[SAOC 2026] ImportC Weekly Update #7

[Niyati Jain]

**SAOC 2026: ImportC Weekly Update #7**

This week I mainly worked on the dead-code scanner from last 
update, the SEH catch-variable linkage, and getting curl, zlib 
and sqlite actually verified against real MSVC headers.

`#24138` (https://github.com/dlang/dmd/pull/24138), from last 
update's dead-code scanner for `#21266`, is now merged, which 
then real `windows.h` broke it: two `__forceinline` functions 
that only call each other, `RtlZeroVolatileMemory` calling 
`RtlFillVolatileMemory`, nothing else calls either one, both got 
kept alive, each vouching for the other, is fixed as `#24144` 
(https://github.com/dlang/dmd/pull/24144), and rewritten as a 
worklist, where a candidate's own body only gets scanned once 
something outside the candidate set has already proven it's used.

But that fix had a bug of its own, which took me a while to 
actually catch it. I'd tracked candidates in a 
`bool[FuncDeclaration]` keyed by class reference that was fine on 
every normal build, and failed only when dmd itself got built by 
an old bootstrap compiler or by LDC, same thing that I'd already 
faced earlier with `importCInlineReferenced`. Plain field on 
`FuncDeclaration` instead, no AA. Tested it against the real 
bootstrap/LDC matrix on a throwaway branch.

While I was on the SEH side of things, Nicholas asked whether one 
can tell a `catch(Throwable)` apart from one that was synthesized 
by ImportC lowering a C file's `__try`/`__except`, versus one 
written in D. Turns out it's `Catch.internalCatch` plus the 
enclosing module's filetype, and chasing that down is what led to 
`#24141` (https://github.com/dlang/dmd/pull/24141): a catch 
variable inside an `extern(C)` function was picking up C linkage 
from the enclosing scope instead of D linkage, which broke 
codegen for exactly this kind of synthesized catch.

`#24139` (https://github.com/dlang/dmd/pull/24139) also merged: 
ImportC's `static` C globals were getting D's `static` instead of 
C's actual internal linkage.

curl's real Windows TLS backend, `lib/vtls/schannel.c`, compiles 
clean through ImportC against real `windows.h`/SSPI now.

sqlite's amalgamation compiles, links, and runs a real 
create/insert/select through ImportC on both Windows and Linux. 
`SQLITE_OMIT_SEH` on Windows, since ImportC can't give 
`__try`/`__except` the exception context it needs, 
`-U__SIZEOF_INT128__` on Linux so glibc's headers stop pulling in 
a type dmd doesn't have, so I built the same amalgamation twice, 
once through ImportC and once through a real `cl.exe` build on 
sqlite's normal SEH path, and diffed the output: byte-for-byte 
identical `sqlite3_libversion()`, with same query result both 
times. zlib against real vcpkg `zlib.h` checks out the same way.

Next week: I'll work on regression tests for the intrinsics that 
only needed declaring, and then start the macro survey for 
Milestone 3.

Feedback and questions are welcome.
Thanks to Nicholas for the guidance throughout.

Best,
Niyati