[LLVMdev] [RFC] Removing static initializers for command line options
Renato Golin
renato.golin at linaro.org
Tue Aug 19 22:28:36 PDT 2014
On 20 August 2014 05:43, Sean Silva <chisophugis at gmail.com> wrote:
> 2. As a way to easily set up a conduit from A to B, where A is the command
> line and B is some place "deep" inside the LLVM library code that will do
> something in response to the command line.
I never liked this as a solution to #2. Heck, I never liked that we do
have #2 in the first place.
> Hence my suggestion of just putting a stringly typed key-value store (or
> whatever) in an easily accessible place (like LLVMContext), and just
> translating any unrecognized command line options (ones that are not for #1)
> into that stringly typed storage.
I fully support this idea, and is in line with my strawman proposal
for FPU/CPU parsing on all tools shared (beyond the boundaries of
LLVM).
String parsing is common and, even being target specific or pass
specific, is still string parsing and should be identical to all users
of the *same* feature.
I also believe a more sane command line option scheme is in order.
Today we have a zillion of options that are completely disconnected,
documented by accident in the initializer, without any bigger context
whatsoever. This is in part to follow what GCC has always done, and
probably we'll still need to support GCC's and our own legacy for
decades, but all that can also live in this commoned up parser.
Specifically to #2, my idea was something like:
--vectorizer-opts="foo,bar,baz=10" --tbaa-opts="...", etc. That could
use a common parser all the way down to parsing "foo", which would be
left by the vectorizer's back-end to the parser to deal with and setup
the right flags in the right structure, used because "vectorizer" in
"vectorizer-opts" tell the factory to return a vectorizer parser's
back-end.
The FPU/CPU parsing (which has to parse command line options and
assembly directives, which happens to have the same syntax), would
have a similar structure.
Such flags in LLVMContext (or whatever) would have to be structured
like a tree and each pass should receive its own tree root, which most
of the time would have just a list of key/values, but some times have
a more nested structure.
cheers,
--renato
More information about the llvm-dev
mailing list