https://gcc.gnu.org/g:09e8d26982bf4280ec41b2e5215a57900517db7c
commit r17-5801-g09e8d26982bf4280ec41b2e5215a57900517db7c Author: Steve Bezalel <[email protected]> Date: Fri Oct 30 02:47:27 2026 +0000 tree-optimization/127211 - Reset SCEV after predcom without unrolling tree_predictive_commoning only resets SCEV when a loop was unrolled or needed loop-closed SSA fixup (changed > 1). When a loop is transformed without either, the SCEV results cached by predcom's data-reference analysis survive the transform. At -O2, where predcom only runs implicitly with -ftree-loop-vectorize and never unrolls, that is the common case. In the testcase, the evolution of s is cached as not analyzable while its step is still a load of a[4]. Predcom then hoists that load, after which s is the affine IV {a_I_lsm0.4_26, +, a_I_lsm0.4_26}_1, but ivopts still reads the cached result, finds no niter and does not strength-reduce the b[s] store. At plain -O2 this is hidden because passes between predcom and ivopts clear the SCEV cache as a side effect: fre4 in pass_pre_slp_scalar_cleanup, and slp1 through the SSA name freelist flush. -fno-tree-slp-vectorize disables both. The fix resets SCEV whenever predcom changed something. Bootstrapped and regtested on x86_64-pc-linux-gnu. OK for trunk? PR tree-optimization/127211 gcc/ChangeLog: * tree-predcom.cc (tree_predictive_commoning): Reset SCEV whenever a loop was transformed, not only when unrolled. gcc/testsuite/ChangeLog: * gcc.dg/tree-ssa/pr127211.c: New test. Signed-off-by: Steve Bezalel <[email protected]> Diff: --- gcc/testsuite/gcc.dg/tree-ssa/pr127211.c | 24 ++++++++++++++++++++++++ gcc/tree-predcom.cc | 7 +++++-- 2 files changed, 29 insertions(+), 2 deletions(-) diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr127211.c b/gcc/testsuite/gcc.dg/tree-ssa/pr127211.c new file mode 100644 index 000000000000..bf9ae22bf219 --- /dev/null +++ b/gcc/testsuite/gcc.dg/tree-ssa/pr127211.c @@ -0,0 +1,24 @@ +/* PR tree-optimization/127211 */ +/* { dg-do compile } */ +/* { dg-options "-O2 -fno-tree-slp-vectorize -fdump-tree-ivopts-details" } */ + +int a[4096], b[4096]; + +void +fn (int m) +{ + int s = 0; + for (int i = 0; i < a[4]; i++) + { + a[3*i + 2] = m; + if (m & 1) + a[3*i + 5] = m - 1; + b[s] = m; + s += a[4]; + } +} + +/* Predictive commoning hoists the load of a[4] out of the loop, after + which s is an affine IV. SCEV results cached by its data-reference + analysis before the transform must not reach ivopts. */ +/* { dg-final { scan-tree-dump-not "niter scev_not_known" "ivopts" } } */ diff --git a/gcc/tree-predcom.cc b/gcc/tree-predcom.cc index 8e2894d46a09..a2043ef3122a 100644 --- a/gcc/tree-predcom.cc +++ b/gcc/tree-predcom.cc @@ -3512,11 +3512,14 @@ tree_predictive_commoning (bool allow_unroll_p) { ret = TODO_update_ssa_only_virtuals; + /* Cached SCEV information is stale after any transformation, not + only after unrolling: the data-reference analysis above queried + SCEV before references were rewritten. */ + scev_reset (); + /* Some loop(s) got unrolled. */ if (changed > 1) { - scev_reset (); - /* Need to fix up loop closed SSA. */ if (changed >= 4) rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
