test: add msvc function-boundary fixtures for #384

[danielplohmann]

Adds tests/msvc_smda384_corpus/: minimal MSVC-built DLLs that reproduce the function-boundary problems in #384, as data for whoever picks up the fix. No code changes and no tests yet. The fixtures are added only, and a regression test belongs with the fix.

What is in it (1.4 MB)

• Three DLLs (x86, x64, arm64) at /O2, 11-14 KB each, with their PDBs. They are xored with byte ^ (index % 256) like the existing fixtures.
• manifest.json per fixture:
  • raw and xored sha256 of DLL and PDB, base address, toolchain and exact flags;
  • expected_functions: the 11 functions of cases.obj from the linker map, with address, name and static / public;
  • known_failures_smda_4_9_0: current results with and without the PDB.
• src/: the C sources. They were written for this purpose; no third-party or malware code.
• README.md: the three behaviours, which cases trigger them, and the build.

Current SMDA 4.9.0 results against the map

build	no PDB	with PDB	cause
x86 /O2	9/11	9/11	a block after ret split off (case_b2_whitespace); a jump table in .text decoded as code swallows case_b4_aggregate
x64 /O2	10/11	10/11	a block after ret split off (case_b2_whitespace)
arm64 /O2	9/11	11/11	shrink-wrapped prologue taken as a function start (case_a2_unistr, case_a3_concat)

Build

• VS 2026 Build Tools (cl 19.51.36260, link 14.51.36260), on a GitHub windows-2025 runner in a neutral directory.
• /O2 /MD /Zi /Brepro and /DLL /DEBUG:FULL /OPT:REF /OPT:NOICF /PDBALTPATH:%_PDB% /MAP.
• The output contains no machine- or user-specific paths, apart from Microsoft's own CRT object paths in the PDBs.

Refs #384.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HDZJnaEz52GwyRZzXP8Mya

[r0ny123]

Thanks, these reproduce exactly the table in the description on master.

One correction for the README: case_b2_whitespace's extra start (+0xd5 on x86, +0x10d on x64) is not a block after ret. The switch is two-level: movzx ecx, byte ptr [ecx + idx]; jmp dword ptr [ecx*4 + table], and SMDA only books the dword table as data. The 218-byte index table behind it (04 04 ... = add al, 4) is then picked up by the gap scan as a function.

With the fix for #384 on top, all six builds are 11/11 with and without the PDB. I'll add a regression test over these fixtures there once this is merged.