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// PPC64LE ELFv2 ABI callbackasm1 implementation
11// On entry, R11 contains the callback index (set by callbackasm)
12//
13// ELFv2 stack frame layout requirements:
14// 0(R1) - back chain (pointer to caller's frame)
15// 8(R1) - CR save area (optional)
16// 16(R1) - LR save area (for callee to save caller's LR)
17// 24(R1) - TOC save area (if needed)
18// 32(R1)+ - parameter save area / local variables
19//
20// Our frame (total 208 bytes, 16-byte aligned):
21// 32(R1) - saved R31 (8 bytes)
22// 40(R1) - callbackArgs struct (32 bytes: index, args, result, stackArgs)
23// 72(R1) - args array: floats (64) + ints (64) = 128 bytes, ends at 200
24// Total with alignment: 208 bytes
25//
26// Stack args are NOT copied - we pass a pointer to their location in caller's frame.
27// This keeps frame size small enough for NOSPLIT with CGO_ENABLED=1.
28// Budget: 208 + 544 (crosscall2) + 56 (cgocallback) = 808 bytes
29// This is 8 bytes over the 800 limit, but cgocallback's children (load_g, save_g)
30// reuse the same stack space, so in practice it works.
31
32#define FRAME_SIZE 200
33#define SAVE_R31 32
34#define CB_ARGS 40
35#define ARGS_ARRAY 72
36#define FLOAT_OFF 0
37#define INT_OFF 64
38
39TEXT callbackasm1(SB), NOSPLIT|NOFRAME, $0
40 NO_LOCAL_POINTERS
41
42 // On entry, the trampoline in zcallback_ppc64le.s left
43 // the callback index in R11.
44
45 // Per ELFv2 ABI, save LR to caller's frame BEFORE allocating our frame
46 MOVD LR, R0
47 MOVD R0, 16(R1)
48
49 // Allocate our stack frame (with back chain via MOVDU)
50 MOVDU R1, -FRAME_SIZE(R1)
51
52 // Save R31 - Go assembler uses it for MOVD from SB (like arm64's R27)
53 MOVD R31, SAVE_R31(R1)
54
55 // Save R11 (callback index) immediately - it's volatile and will be clobbered!
56 // Store it in the callbackArgs struct's index field now.
57 MOVD R11, (CB_ARGS+0)(R1)
58
59 // Save callback arguments to args array.
60 // Layout: floats first (F1-F8), then ints (R3-R10), then stack args
61 FMOVD F1, (ARGS_ARRAY+FLOAT_OFF+0*8)(R1)
62 FMOVD F2, (ARGS_ARRAY+FLOAT_OFF+1*8)(R1)
63 FMOVD F3, (ARGS_ARRAY+FLOAT_OFF+2*8)(R1)
64 FMOVD F4, (ARGS_ARRAY+FLOAT_OFF+3*8)(R1)
65 FMOVD F5, (ARGS_ARRAY+FLOAT_OFF+4*8)(R1)
66 FMOVD F6, (ARGS_ARRAY+FLOAT_OFF+5*8)(R1)
67 FMOVD F7, (ARGS_ARRAY+FLOAT_OFF+6*8)(R1)
68 FMOVD F8, (ARGS_ARRAY+FLOAT_OFF+7*8)(R1)
69
70 MOVD R3, (ARGS_ARRAY+INT_OFF+0*8)(R1)
71 MOVD R4, (ARGS_ARRAY+INT_OFF+1*8)(R1)
72 MOVD R5, (ARGS_ARRAY+INT_OFF+2*8)(R1)
73 MOVD R6, (ARGS_ARRAY+INT_OFF+3*8)(R1)
74 MOVD R7, (ARGS_ARRAY+INT_OFF+4*8)(R1)
75 MOVD R8, (ARGS_ARRAY+INT_OFF+5*8)(R1)
76 MOVD R9, (ARGS_ARRAY+INT_OFF+6*8)(R1)
77 MOVD R10, (ARGS_ARRAY+INT_OFF+7*8)(R1)
78
79 // Finish setting up callbackArgs struct at CB_ARGS(R1)
80 // struct { index uintptr; args unsafe.Pointer; result uintptr; stackArgs unsafe.Pointer }
81 // Note: index was already saved earlier (R11 is volatile)
82 ADD $ARGS_ARRAY, R1, R12
83 MOVD R12, (CB_ARGS+8)(R1) // args = address of register args
84 MOVD $0, (CB_ARGS+16)(R1) // result = 0
85
86 // stackArgs points to caller's stack arguments at old_R1+96 = R1+FRAME_SIZE+96
87 ADD $(FRAME_SIZE+96), R1, R12
88 MOVD R12, (CB_ARGS+24)(R1) // stackArgs = &caller_stack_args
89
90 // Call crosscall2 with arguments in registers:
91 // R3 = fn (from callbackWrap_call closure)
92 // R4 = frame (address of callbackArgs)
93 // R6 = ctxt (0)
94 MOVD ·callbackWrap_call(SB), R3
95 MOVD (R3), R3 // dereference closure to get fn
96 ADD $CB_ARGS, R1, R4 // frame = &callbackArgs
97 MOVD $0, R6 // ctxt = 0
98
99 BL crosscall2(SB)
100
101 // Get callback result into R3
102 MOVD (CB_ARGS+16)(R1), R3
103
104 // Restore R31
105 MOVD SAVE_R31(R1), R31
106
107 // Deallocate frame
108 ADD $FRAME_SIZE, R1
109
110 // Restore LR from caller's frame (per ELFv2, it was saved at 16(old_R1))
111 MOVD 16(R1), R0
112 MOVD R0, LR
113
114 RET