On 9/9/26 6:07 PM, Andrea Pinski wrote:
I noticed this while working on std::optional, the full blown cselim with data
references
hanldes the case where there are a few stores on either side of the branch that
don't
interfer with the other stores. We could just support clobbers but that would
have
a similar code except for an extra check for clobbers. So let's just add
support
for skipping on store on either side of the conditional too.
The change to trailing_store_in_bb simplifies some things as already the other
side would check on the store being a simple assign, this does not change
anything.
Note cselim-1.c was reduced from a draft version of this patch which was
miscompiling isl.
In the whole testsuite the only regressions were in graphite. So I thought it
would
be a good idea to have a testcase to be added for the miscompiling case. I had
forgot
some stores could also be loads so another check was needed in that case.
This was found by looking into std::optional::reset code gen.
Changes since v1:
* v2: skip stores and not just clobbers
Bootstrapped and tested on x86_64-linux-gnu.
PR tree-optimization/127053
gcc/ChangeLog:
* tree-ssa-phiopt.cc (trailing_store_in_bb): Rename vphi to
vuse_only.
(cond_if_else_store_replacement_limited): Add case if
cond_if_else_store_replacement_1 failed, allowing
for skipping one store on each side.
(trailing_store_in_bb): Require the store stmt
to be a gimple simple assign.
gcc/testsuite/ChangeLog:
* g++.dg/tree-ssa/cselim-2.C: New test.
* gcc.dg/torture/cselim-1.c: New test.
Signed-off-by: Andrea Pinski <[email protected]>
OK. Sorry about the long wait.
jeff
