nx_pe_dll_d3d11.nx source
↩ module page · 229 lines · 13847 B
1// nx_pe_dll_d3d11.nx -- the REAL game entry point. Superset of nx_pe_dll_id3d11: same data-driven 43-slot ID3D11Device
2// vtable (compiled IUnknown+GetFeatureLevel bodies) PLUS the actual export D3D11CreateDeviceAndSwapChain (the literal
3// first function a game calls). That export is MS-x64 glue that reads the out-params from the STACK (args 8-11 at
4// [rsp+0x48..0x60] -- proving stack-arg access) and returns a STATIC device (real D3D allocates internally; a singleton
5// static device in a new RW .data section is the honest minimal form). Verified against Microsoft's own
6// PFN_D3D11_CREATE_DEVICE_AND_SWAP_CHAIN typedef so the arg->stack mapping comes from <d3d11.h>, not us.
7// Method bodies = sovereign nx_cc NishiLang; vtable+glue = sovereign emitter (NO C++). ⚠static vtable/device baked @ fixed
8// base 0x180000000 (.reloc TODO). Input /tmp/nishi_export_lib.s Output knowledge/nishi_d3d11.dll license_tier: ORIGINAL expect_exit: 0
9import "nx_syscalls.nx"
10import "nx_itoa_lib.nx" // shared MSB-first emitter (zero-alloc)
11import "nxasm_x86.nx"
12import "nx_pe_writer.nx"
13
14const N_SLOTS: i64 = 43
15
16func d11_puts(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 }
17// MIGRATED to the shared emitter (debt 1785563586). The old body mmapped a scratch buffer
18// per call and never freed it. At PAGE granularity that is 4096B leaked PER CALL -- the
19// defect that took 28.5GB of a 36GB host in nx_ts_lumadiff (2MB input, ~3.66M calls).
20// nxi_* is MSB-first, allocates NOTHING, and emits identical bytes including the sign.
21func d11_putn(v: i64) -> i64 { nxi_out(v); return 0 }
22func d11_wstr(buf: *u8, off: i64, s: *u8) -> i64 { var i: i64=0; while s[i]!=(0 as u8){ buf[off+i]=s[i]; i=i+1 } buf[off+i]=0 as u8; return i+1 }
23
24func d11_assemble(src: *u8, n: i64, out: *u8, out_cap: i64, names: *i64, offs: *i64, cnt: i64) -> i64 {
25 let lab_off: *i64 = sys_mmap(ASM_MAX_LABELS * 8) as *i64
26 let lab_len: *i64 = sys_mmap(ASM_MAX_LABELS * 8) as *i64
27 let lab_addr: *i64 = sys_mmap(ASM_MAX_LABELS * 8) as *i64
28 let lab_sec: *i64 = sys_mmap(ASM_MAX_LABELS * 8) as *i64
29 let op0: *i64 = sys_mmap(72) as *i64 // 9 slots for SIB (matches nxasm_x86)
30 let op1: *i64 = sys_mmap(72) as *i64
31 let op2: *i64 = sys_mmap(72) as *i64 // API DRIFT FIX: axc_pass gained op2
32 let scratch: *u8 = sys_mmap(64)
33 let posbox: *i64 = sys_mmap(16) as *i64
34 let n_lab_box: *i64 = sys_mmap(16) as *i64
35 n_lab_box[0] = 0
36 let lh: *i64 = sys_mmap(ASM_LH_SIZE * 8) as *i64
37 let text_size: i64 = axc_pass(src, n, out, 0, 0, lab_off, lab_len, lab_addr, lab_sec, n_lab_box, lh, op0, op1, op2, scratch, posbox)
38 if text_size < 0 { return text_size }
39 let n_lab: i64 = n_lab_box[0]
40 var k: i64 = 0
41 while k < n_lab {
42 if lab_sec[k] == 1 { lab_addr[k] = lab_addr[k] + text_size }
43 var m: i64 = 0
44 while m < cnt { if axc_tok_is(src, lab_off[k], lab_len[k], names[m] as *u8) == 1 { offs[m] = lab_addr[k] } m = m + 1 }
45 k = k + 1
46 }
47 axc_lh_build(src, lab_off, lab_len, n_lab, lh)
48 let total: i64 = axc_pass(src, n, out, text_size, 1, lab_off, lab_len, lab_addr, lab_sec, n_lab_box, lh, op0, op1, op2, scratch, posbox)
49 if total < 0 { return total }
50 if total > out_cap { return 0 - 200 }
51 return total
52}
53
54func d11_thunk(buf: *u8, tco: i64, moff: i64) -> i64 {
55 let t: i64 = 0x200 + tco
56 _w8(buf,t+0,0x57); _w8(buf,t+1,0x56)
57 _w8(buf,t+2,0x48); _w8(buf,t+3,0x89); _w8(buf,t+4,0xCF)
58 _w8(buf,t+5,0x48); _w8(buf,t+6,0x89); _w8(buf,t+7,0xD6)
59 _w8(buf,t+8,0x4C); _w8(buf,t+9,0x89); _w8(buf,t+10,0xC2)
60 _w8(buf,t+11,0x4C); _w8(buf,t+12,0x89); _w8(buf,t+13,0xC9)
61 _w8(buf,t+14,0x48); _w8(buf,t+15,0x83); _w8(buf,t+16,0xEC); _w8(buf,t+17,0x08)
62 _w8(buf,t+18,0xE8); _w32(buf, t+19, moff - (tco + 23))
63 _w8(buf,t+23,0x48); _w8(buf,t+24,0x83); _w8(buf,t+25,0xC4); _w8(buf,t+26,0x08)
64 _w8(buf,t+27,0x5E); _w8(buf,t+28,0x5F); _w8(buf,t+29,0xC3)
65 return 0
66}
67
68func d11_slot_method(slot: i64) -> i64 {
69 if slot == 0 { return 0 }
70 if slot == 1 { return 1 }
71 if slot == 2 { return 2 }
72 if slot == 37 { return 3 }
73 return 4
74}
75
76func d11_emit(buf: *u8, code: *u8, code_len: i64, offs: *i64) -> i64 {
77 if (buf as i64) == 0 { return 0 - NX_PE_BAD_INPUT }
78 let FF: i64 = code_len + 150 // factory create_device (22B)
79 let CG: i64 = FF + 22 // D3D11CreateDeviceAndSwapChain glue (73B)
80 let VT: i64 = ((CG + 73 + 7) / 8) * 8 // 43-slot vtable (8-aligned)
81 let text_used: i64 = VT + N_SLOTS * 8
82 let text_pages: i64 = (text_used + 4095) / 4096
83 let text_raw: i64 = text_pages * 4096
84 let data_rva: i64 = 0x1000 + text_raw // .data (static device) -- contiguous after .text
85 let data_file: i64 = 0x200 + text_raw
86 let edata_rva: i64 = data_rva + 0x1000
87 let edata_file: i64 = data_file + 0x200
88 let size_of_image: i64 = edata_rva + 0x1000
89 let file_size: i64 = edata_file + 0x200
90 let vt_rva: i64 = 0x1000 + VT
91
92 // headers (3 sections)
93 _w16(buf, 0, 0x5A4D); _w32(buf, 0x3C, 0x80); _w32(buf, 0x80, 0x00004550)
94 _w16(buf, 0x84 + 0, PE_MACHINE_AMD64); _w16(buf, 0x84 + 2, 3); _w16(buf, 0x84 + 16, 0xF0); _w16(buf, 0x84 + 18, 0x2022)
95 let O: i64 = 0x98
96 _w16(buf, O + 0, PE_OH_MAGIC_PEPLUS); _w8(buf, O + 2, 1)
97 _w32(buf, O + 4, text_raw); _w32(buf, O + 8, 0x400)
98 _w32(buf, O + 16, 0); _w32(buf, O + 20, 0x1000)
99 _w64(buf, O + 24, 0x80000000, 0x1)
100 _w32(buf, O + 32, 0x1000); _w32(buf, O + 36, 0x200)
101 _w16(buf, O + 40, 6); _w16(buf, O + 48, 6)
102 _w32(buf, O + 56, size_of_image); _w32(buf, O + 60, 0x200)
103 _w16(buf, O + 68, PE_SUBSYSTEM_CONSOLE)
104 _w64(buf, O + 72, 0x100000, 0); _w64(buf, O + 80, 0x1000, 0); _w64(buf, O + 88, 0x100000, 0); _w64(buf, O + 96, 0x1000, 0)
105 _w32(buf, O + 108, 16)
106 _w32(buf, O + 112 + 0, edata_rva); _w32(buf, O + 112 + 4, 0x7E) // DataDir[0] Export
107
108 _emit_section_header(buf, FOFF_SECT_TBL, 46,116,101,120,116,0,0,0, text_used, 0x1000, text_raw, 0x200, PE_SECT_CODE_X_R) // .text
109 _emit_section_header(buf, FOFF_SECT_TBL + 40, 46,100,97,116,97,0,0,0, 16, data_rva, 0x200, data_file, PE_SECT_DATA_RW) // .data
110 _emit_section_header(buf, FOFF_SECT_TBL + 80, 46,101,100,97,116,97,0,0, 0x7E, edata_rva, 0x200, edata_file, PE_SECT_DATA_R) // .edata
111
112 var i: i64 = 0
113 while i < code_len { buf[0x200 + i] = code[i]; i = i + 1 }
114
115 // 5 uniform thunks
116 var j: i64 = 0
117 while j < 5 { d11_thunk(buf, code_len + j*30, offs[j]); j = j + 1 }
118
119 // factory create_device @ FF (builds device in caller mem; 22B): mov rax,rcx; lea rdx,[rip+dV]; mov[rcx],rdx; mov qword[rcx+8],1; ret
120 let f: i64 = 0x200 + FF
121 _w8(buf,f+0,0x48); _w8(buf,f+1,0x89); _w8(buf,f+2,0xC8)
122 _w8(buf,f+3,0x48); _w8(buf,f+4,0x8D); _w8(buf,f+5,0x15); _w32(buf, f+6, VT - (FF + 10))
123 _w8(buf,f+10,0x48); _w8(buf,f+11,0x89); _w8(buf,f+12,0x11)
124 _w8(buf,f+13,0x48); _w8(buf,f+14,0xC7); _w8(buf,f+15,0x41); _w8(buf,f+16,0x08); _w32(buf, f+17, 1)
125 _w8(buf,f+21,0xC3)
126
127 // D3D11CreateDeviceAndSwapChain glue @ CG (MS-ABI, reads out-params from stack; 73B). ppSwapChain=[rsp+0x48]
128 // ppDevice=[rsp+0x50] pFeatureLevel=[rsp+0x58] ppImmediateContext=[rsp+0x60]. static device @ .data (data_rva).
129 let g: i64 = 0x200 + CG
130 // *ppDevice = &static_device
131 _w8(buf,g+0,0x48); _w8(buf,g+1,0x8B); _w8(buf,g+2,0x44); _w8(buf,g+3,0x24); _w8(buf,g+4,0x50) // mov rax,[rsp+0x50]
132 _w8(buf,g+5,0x48); _w8(buf,g+6,0x85); _w8(buf,g+7,0xC0) // test rax,rax
133 _w8(buf,g+8,0x74); _w8(buf,g+9,0x0A) // je +10 (after_dev)
134 _w8(buf,g+10,0x48); _w8(buf,g+11,0x8D); _w8(buf,g+12,0x15); _w32(buf, g+13, data_rva - (vt_rva - VT + CG + 17)) // lea rdx,[rip+dD]
135 _w8(buf,g+17,0x48); _w8(buf,g+18,0x89); _w8(buf,g+19,0x10) // mov [rax],rdx
136 // *pFeatureLevel = 0xb000
137 _w8(buf,g+20,0x48); _w8(buf,g+21,0x8B); _w8(buf,g+22,0x44); _w8(buf,g+23,0x24); _w8(buf,g+24,0x58) // mov rax,[rsp+0x58]
138 _w8(buf,g+25,0x48); _w8(buf,g+26,0x85); _w8(buf,g+27,0xC0) // test rax,rax
139 _w8(buf,g+28,0x74); _w8(buf,g+29,0x06) // je +6 (after_fl)
140 _w8(buf,g+30,0xC7); _w8(buf,g+31,0x00); _w32(buf, g+32, 0xB000) // mov dword[rax],0xb000
141 // *ppSwapChain = 0
142 _w8(buf,g+36,0x48); _w8(buf,g+37,0x8B); _w8(buf,g+38,0x44); _w8(buf,g+39,0x24); _w8(buf,g+40,0x48) // mov rax,[rsp+0x48]
143 _w8(buf,g+41,0x48); _w8(buf,g+42,0x85); _w8(buf,g+43,0xC0) // test rax,rax
144 _w8(buf,g+44,0x74); _w8(buf,g+45,0x07) // je +7 (after_sc)
145 _w8(buf,g+46,0x48); _w8(buf,g+47,0xC7); _w8(buf,g+48,0x00); _w32(buf, g+49, 0) // mov qword[rax],0
146 // *ppImmediateContext = 0
147 _w8(buf,g+53,0x48); _w8(buf,g+54,0x8B); _w8(buf,g+55,0x44); _w8(buf,g+56,0x24); _w8(buf,g+57,0x60) // mov rax,[rsp+0x60]
148 _w8(buf,g+58,0x48); _w8(buf,g+59,0x85); _w8(buf,g+60,0xC0) // test rax,rax
149 _w8(buf,g+61,0x74); _w8(buf,g+62,0x07) // je +7 (after_ctx)
150 _w8(buf,g+63,0x48); _w8(buf,g+64,0xC7); _w8(buf,g+65,0x00); _w32(buf, g+66, 0) // mov qword[rax],0
151 _w8(buf,g+70,0x31); _w8(buf,g+71,0xC0) // xor eax,eax (S_OK)
152 _w8(buf,g+72,0xC3) // ret
153
154 // static vtable @ VT (43 baked-absolute thunk pointers)
155 var s: i64 = 0
156 while s < N_SLOTS {
157 let mj: i64 = d11_slot_method(s)
158 _w64(buf, 0x200 + VT + s*8, 0x80000000 + 0x1000 + code_len + mj*30, 1)
159 s = s + 1
160 }
161
162 // static device @ .data (data_file): device[0]=&vtable (abs), device[1]=refcount 1
163 _w64(buf, data_file + 0, 0x80000000 + vt_rva, 1)
164 _w64(buf, data_file + 8, 1, 0)
165
166 // .edata: 2 exports SORTED ("D3D11CreateDeviceAndSwapChain" < "nishi_d3d_create_device")
167 let E: i64 = edata_file
168 _w32(buf, E + 12, edata_rva + 0x72) // Name -> dllname
169 _w32(buf, E + 16, 1); _w32(buf, E + 20, 2); _w32(buf, E + 24, 2)
170 _w32(buf, E + 28, edata_rva + 0x28); _w32(buf, E + 32, edata_rva + 0x30); _w32(buf, E + 36, edata_rva + 0x38)
171 _w32(buf, E + 0x28, 0x1000 + FF) // func[0] = create_device factory
172 _w32(buf, E + 0x2C, 0x1000 + CG) // func[1] = D3D11CreateDeviceAndSwapChain glue
173 _w32(buf, E + 0x30, edata_rva + 0x3C) // name[0] = "D3D11Create..."
174 _w32(buf, E + 0x34, edata_rva + 0x5A) // name[1] = "nishi_d3d_create_device"
175 _w16(buf, E + 0x38, 1) // "D3D11Create..." -> func[1]
176 _w16(buf, E + 0x3A, 0) // "nishi_d3d..." -> func[0]
177 d11_wstr(buf, E + 0x3C, "D3D11CreateDeviceAndSwapChain\x00" as *u8)
178 d11_wstr(buf, E + 0x5A, "nishi_d3d_create_device\x00" as *u8)
179 d11_wstr(buf, E + 0x72, "nishicc.dll\x00" as *u8)
180 return file_size
181}
182
183func main() -> i64 {
184 d11_puts("=== nx_pe_dll_d3d11 -- real D3D11CreateDeviceAndSwapChain export + conformant ID3D11Device ===\n" as *u8)
185 let lenbox: *i64 = sys_mmap(16) as *i64
186 let src: *u8 = sys_read_file("/tmp/nishi_export_lib.s" as *u8, lenbox)
187 if (src as i64) == 0 { d11_puts("FAIL no /tmp/nishi_export_lib.s\n" as *u8); sys_exit(1); return 1 }
188 let n: i64 = lenbox[0]
189 if n <= 0 { d11_puts("FAIL empty .s\n" as *u8); sys_exit(2); return 2 }
190
191 let names: *i64 = sys_mmap(8*5) as *i64
192 names[0] = "nx_dev_qi" as *u8 as i64
193 names[1] = "nx_dev_addref" as *u8 as i64
194 names[2] = "nx_dev_release" as *u8 as i64
195 names[3] = "nx_dev_getfl" as *u8 as i64
196 names[4] = "nx_dev_notimpl" as *u8 as i64
197 let offs: *i64 = sys_mmap(8*5) as *i64
198 var z: i64 = 0
199 while z < 5 { offs[z] = 0 - 1; z = z + 1 }
200
201 let code: *u8 = sys_mmap(1048576)
202 let code_len: i64 = d11_assemble(src, n, code, 1048576, names, offs, 5)
203 if code_len < 0 { d11_puts("FAIL assemble rc="); d11_putn(code_len); d11_puts("\n" as *u8); sys_exit(3); return 3 }
204 var miss: i64 = 0
205 var q: i64 = 0
206 while q < 5 { if offs[q] < 0 { miss = miss + 1 } q = q + 1 }
207 if miss > 0 { d11_puts("FAIL method(s) not found\n" as *u8); sys_exit(4); return 4 }
208 d11_puts(" code_len="); d11_putn(code_len); d11_puts(" factory@"); d11_putn(code_len+150); d11_puts(" create-glue@"); d11_putn(code_len+172); d11_puts("\n" as *u8)
209
210 let buf: *u8 = sys_mmap(0x20000)
211 let fsz: i64 = d11_emit(buf, code, code_len, offs)
212 if fsz < 0 { d11_puts("FAIL emit rc="); d11_putn(fsz); d11_puts("\n" as *u8); sys_exit(5); return 5 }
213
214 let fd: i64 = sys_openat_wr("knowledge/nishi_d3d11.dll\x00" as *u8, 0x1a4)
215 if fd < 0 { d11_puts("FAIL open output\n" as *u8); sys_exit(6); return 6 }
216 let wrote: i64 = sys_write(fd, buf, fsz)
217 sys_close(fd)
218 d11_puts(" wrote "); d11_putn(wrote); d11_puts(" bytes -> knowledge/nishi_d3d11.dll (D3D11CreateDeviceAndSwapChain + ID3D11Device)\n" as *u8)
219
220 var fails: i64 = 0
221 if (buf[0] as i64) != 0x4D { fails=fails+1 }
222 if (buf[0x84+18] as i64) != 0x22 { fails=fails+1 }
223 if (buf[0x84+2] as i64) != 3 { fails=fails+1 }
224 if wrote != fsz { fails=fails+1 }
225 if fails == 0 { d11_puts("NX-PE-D3D11 GREEN -- real create export + conformant device; harness must call the real signature\n" as *u8); sys_exit(0); return 0 }
226 d11_puts("NX-PE-D3D11 RED fails="); d11_putn(fails); d11_puts("\n" as *u8)
227 sys_exit(1)
228 return 1
229}