[Bug rtl-optimization/116039] [15 Regression] rv64gc miscompile at -O3 with -fno-strict-aliasing since r15-1901-g98914f9eba5

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116039

--- Comment #3 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jeff Law <law@gcc.gnu.org>:

https://gcc.gnu.org/g:34fb0feca71f763b2fbe832548749666d34a4a76

commit r15-2321-g34fb0feca71f763b2fbe832548749666d34a4a76
Author: Jeff Law <jlaw@ventanamicro.com>
Date:   Thu Jul 25 12:32:28 2024 -0600

    [PR rtl-optimization/116039] Fix life computation for promoted subregs

    So this turned out to be a neat little test and while the fuzzer found it
on
    RISC-V, I wouldn't be surprised if the underlying issue is also the root
cause
    of the loongarch issue with ext-dce.

    The key issue is that if we have something like

    (set (dest) (any_extend (subreg (source))))

    If the subreg object is marked with SUBREG_PROMOTED and the sign/unsigned
state
    matches the any_extend opcode, then combine (and I guess anything using
    simplify-rtx) may simplify that to

    (set (dest) (source))

    That implies that bits outside the mode of the subreg are actually live and
    valid.  This needs to be accounted for during liveness computation.

    We have to be careful here though. If we're too conservative about setting
    additional bits live, then we'll inhibit the desired optimization in the
    coremark examples.  To do a good job we need to know the extension opcode.

    I'm extremely unhappy with how the use handling works in ext-dce.  It mixes
    different conceptual steps and has horribly complex control flow.  It only
    handles a subset of the unary/binary opcodes, etc etc.  It's just damn
mess.
    It's going to need some more noodling around.

    In the mean time this is a bit hacky in that it depends on non-obvious
behavior
    to know it can get the extension opcode, but I don't want to leave the
trunk in
    a broken state while I figure out the refactoring problem.

    Bootstrapped and regression tested on x86 and tested on the crosses. 
Pushing to the trunk.

            PR rtl-optimization/116039
    gcc/
            * ext-dce.cc (ext_dce_process_uses): Add some comments about
concerns
            with current code.  Mark additional bit groups as live when we have
            an extension of a suitably promoted subreg.

    gcc/testsuite
            * gcc.dg/torture/pr116039.c: New test.