On Fri, Aug 1, 2025 at 9:01 PM Andrew Pinski <quic_apin...@quicinc.com> wrote: > > This is version 2 of the following patches: > https://inbox.sourceware.org/gcc-patches/20250608175102.1476184-1-quic_apin...@quicinc.com/ > https://inbox.sourceware.org/gcc-patches/20250609061323.1772598-1-quic_apin...@quicinc.com/T/#u > > For the memset patch, I had forgot to include the limit for the walker. > This has now been added back and some minor formating issues has been fixed. > The first one is unchanged; see below for stats. > > The other 2 patches here are a bug fix for non-call exceptions and an > improvement dealing with zeroing case; recognizing integer zero stores > like a memset of zero (also vector and complex zeros). > > As I mentioned the first patch makes it easier to prop not just into other > copy statements but into arguments and returns and even the possibility of > handling partial loads in the future. > > As far as stats are concerned, the current copy prop for aggregates and this > one, > does not change code generation in many cases. Though this is only because > SRA was > handling these cases. Now what I have seen is in heavily C++ code, the number > of decls > created have decreased significantly. For an example while compiling > highway's benchmark > code, the number of decls created decreased over 2000; this is because SRA > does not need > to do a full scalarization for many of the copies. > Also the places where this has improved code generation is SRA is not doing > full scalarization > which made things worse.
Just an FYI, I found a bug in the second patch of this patch set. Well the bug there is on the trunk; I don't know if it affects anything. The bug is in the creation of ao_ref for the memset (memcpy on the trunk). Since ao_ref_init is used on the ADDR_EXPR; the size could be wrong. Though I can't find a testcase for the new forward code, I did find one for backwards version, filed as https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121422 . Thanks, Andrew Pinski > > > > Andrew Pinski (4): > forwprop: Change optimize_agr_copyprop into forward walk instead of > backwards > forwprop: Change proping memset into memcpy into a forwprop rather > than a backwalk > forwprop: Don't do copy-prop-aggregates from statements that could > throw [PR120599] > forwprop: Recongize a store of integral zero for > optimize_aggr_zeroprop. > > gcc/testsuite/g++.dg/torture/noncall-eh-1.C | 26 ++ > gcc/testsuite/gcc.dg/pr118946-1.c | 15 + > .../gcc.dg/torture/copy-prop-aggr-zero-1.c | 28 ++ > .../gcc.dg/torture/copy-prop-aggr-zero-2.c | 28 ++ > .../tree-ssa/copy-prop-aggregate-zero-1.c | 28 ++ > .../tree-ssa/copy-prop-aggregate-zero-2.c | 25 ++ > .../tree-ssa/copy-prop-aggregate-zero-3.c | 25 ++ > gcc/tree-ssa-forwprop.cc | 425 ++++++++++-------- > 8 files changed, 413 insertions(+), 187 deletions(-) > create mode 100644 gcc/testsuite/g++.dg/torture/noncall-eh-1.C > create mode 100644 gcc/testsuite/gcc.dg/pr118946-1.c > create mode 100644 gcc/testsuite/gcc.dg/torture/copy-prop-aggr-zero-1.c > create mode 100644 gcc/testsuite/gcc.dg/torture/copy-prop-aggr-zero-2.c > create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/copy-prop-aggregate-zero-1.c > create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/copy-prop-aggregate-zero-2.c > create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/copy-prop-aggregate-zero-3.c > > -- > 2.43.0 >
