1// SPDX-License-Identifier: Apache-2.0
2// SPDX-FileCopyrightText: 2026 The Ebitengine Authors
3
4//go:build linux
5
6#include "textflag.h"
7#include "go_asm.h"
8#include "funcdata.h"
9
10// S390X ELF ABI callbackasm1 implementation
11// On entry, R0 contains the callback index (set by callbackasm)
12// NOTE: We use R0 instead of R11 because R11 is callee-saved on S390X.
13//
14// S390X stack frame layout:
15// 0(R15) - back chain
16// 48(R15) - register save area (R6-R15)
17// 160(R15) - parameter area
18//
19// S390X uses R2-R6 for integer arguments (5 registers) and F0,F2,F4,F6 for floats (4 registers).
20//
21// Our frame layout (total 200 bytes, 8-byte aligned). To stay under the
22// 800-byte nosplit limit on Go 1.27 the callbackArgs struct is stored in the
23// caller's free linkage slot (old_R15+16..48) instead of our own frame:
24// 0(R15) - back chain
25// 48(R15) - saved R6-R15 (done by STMG)
26// 128(R15) - args array start
27//
28// Args array layout:
29// - floats F0,F2,F4,F6 (32 bytes)
30// - ints R2-R6 (40 bytes)
31// Total args array: 72 bytes, ends at 200
32//
33// callbackArgs struct lives at old_R15+16 (32 bytes: index, args, result, stackArgs)
34// Stack args in caller's frame start at old_R15+160
35
36#define FRAME_SIZE 200
37#define ARGS_ARRAY 128
38#define CB_ARGS (FRAME_SIZE+16)
39#define FLOAT_OFF 0
40#define INT_OFF 32
41
42TEXT callbackasm1(SB), NOSPLIT|NOFRAME, $0
43 NO_LOCAL_POINTERS
44
45 // On entry, the trampoline in zcallback_s390x.s left
46 // the callback index in R0 (NOT R11, since R11 is callee-saved).
47 // R6 contains the 5th integer argument.
48
49 // Save R6-R15 in caller's frame (per S390X ABI) BEFORE allocating our frame
50 // STMG stores R6's current value (the 5th arg) at 48(R15)
51 STMG R6, R15, 48(R15)
52
53 // Save current stack pointer (will be back chain)
54 MOVD R15, R1
55
56 // Allocate our stack frame
57 SUB $FRAME_SIZE, R15
58 MOVD R1, 0(R15) // back chain
59
60 // Save R0 (callback index) immediately - it's volatile
61 MOVD R0, (CB_ARGS+0)(R15)
62
63 // Save callback arguments to args array.
64 // Layout: floats first (F0,F2,F4,F6), then ints (R2-R6)
65 FMOVD F0, (ARGS_ARRAY+FLOAT_OFF+0*8)(R15)
66 FMOVD F2, (ARGS_ARRAY+FLOAT_OFF+1*8)(R15)
67 FMOVD F4, (ARGS_ARRAY+FLOAT_OFF+2*8)(R15)
68 FMOVD F6, (ARGS_ARRAY+FLOAT_OFF+3*8)(R15)
69
70 MOVD R2, (ARGS_ARRAY+INT_OFF+0*8)(R15)
71 MOVD R3, (ARGS_ARRAY+INT_OFF+1*8)(R15)
72 MOVD R4, (ARGS_ARRAY+INT_OFF+2*8)(R15)
73 MOVD R5, (ARGS_ARRAY+INT_OFF+3*8)(R15)
74
75 // R6 (5th int arg) was saved at 48(old_R15) by STMG
76 // old_R15 = current R15 + FRAME_SIZE
77 MOVD (48+FRAME_SIZE)(R15), R1
78 MOVD R1, (ARGS_ARRAY+INT_OFF+4*8)(R15)
79
80 // Finish setting up callbackArgs struct at CB_ARGS(R15)
81 // struct { index uintptr; args unsafe.Pointer; result uintptr; stackArgs unsafe.Pointer }
82 // Note: index was already saved earlier
83 ADD $ARGS_ARRAY, R15, R1
84 MOVD R1, (CB_ARGS+8)(R15) // args = address of register args
85 MOVD $0, (CB_ARGS+16)(R15) // result = 0
86
87 // stackArgs points to caller's stack arguments at old_R15+160 = R15+FRAME_SIZE+160
88 ADD $(FRAME_SIZE+160), R15, R1
89 MOVD R1, (CB_ARGS+24)(R15) // stackArgs = &caller_stack_args
90
91 // Call crosscall2 with arguments in registers:
92 // R2 = fn (from callbackWrap_call closure)
93 // R3 = frame (address of callbackArgs)
94 // R5 = ctxt (0)
95 MOVD ·callbackWrap_call(SB), R2
96 MOVD (R2), R2 // dereference closure to get fn
97 ADD $CB_ARGS, R15, R3 // frame = &callbackArgs
98 MOVD $0, R5 // ctxt = 0
99
100 BL crosscall2(SB)
101
102 // Get callback result into R2
103 MOVD (CB_ARGS+16)(R15), R2
104
105 // Deallocate frame
106 ADD $FRAME_SIZE, R15
107
108 // Restore R6-R15 from caller's frame
109 LMG 48(R15), R6, R15
110
111 RET