--disable-verify should be on by default in LDC

[brianush1]

https://llvm.org/docs/Passes.html#verify-module-verifier

Module verification turns segfaults into error messages that should never show up in the first place. I tested the effect --disable-verify has on a pretty large project of mine, here are the results:

ldc with --disable-verify
real    0m5.505s
user    0m4.788s
sys     0m0.705s

ldc without --disable-verify
real    0m18.184s
user    0m17.292s
sys     0m0.849s

dmd
real    0m4.098s
user    0m3.138s
sys     0m0.767s

[andrey-zherikov]

I'm not familiar with ldc so is there --enable-verify option? If so then assuming that proper toolchain should set all important compiler options for each build type (like debug or release), I'd say yes, --disable-verify should be default.

[brianush1]

there isn't one right now to my knowledge, but it probably could be added with close to zero effort

also apparently clang disables verification when compiling C files by default, so it's clearly good enough for them

[quickfur]

Sorry for my ignorance, what does --disable-verify do?

[brianush1]

LLVM has some verifications to make sure that it's not being fed malformed LLVM bitcode; --disable-verify disables those

since LDC should (presumably) not generate buggy bitcode, there's no point in having verification enabled by default

[quickfur]

Are those times you posted compile times, or runtimes? Adam was planning to standardize on ldc for release builds and dmd for dev builds; so I'd say compile times are not as important for ldc. Of course, having faster ldc compile times would be welcome, as long as it doesn't affect the correctness of the output.

[pinver]

For macOS Silicon actually you are forced to use LDC also for debug builds, as DMD is only available as x86_64; Rosetta is used, and the compile time are horrible compared to LDC compile time, witch is aarm64.

[quickfur]

Hmm if so, this is probably worth it.

[adamdruppe]

so how do we do it?

[kassane]

I'm currently using the latest version of ldc2, and I haven't noticed much difference using this flag at the moment.

I've also improved ldc2 support in my sokol-d project using zig build system parallelism. Attaching the caching build of both tools.
kassane/sokol-d@c557840

I remember reading an old post by one of the LDC developers about using -ir2obj-cache [renamed to -cache] (caching build).

Fixed link (edited)
https://johanengelen.github.io/ldc/2016/09/17/LDC-object-file-caching.html

Wouldn't using the -cache build flag progressively speed up compilation?