After tail-merge merges duplicate blocks, the predecessors that branched to
them now branch to a single successor. This series re-runs ifcombine on those
predecessors so that sequential comparisons feeding a common successor are
combined into one condition; on targets with conditional compares the backend
then emits ccmp, and elsewhere it helps wherever ifcombine's cost model already
prefers the combined form.
1/3 exports the ifcombine helpers via a new header.
2/3 relaxes the uninit predicate analysis so a combined-condition shape does
not produce bogus -Wmaybe-uninitialized warnings.
3/3 collects the candidate predecessors after tail-merge and runs ifcombine
on them to combine the conditions.
Bootstrapped and regression-tested on aarch64-linux-gnu and
x86_64-linux-gnu.
Changes in v3:
- Drop the targetm.have_ccmp () gate; the post-tail-merge ifcombine now runs on
all targets, leaving profitability to ifcombine's own cost model.
- Replace the maybe_undef_at_dom_frontier_p work-around with a fix in the uninit
predicate analysis (relax the definition predicate by the maybe-undef
operands' incoming-edge conditions).
- apply_clusters: hoist the immediate-dominator check ahead of
recognize_if_then_else and assert the gcond.
- Add gcc.dg/uninit-pred-13.c covering the uninit interaction.
- Make the pr102793 tests target-independent (force the combine with
--param logical-op-non-short-circuit=1, drop the ccmp dg-skip-if) and use
__UINT64_TYPE__.
- ifcombine: do not combine the conditions guarding a loop. Extend the
existing niter-safety check (which already leaves integer/pointer loop exit
conditions alone) to loop guards too, so the post-tail-merge combine no
longer makes a guarded loop's number-of-iterations unanalyzable; this
restores the IV elimination in gcc.dg/tree-ssa/ivopt_mult_1.c and
ivopt_mult_2.c.
- pr117123.C: force --param logical-op-non-short-circuit=0 so this
value-numbering test keeps its short-circuit form (3 ifs) on all targets,
rather than being combined to 2 where the target defaults to
non-short-circuit (e.g. x86).
Konstantinos Eleftheriou (3):
ifcombine: Add tree-ssa-ifcombine.h and update function signatures
uninit: Relax PHI def predicate by maybe-undef edge conditions
tail-merge: Combine conditions of merged blocks for ccmp [PR102793]
gcc/gimple-predicate-analysis.cc | 96 ++++++++++++++++++++--
gcc/gimple-predicate-analysis.h | 2 +
gcc/testsuite/g++.dg/tree-ssa/pr117123.C | 2 +-
gcc/testsuite/gcc.dg/tree-ssa/pr102793-1.c | 49 +++++++++++
gcc/testsuite/gcc.dg/tree-ssa/pr102793-2.c | 50 +++++++++++
gcc/testsuite/gcc.dg/uninit-pred-13.c | 60 ++++++++++++++
gcc/tree-ssa-ifcombine.cc | 41 +++++++--
gcc/tree-ssa-ifcombine.h | 27 ++++++
gcc/tree-ssa-tail-merge.cc | 60 +++++++++++++-
9 files changed, 370 insertions(+), 17 deletions(-)
create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr102793-1.c
create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr102793-2.c
create mode 100644 gcc/testsuite/gcc.dg/uninit-pred-13.c
create mode 100644 gcc/tree-ssa-ifcombine.h
--
2.52.0
