On Thu, 9 Feb 2017, Jakub Jelinek wrote:
> Hi!
>
> This is a big hammer approach for (ab) SSA_NAME issues in reassoc, on
> the following testcase we extend the lifetime of (ab) SSA_NAME and ICE
> because it can't be coalesced anymore. In theory we could try to be careful
> and whenever changing something that could extend (ab) lifetime, instead
> insert a non-(ab) SSA_NAME to copy the (ab) SSA_NAME to at the point of use
> and then use the temporary later on, but after looking for a while, it would
> mean changing way too many places, so this patch just punts on them or
> stops linearing when seeing them, so the ops vectors should not contain
> (ab) SSA_NAMEs, the various negations/undistributions etc. should also avoid
> them etc.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Ok.
I suppose at the point we add the ability to associate undefined-overflow
ops (and have to be careful to not introduce new undefined-overflow cases)
we'll have enough infrastructure in place to improve on this. Like
for undefined-overflow re-writing everything to unsigned arithmetic and
for the abnormal case inserting copies at the original point of uses.
(both less than ideal but working)
Thanks,
Richard.
> 2017-02-09 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/79411
> * tree-ssa-reassoc.c (is_reassociable_op): Return false if
> stmt operands are SSA_NAMEs used in abnormal phis.
> (can_reassociate_p): Return false if op is SSA_NAME used in abnormal
> phis.
>
> * gcc.c-torture/compile/pr79411.c: New test.
>
> --- gcc/tree-ssa-reassoc.c.jj 2017-01-24 23:29:09.000000000 +0100
> +++ gcc/tree-ssa-reassoc.c 2017-02-09 13:14:23.711878845 +0100
> @@ -605,7 +605,18 @@ is_reassociable_op (gimple *stmt, enum t
> if (is_gimple_assign (stmt)
> && gimple_assign_rhs_code (stmt) == code
> && has_single_use (gimple_assign_lhs (stmt)))
> - return true;
> + {
> + tree rhs1 = gimple_assign_rhs1 (stmt);
> + tree rhs2 = gimple_assign_rhs1 (stmt);
> + if (TREE_CODE (rhs1) == SSA_NAME
> + && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1))
> + return false;
> + if (rhs2
> + && TREE_CODE (rhs2) == SSA_NAME
> + && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs2))
> + return false;
> + return true;
> + }
>
> return false;
> }
> @@ -4989,6 +5000,8 @@ static bool
> can_reassociate_p (tree op)
> {
> tree type = TREE_TYPE (op);
> + if (TREE_CODE (op) == SSA_NAME && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op))
> + return false;
> if ((ANY_INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_WRAPS (type))
> || NON_SAT_FIXED_POINT_TYPE_P (type)
> || (flag_associative_math && FLOAT_TYPE_P (type)))
> --- gcc/testsuite/gcc.c-torture/compile/pr79411.c.jj 2017-02-09
> 13:34:06.016485562 +0100
> +++ gcc/testsuite/gcc.c-torture/compile/pr79411.c 2017-02-09
> 13:33:53.000000000 +0100
> @@ -0,0 +1,22 @@
> +/* PR tree-optimization/79411 */
> +
> +typedef struct __jmp_buf_tag { char buf[1024]; } jmp_buf[1];
> +extern int setjmp (jmp_buf);
> +extern int bar (unsigned int *);
> +extern jmp_buf *baz (void);
> +struct C { int c1; unsigned int c2, c3, c4; };
> +
> +void
> +foo (struct C *x, const int *y, unsigned int *z, unsigned int e, unsigned
> int g)
> +{
> + unsigned int d = 0;
> + unsigned long f;
> + setjmp (*baz ());
> + f = 1 + d;
> + if ((x->c1 || x->c2) && g && (!e || d >= 8))
> + d = 16;
> + else
> + d = 8;
> + if ((!x->c3 && !x->c4 || *y == 0) && !e && bar (z))
> + *z = 1 + f;
> +}
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB
21284 (AG Nuernberg)
