On Fri, Mar 24, 2023 at 11:08:54AM -0400, Andrew MacLeod wrote:
> Before floating point relations were added, we tried to sanitize
> value-relation records to not include non-sensensical records... ie x != x
> or x < x. Instead, we made a VREL_VARYING record with no operands.
>
> When floating point relation support was added, some of these were no longer
> non-sensical, AND we expanded the use of value_relation records into GORI
> shortly thereafter.
>
> As a result, this sanitization is no longer needed, nor desired. The Oracle
> does not create records with op1 == op2 already, so its only within GORI
> that these records can exist, and we shouldn't try to interpret them.
>
> The bug occurs because the "sanitized" records doesn't set op1 and op2, and
> changes the relation to VARYING.. and we expected the operands it to be set
> the way they were specified. We should not be setting a VREL_VARYING record
> if asked to set something else. In fact, we are missing some opportunities
> because we are trying to FP range-ops that op1 != op1 but its getting
> transformed into a VREL_VARYING record and not communicated properly.
>
> Currently bootstrapping on x86_64-pc-linux-gnu and assuming no regressions,
> OK for trunk?
>
> Andrew
> commit 1f02961b23976d35b10e2399708c6eb00632f9d6
> Author: Andrew MacLeod <amacl...@redhat.com>
> Date: Fri Mar 24 09:18:33 2023 -0400
>
> Don't interpret contents of a value_relation record.
>
> before floating point relations were added, we tried to sanitize
> value-relation records to not include non-sensensical records... ie
> x != x or x < x. INstead, we made a VREL_VARYING record with no
s/IN/In/
> operands.
>
> When floating point relations were supported, some of these were no
> longer non-sensical, AND we expanded the use of value_relation records
> into GORI.
>
> As a result, this sanitization is no longer needed. The Oracle
> does not create records with op1 == op2, so its only within GORI
> that these records can exist, and we shouldnt try to interpret them.
s/shouldnt/shouldn't/
>
> The bug occurs because the "sanitized" records doesnt set op1 anmd op2,
s/doesnt/doesn't/
> but we have a record so expected it to be set.
>
> PR tree-optimization/109265
> PR tree-optimization/109274
> gcc/
> * value-relation.h (value_relation::set_relation): Always create
> the
> record that is requested.
>
> gcc/testsuite/
> * gcc.dg/pr109274.c: New.
LGTM, indeed with floating point a != a isn't nonsensical but basically
__builtin_isnan (a) check.
I'll commit the Fortran testcase I've added in my version of the patch
incrementally when you commit.
Jakub
