On Tue, 4 Feb 2014, Jakub Jelinek wrote:
> Hi!
>
> If a load or store is in EH tables, but doesn't have EH edges out of it,
> e.g. for must-not-throw regions, we weren't cleaning it up and adding
> MASK_{LOAD,STORE} into EH region instead.
> Similarly, the vectorizer needs to add the vectorized stmts to the same
> EH region as the corresponding {,masked} load/store (second hunk).
> The rest is just making the compiler more robust and not crash if DCE etc.
> decides to remove the lhs of MASK_LOAD.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> P.S.: the patch doesn't fix the old issue that -fnon-call-exceptions
> loads are never removed from the vectorized loops, stores aren't actually
> removed when transforming the stmt, but later on and handling scalar loads
> similarly (turning them into say assignments of zero) didn't look trivial.
>
> 2014-02-04 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/60023
> * tree-if-conv.c (predicate_mem_writes): Pass true instead of
> false to gsi_replace.
> * tree-vect-stmts.c (vect_finish_stmt_generation): If stmt
> has been in some EH region and vec_stmt could throw, add
> vec_stmt into the same EH region.
> * tree-data-ref.c (get_references_in_stmt): If IFN_MASK_LOAD
> has no lhs, ignore it.
> * internal-fn.c (expand_MASK_LOAD): Likewise.
>
> * g++.dg/vect/pr60023.cc: New test.
>
> --- gcc/tree-if-conv.c.jj 2014-01-03 11:40:57.000000000 +0100
> +++ gcc/tree-if-conv.c 2014-02-03 16:01:40.250716523 +0100
> @@ -1723,7 +1723,7 @@ predicate_mem_writes (loop_p loop)
> new_stmt
> = gimple_build_call_internal (IFN_MASK_STORE, 4, addr, ptr,
> mask, rhs);
> - gsi_replace (&gsi, new_stmt, false);
> + gsi_replace (&gsi, new_stmt, true);
> }
> else if (gimple_vdef (stmt))
> {
> --- gcc/tree-vect-stmts.c.jj 2014-01-16 20:54:59.000000000 +0100
> +++ gcc/tree-vect-stmts.c 2014-02-03 16:29:00.881394627 +0100
> @@ -1691,6 +1691,13 @@ vect_finish_stmt_generation (gimple stmt
> }
>
> gimple_set_location (vec_stmt, gimple_location (stmt));
> +
> + /* While EH edges will generally prevent vectorization, stmt might
> + e.g. be in a must-not-throw region. Ensure newly created stmts
> + that could throw are part of the same region. */
> + int lp_nr = lookup_stmt_eh_lp (stmt);
> + if (lp_nr != 0 && stmt_could_throw_p (vec_stmt))
> + add_stmt_to_eh_lp (vec_stmt, lp_nr);
Can you instead use maybe_clean_or_replace_eh_stmt (stmt, vec_stmt)?
That should (maybe) take care of the missed DCE as well (maybe only
for the first scalar stmt though).
Otherwise ok.
Thanks,
Richard.
> }
>
> /* Checks if CALL can be vectorized in type VECTYPE. Returns
> --- gcc/tree-data-ref.c.jj 2014-01-03 11:40:57.000000000 +0100
> +++ gcc/tree-data-ref.c 2014-02-03 16:05:42.111480787 +0100
> @@ -4401,6 +4401,8 @@ get_references_in_stmt (gimple stmt, vec
> switch (gimple_call_internal_fn (stmt))
> {
> case IFN_MASK_LOAD:
> + if (gimple_call_lhs (stmt) == NULL_TREE)
> + break;
> ref.is_read = true;
> case IFN_MASK_STORE:
> ref.ref = fold_build2 (MEM_REF,
> --- gcc/internal-fn.c.jj 2014-02-03 11:48:17.000000000 +0100
> +++ gcc/internal-fn.c 2014-02-03 16:08:55.606492962 +0100
> @@ -820,6 +820,8 @@ expand_MASK_LOAD (gimple stmt)
>
> maskt = gimple_call_arg (stmt, 2);
> lhs = gimple_call_lhs (stmt);
> + if (lhs == NULL_TREE)
> + return;
> type = TREE_TYPE (lhs);
> rhs = fold_build2 (MEM_REF, type, gimple_call_arg (stmt, 0),
> gimple_call_arg (stmt, 1));
> --- gcc/testsuite/g++.dg/vect/pr60023.cc.jj 2014-02-03 16:39:17.996271473
> +0100
> +++ gcc/testsuite/g++.dg/vect/pr60023.cc 2014-02-03 16:39:36.455176379
> +0100
> @@ -0,0 +1,80 @@
> +// PR tree-optimization/60023
> +// { dg-do compile }
> +// { dg-additional-options "-O3 -std=c++11 -fnon-call-exceptions" }
> +// { dg-additional-options "-mavx2" { target { i?86-*-* x86_64-*-* } } }
> +
> +struct A { A (); ~A (); };
> +
> +void
> +f1 (int *p, int *q, int *r) noexcept (true)
> +{
> + int i;
> + for (i = 0; i < 1024; i++)
> + if (r[i])
> + p[i] = q[i] + 1;
> +}
> +
> +void
> +f2 (int *p, int *q, int *r)
> +{
> + int i;
> + for (i = 0; i < 1024; i++)
> + if (r[i])
> + p[i] = q[i] + 1;
> +}
> +
> +void
> +f3 (int *p, int *q) noexcept (true)
> +{
> + int i;
> + for (i = 0; i < 1024; i++)
> + p[i] = q[i] + 1;
> +}
> +
> +void
> +f4 (int *p, int *q)
> +{
> + int i;
> + for (i = 0; i < 1024; i++)
> + p[i] = q[i] + 1;
> +}
> +
> +void
> +f5 (int *p, int *q, int *r) noexcept (true)
> +{
> + int i;
> + A a;
> + for (i = 0; i < 1024; i++)
> + if (r[i])
> + p[i] = q[i] + 1;
> +}
> +
> +void
> +f6 (int *p, int *q, int *r)
> +{
> + int i;
> + A a;
> + for (i = 0; i < 1024; i++)
> + if (r[i])
> + p[i] = q[i] + 1;
> +}
> +
> +void
> +f7 (int *p, int *q) noexcept (true)
> +{
> + int i;
> + A a;
> + for (i = 0; i < 1024; i++)
> + p[i] = q[i] + 1;
> +}
> +
> +void
> +f8 (int *p, int *q)
> +{
> + int i;
> + A a;
> + for (i = 0; i < 1024; i++)
> + p[i] = q[i] + 1;
> +}
> +
> +// { dg-final { cleanup-tree-dump "vect" } }
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
GF: Jeff Hawn, Jennifer Guild, Felix Imend"orffer
