BoF: Register pressure sensitivity in the GCC middle-end

Chip Kerchner Q:

    This may be a silly question but for some of these issues has anyone considered using an AI for register pressure, etc.?


Notes:

    Historically GCC created larger statements for expand.
    What is a good-enough proxy for pressure in the middle-end? SSA names?
    GIMPLE to RTL translation could look at register use of register classes for statements.
    Expand should be simple lowering or at least remove optimization.
    Can register allocator create sub-regions?
    Experiment with disabling TR and replace expand optimization with earlier middle-end optimization to avoid unnecessarily large and unrecognized register usage.
    Piggy back on const IL walking prior to expand.
    LICM impact on register pressure is easier to predict
    Undo reassociation
    Inlining impact difficult to predict.
    Extend regions that global register allocator forms.
    Identify strongly connected components and partition.


----------------------------------------
RISC-V BoF
Jose E. Marchesi

    So what simulator should make it to GDB :)

Jamie Lokier

    After Glibc has used the vector unit in memcpy(), is there not a fast way to say "no longer care about preserving vector state from here" so that context switch doesn't have to save it?

Philipp Tomsich

    Can we first move from the guthub integration branches to vendor branches in the GCC git?

Joseph Myers

    With these extensions in master (which I think is a good idea), at some point you'll need to deal with obsolescing/removing support for extensions that's no longer maintained and causing maintenance problems (cf. the removal of e500 support from the GCC powerpc port, for example).

Pedro Alves

    it's useful to be able to install multiple versions of the toolchain one masking others. e.g., x86-64-akme-gcc and x86-64-msft-gcc both on the PATH. otherwise, should be cosmetic, IMHO.
        WITHOUT one masking...

Joseph Myers

    I'm not convinced it's the point of the vendor field in target triplets (that's more a relic of m68k-<some-unix-box-vendor>-sysv), but certainly plenty of targets have -m options for vendor extensions.

Kito Cheng
- RISC-V using -march as an unify interface to control which extensions are enabled, I think vendor ext. should just follow that could be easier.
Jose E. Marchesi

    where is embecosm' cpu description anyway? It is not in upstream cpu/

----------------------------------------

    Richard Biener

     The SUSE packages are built from the   gcc-11 branch, not OG11
     

    Vladimir Mezentsev

    Can gdb attach to the multi-thread application ?