short: introduce shorter template for C/C++

Enables 156KB/0ms with the shorter generation paradigm.

scripts/static-pie-gen.py

@@ -74,6 +74,10 @@
 code_raw_b91 = base91.encode(code_raw, use_rle=True).decode('ascii')
 code_raw_b91_len = len(code_raw_b91)
 code_raw_b91 = '"' + code_raw_b91 + '"'
+if lang_name == "C":
+    # Escape '\' and '?'
+    code_raw_b91 = code_raw_b91.replace('\\', '\\\\')
+    code_raw_b91 = code_raw_b91.replace('?', '\\?')
 
 # binary
 with open(compressed_binary_path, "rb") as f:
@@ -124,7 +128,7 @@
 
 # template
 template_candidates = [template_path]
-if lang_name == "Rust" and "x86_64" in target_name and "short" in template_path and len(code_raw) <= 4096 - 256:
+if lang_name in ["C", "Rust"] and "x86_64" in target_name and "short" in template_path and len(code_raw) <= 4096 - 256:
     template_candidates.append(template_path.replace("short", "shorter"))
 
 out = None

scripts/static-pie-prestub-amd64-print.py

@@ -30,6 +30,14 @@
     SPECIFIER = ".quad"
     CHUNK_SIZE = 8
     ENTRIES_PER_LINE = 4
+if "--c" in sys.argv:
+    PREFIX = "0x"
+    SUFFIX = ""
+    SPECIFIER = ""
+    ENTRIES_PER_LINE = 100
+else:
+    PREFIX = ""
+    SUFFIX = "h"
 
 # pad to align at `CHUNK_SIZE`-byte boundary
 while len(prestub) % CHUNK_SIZE != 0:
@@ -48,8 +56,8 @@ def to_hex_short(y):
         nonzero_idx = len(out)
         while nonzero_idx > 1 and out[nonzero_idx-1] == '0':
             nonzero_idx -= 1
-        out2 = out[:nonzero_idx] + "h<<" + str((len(out) - nonzero_idx) * 4)
-        out = out + "h"
+        out2 = PREFIX + out[:nonzero_idx] + SUFFIX + "<<" + str((len(out) - nonzero_idx) * 4)
+        out = PREFIX + out + SUFFIX
         if len(out2) < len(out):
             out = out2
         if ord(out[0]) >= ord('a'):

scripts/static-pie-prestub-amd64-shorter-c.asm

@@ -0,0 +1,98 @@
+; -*- tab-width: 4 -*-
+; 
+; The prestub for amd64-rust target
+; (prestub: the code that runs before the stub and sets the stage)
+;
+; build: nasm -f bin -O9 static-pie-prestub-amd64-shorter-c.asm -o static-pie-prestub-amd64-shorter-c.bin
+; note: after building with the above command, run static-pie-prestub-amd64-print.py static-pie-prestub-amd64-shorter-c.bin --c
+;       to obtain the form that can be embedded in Rust as inline assembly.
+
+BITS 64
+ORG 0
+section .text
+
+; Reserve space on stack
+    and     rsp, 0xffffffffffffff80 ; ensures at least 128 bytes
+
+; mprotect: make stack executable
+    mov     eax, 10                 ; mprotect
+    mov     esi, 0x1000             ; len
+    push    rdi                     ; Save binary_raw_base91
+    lea     rdi, [rsp + 8]          ; addr
+    push    7                       ; protect (RWX)
+    pop     rdx
+    and     rdi, 0xfffffffffffff000 ; align to page boundary (4K)
+    syscall
+
+; Relocate to stack
+    lea     rsi, [rel _start]
+    lea     rdi, [rsp + 8]
+    push    rdi                     ; _start of relocated stub
+    mov     ecx, _end - _start
+    rep     movsb
+
+; Jump to stack
+    pop     rax                     ; _start of relocated stub
+    call    rax
+
+_start:
+
+; Free the .text section
+    pop     rdi                     ; Get RIP saved on stack by call instruction
+    and     rdi, 0xfffffffffffff000
+    mov     esi, 0x1000
+    mov     eax, 11
+    syscall
+
+; svc_alloc_rwx for Linux
+_svc_alloc_rwx:
+    push    9
+    pop     rax                     ; syscall id of x64 mmap
+    cdq                             ; rdx=0
+    xor     r9d, r9d                ; offset
+    xor     edi, edi                ; rdi=0
+    mov     esi, eax                ; size (anything in [1, 4096])
+    mov     dl, 7                   ; protect (safe since we have ensured rdx=0)
+    push    0x22
+    pop     r10                     ; flags
+    push    -1
+    pop     r8                      ; fd
+    syscall
+    pop     rsi                     ; restore rsi
+
+; Current state: rax = new buffer
+    xchg    rax, rdi                ; rdi = new buffer
+
+; Base91 decoder
+_decode:
+    mov     al, 0x1f                ; syscall preserves all registers except rcx, r11, rax; hence at this point rax=(previous rdi)=0
+_decode_loop:
+    shl     eax, 13
+_decode_loop_2:
+    lodsb
+    xor     ecx, ecx                ; ecx = 0
+    sub     al, 0x23
+    jbe     _decode_zeros
+    dec     al
+    xchg    eax, ecx
+    lodsb
+    sub     al, 0x24
+    imul    eax, eax, 91
+    add     eax, ecx
+_decode_output:
+    stosb
+    shr     eax, 8
+    test    ah, 16
+    jnz     _decode_output
+    jmp     _decode_loop
+_decode_zeros:
+    xchg    byte [rdi-1], cl        ; ecx = cl = ((number of zeros) - 1), byte [rdi-1] = 0
+    rep     stosb                   ; we have made sure the last byte is zero (in the packager)
+    jz      _decode_loop_2
+
+; Jump to entrypoint
+_jump_to_entrypoint:
+    sub     rdi, qword [rdi-8]
+    call    rdi
+
+_end:

scripts/static-pie-template-amd64-shorter.c

@@ -0,0 +1,10 @@
+// Generated with https://github.com/kiwiyou/basm-rs
+// Learn rust (https://doc.rust-lang.org/book/) and get high performance out of the box!
+// SOLUTION BEGIN
+$$$$solution_src$$$$
+// SOLUTION END
+__attribute__((section(".text#")))unsigned long long s[]={0xab880e48348,0x485700001000be00,0x485a076a08247c8d,0x50ffffff000e781,0x4800000010358d48,0x5db95708247c8d,0x5fd0ff58a4f30000,0xbefffff000e78148,0xbb800001000,0x459958096a050f00,0x7b2c689ff31c931,0x5841ff6a5a41226a,0xc11fb097485e050f,0x76232cc931ac0de0,0x6b242cac91c8fe16,0x8e8c1aac8015bc0,0x86e0ebf77510c4f6,0x2b48dc74aaf3ff4f,0xd7fff87f};char t[]=$$$$binary_raw_base91$$$$;int main(){return 0;}
+#if defined(__cplusplus)
+extern "C"
+#endif
+int __libc_start_main(){return((int(*)(void*))s)(t);}