[PATCH 1/5] cprop: Do not queue dead CC setters when bypassing jumps [PR127661]
Jeff Law
jeffrey.law@oss.qualcomm.com
Thu Oct 8 02:07:48 GMT 2026
On 10/6/26 3:33 AM, ktkachov@nvidia.com wrote:
> From: Kyrylo Tkachov <ktkachov@nvidia.com>
>
> The testcase in the PR is a long chain of
>
> if (v == 0) c->hit (0);
> if (v == 1) c->hit (1);
> ...
>
> At -O2 almost all of its compile time goes to jump bypassing in cprop1.
> The GIMPLE threaders move the edge after each call only a few tests
> forward. bypass_block then redirects each such edge to the next
> conditional block, which the forward sweep of bypass_conditional_jumps
> visits later, so a chain of N tests gives about N*N/2 redirections.
>
> Each redirection queues a copy of the flags setter of the bypassed
> block on the edge. The copies are dead, because the next block sets
> the flags again, but reg_killed_on_edge and insert_insn_on_edge walk the
> whole queue of the edge at each step, and commit_edge_insertions adds
> about N*N/2 dead compares. In gcc.target/aarch64/pr127661.c the block
> of the first call ends up as:
>
> (call_insn 10 ... (call (mem:DI (symbol_ref:DI ("f")) ...) ...))
> (insn 76 ... (set (reg:CC 66 cc)
> (compare:CC (reg/v:SI 101 [ v ]) (const_int 1))))
> (insn 78 ... (set (reg:CC 66 cc)
> (compare:CC (reg/v:SI 101 [ v ]) (const_int 2))))
> ... five more compares, with the constants 3 to 7 ...
>
> This patch copies the setter only if it is needed on entry to the new
> destination, that is if it has side effects, can throw, or sets a
> register that is live there or is not a condition-code register.
> Setters of other registers are always copied, because debug insns and
> REG_EQUAL notes, which the liveness information ignores, can refer to
> them. The comment of bypass_setcc_needed_p explains why the liveness
> information, which predates the global propagation and the
> redirections, is safe to use. With the patch the block above contains
> only the call. gcc.dg/pr127661.c and gcc.target/aarch64/pr127661-2.c
> check that the copy stays when the flags are live at the new
> destination.
>
> For the testcase at N=2048 on Grace the compile time drops by 95% and
> the peak memory use by 78%. For the variant with the calls inlined the
> compile time at N=1024 drops by 92%. Dead copies no longer stop the
> bypass of an outgoing edge, so the patch can add redirections, but it
> adds none for the testcase, and the assembly output does not change for
> N from 64 to 2048.
>
> Bootstrapped and tested as a series on aarch64-linux-gnu and
> x86_64-linux-gnu.
>
> Ok for trunk?
> Thanks,
> Kyrill
>
> Cc: Jeff Law <jeffrey.law@oss.qualcomm.com>
>
> gcc/ChangeLog:
>
> PR rtl-optimization/127661
> * cprop.cc (bypass_setcc_needed_p): New function.
> (bypass_block): Use it to decide whether to copy SETCC onto a
> redirected edge.
> (bypass_conditional_jumps): Update comment.
>
> gcc/testsuite/ChangeLog:
>
> PR rtl-optimization/127661
> * gcc.dg/pr127661.c: New test.
> * gcc.target/aarch64/pr127661.c: New test.
> * gcc.target/aarch64/pr127661-2.c: New test.
> * gcc.target/i386/pr127661.c: New test.
>
> Signed-off-by: Kyrylo Tkachov <ktkachov@nvidia.com>
OK
jeff
More information about the Gcc-patches
mailing list