This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: RFA: fix handling of 'p' in constrain_operands

> Does this patch fix a bug? Do you have a testcase? Why does this patch make a difference? I'd like to know why the patch is correct before I approve it.

Sorry, when I tried to dig up the details, I found this patch was actually
papering over a bug introduced by a workaround for another bug which has
since been fixed.  Thus, I withdraw the patch.
The original problem was with some SHmedia patterns that I didn't add
in the fsf repository because of design divergence.

I thought this was the original bug because I mistook
recog_data.operand_mode[opno] for the mode of the operand.  It is
actually the mode of the construct that matches the operand.
Most of the time, these are the same when the predicate matches
(for CONST_INTs we currently have VOIDmode for the rtl of the operand, but
the mode of the matching construct is the actual mode of the data),
but for address_operand the difference is crucial.

I wonder if this is clear enough for other people, or if we should add
a comment there.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]