On 11/20/2015 11:04 AM, Manuel López-Ibáñez wrote:
On 20 November 2015 at 17:42, Jeff Law <l...@redhat.com> wrote:
So we have to detangle the operand shortening from warning detection. Kai's
idea was to first make the shortening code "pure" in the sense that it would
have no side effects other than to generate the warnings. Canonicalization
and other transformations would still occur internally, but not be reflected
in the IL.
That was the overall plan and he posted a patch for that. But that patch
didn't do the due diligence to verify that once the shortening code was made
"pure" that we didn't regress on the quality of the code we generated.
I thought that the original plan was to make the warning code also use
match.pd. That is, that all folding, including FE folding, will be
match.pd-based. Has this changed?
I don't think that's changed.
Detangling the two is the first step. Once detangled we can then look to
move the warning to a more suitable location -- right now it's in the
C/C++ front-ends, firing way too early. Instead it ought to be checked
in gimple form, after match.pd canonicalization and simplifications.
Jeff
