@ -657,13 +674,11 @@ FFI_HIDDEN(ffi_closure_THISCALL)
 	.def	_ffi_closure_THISCALL;	.scl	2;	.type	32;	.endef
 #endif
 USCORE_SYMBOL(ffi_closure_THISCALL):
-	pushl	%ebp
-	movl	%esp, %ebp
-	subl	$40, %esp
-	leal	-24(%ebp), %edx
-	movl	%edx, -12(%ebp)	/* resp */
-	leal	12(%ebp), %edx  /* account for stub return address on stack */
-	jmp	.stub
+	/* Insert the register argument on the stack as the first argument */
+	xchg	%ecx, 4(%esp)
+	xchg	%ecx, (%esp)
+	push	%ecx
+	jmp	.ffi_closure_STDCALL_internal
 .LFE1:
 
         # This assumes we are using gas.
@@ -685,7 +700,6 @@ USCORE_SYMBOL(ffi_closure_SYSV):
 	leal	-24(%ebp), %edx
 	movl	%edx, -12(%ebp)	/* resp */
 	leal	8(%ebp), %edx
-.stub:
 	movl	%edx, 4(%esp)	/* args = __builtin_dwarf_cfa () */
 	leal	-12(%ebp), %edx
 	movl	%edx, (%esp)	/* &resp */
@@ -811,6 +825,8 @@ USCORE_SYMBOL(ffi_closure_SYSV):
 #define RAW_CLOSURE_FUN_OFFSET (RAW_CLOSURE_CIF_OFFSET + 4)
 #define RAW_CLOSURE_USER_DATA_OFFSET (RAW_CLOSURE_FUN_OFFSET + 4)
 #define CIF_FLAGS_OFFSET 20
+
+#ifdef X86_WIN32
         .balign 16
 FFI_HIDDEN(ffi_closure_raw_THISCALL)
 	.globl	USCORE_SYMBOL(ffi_closure_raw_THISCALL)
@@ -827,6 +843,8 @@ USCORE_SYMBOL(ffi_closure_raw_THISCALL):
 	movl	%edx, 12(%esp)	/* user_data */
 	leal	12(%ebp), %edx	/* __builtin_dwarf_cfa () */
 	jmp	.stubraw
+#endif /* X86_WIN32 */
+
         # This assumes we are using gas.
         .balign 16
 #if defined(X86_WIN32)
@@ -958,6 +976,7 @@ FFI_HIDDEN(ffi_closure_STDCALL)
 	.def	_ffi_closure_STDCALL;	.scl	2;	.type	32;	.endef
 #endif
 USCORE_SYMBOL(ffi_closure_STDCALL):
+.ffi_closure_STDCALL_internal:
 .LFB5:
 	pushl	%ebp
 .LCFI9:
@@ -1070,7 +1089,15 @@ USCORE_SYMBOL(ffi_closure_STDCALL):
 .Lscls_epilogue:
 	movl	%ebp, %esp
 	popl	%ebp
-	ret
+	popl	%ecx
+	movl	(CLOSURE_CIF_OFFSET-10)(%ecx), %ecx
+	cmpl	$3, CIF_ABI_OFFSET(%ecx) /* FFI_THISCALL */
+	movl	CIF_BYTES_OFFSET(%ecx), %ecx
+	jne	1f
+	addl	$4, %ecx
+1:	popl	%edx
+	addl	%ecx, %esp
+	jmp	*%edx
 .ffi_closure_STDCALL_end:
 .LFE5: