On 04/08/2016 07:05 PM, Jakub Jelinek wrote:
After IRC discussions, I've bootstrapped/regtested following patch that
just punts if *loc contains any paradoxical subregs, together with
additional statistics gathering that proved that the new testcase is
the only spot in which this patch makes a difference on x86_64-linux
and i686-linux bootstrap/regtest.
Of course on other targets it might affect more.
I ran the test with just not generating any REG_EQUAL notes here in
fwprop, and out of 4492 source files only three showed code generation
differences, which suggests we're not getting much value out of these.
Two of the cases where the same missed tree optimization, for which I've
filed PR70600.
2016-04-08 Jakub Jelinek <ja...@redhat.com>
PR rtl-optimization/70574
* fwprop.c (forward_propagate_and_simplify): Don't add
REG_EQUAL note if DF_REF_REG (use) is a paradoxical subreg.
(try_fwprop_subst): Don't add REG_EQUAL note if there are any
paradoxical subregs within *loc.
* gcc.target/i386/avx2-pr70574.c: New test.
Ok.
Bernd
