On Wed, 18 Sep 2024, Richard Biener wrote: > The following adds --param vect-force-slp to enable the transition > to full SLP. Full SLP is enforced during stmt analysis where it > detects failed SLP discovery and at loop analysis time where it > avoids analyzing a loop with SLP disabled. Failure to SLP results > in vectorization to fail. > > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress; I'll > push after that finishes. > > This gives you the ability to dial ahead in time and evaluate > your target as to what happens when we pull the trigger.
Note this isn't a good place to put in ICEs looking for regressions since you'd have to make sure that we could have vectorized without SLP. There isn't an easy way to put in (guarded) asserts for that also given we'd really want to compare to GCC 14 with hybrid SLP and all such. So for now we have to simply rely on the testsuite to uncover regressions. Note there are still known gaps in SLP, notably early break vectorization and some OMP SIMD features. Richard. > Richard. > > * params.def (vect-force-slp): New param, default 0. > * doc/invoke.texi (--param vect-force-slp): Document. > * tree-vect-loop.cc (vect_analyze_loop_2): When analyzing > without SLP but --param vect-force-slp is 1 fail. > * tree-vect-stmts.cc (vect_analyze_stmt): Fail vectorization > for non-SLP stmts when --param vect-force-slp is 1. > --- > gcc/doc/invoke.texi | 3 +++ > gcc/params.opt | 4 ++++ > gcc/tree-vect-loop.cc | 6 ++++++ > gcc/tree-vect-stmts.cc | 6 ++++++ > 4 files changed, 19 insertions(+) > > diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi > index 511b8c8d311..b91fb9c9cca 100644 > --- a/gcc/doc/invoke.texi > +++ b/gcc/doc/invoke.texi > @@ -16955,6 +16955,9 @@ this parameter. The default value of this parameter > is 50. > @item vect-induction-float > Enable loop vectorization of floating point inductions. > > +@item vect-force-slp > +Force the use of SLP when vectorizing, fail if not possible. > + > @item vrp-block-limit > Maximum number of basic blocks before VRP switches to a lower memory > algorithm. > > diff --git a/gcc/params.opt b/gcc/params.opt > index c17ba17b91b..949b4754498 100644 > --- a/gcc/params.opt > +++ b/gcc/params.opt > @@ -1198,6 +1198,10 @@ The maximum factor which the loop vectorizer applies > to the cost of statements i > Common Joined UInteger Var(param_vect_induction_float) Init(1) > IntegerRange(0, 1) Param Optimization > Enable loop vectorization of floating point inductions. > > +-param=vect-force-slp= > +Common Joined UInteger Var(param_vect_force_slp) Init(0) IntegerRange(0, 1) > Param Optimization > +Force the use of SLP when vectorizing, fail if not possible. > + > -param=vrp-block-limit= > Common Joined UInteger Var(param_vrp_block_limit) Init(150000) Optimization > Param > Maximum number of basic blocks before VRP switches to a fast model with less > memory requirements. > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc > index 62c7f90779f..d42694d1974 100644 > --- a/gcc/tree-vect-loop.cc > +++ b/gcc/tree-vect-loop.cc > @@ -2891,6 +2891,12 @@ vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool > &fatal, > /* This is the point where we can re-start analysis with SLP forced off. > */ > start_over: > > + /* When we arrive here with SLP disabled and we are supposed > + to use SLP for everything fail vectorization. */ > + if (!slp && param_vect_force_slp) > + return opt_result::failure_at (vect_location, > + "may need non-SLP handling\n"); > + > /* Apply the suggested unrolling factor, this was determined by the backend > during finish_cost the first time we ran the analyzis for this > vector mode. */ > diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc > index b1353c91fce..495f45e40e6 100644 > --- a/gcc/tree-vect-stmts.cc > +++ b/gcc/tree-vect-stmts.cc > @@ -13353,6 +13353,12 @@ vect_analyze_stmt (vec_info *vinfo, > return opt_result::success (); > } > > + /* When we arrive here with a non-SLP statement and we are supposed > + to use SLP for everything fail vectorization. */ > + if (!node && param_vect_force_slp) > + return opt_result::failure_at (stmt_info->stmt, > + "needs non-SLP handling\n"); > + > ok = true; > if (!bb_vinfo > && (STMT_VINFO_RELEVANT_P (stmt_info) > -- Richard Biener <rguent...@suse.de> SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg, Germany; GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
