https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108571
Bug ID: 108571
Summary: Fix for PR96373 regresses fabd_1.c with
-ftrapping-math
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Target Milestone: ---
Target: aarch64*-*-*
The upcoming fix for PR96373 means that, without -fno-trapping-math,
we can no longer optimise abs(a - b) to absolute-difference(a, b)
for predicated SVE loops. The subtraction becomes predicated,
and so we have the equivalent of:
abs(IFN_COND_SUB(loopmask, a, b, c))
where c is the fallback/else value. It isn't possible to tell from
this expression alone that the value of the abs result doesn't
matter for false loopmask elements (and that c will do just as
well as abs(c)). It therefore isn't possible to fold this into
a conditional absolute difference without further information.
One possible fix would be to add a gimple value and RTL object that
represents an unknown/undefined/don't-care value, a bit like LLVM's
undef. Another would be to make absolute difference a recognised
gimple operation (via ifns) and make the vectoriser produce it
directly. Or we could add a pass that back-propagates information
about uses to discover which elements of a vector actually matter
(which I think would also require gimple-level absolute difference).
