On Mon, Jan 12, 2015 at 02:29:53PM -0700, Jeff Law wrote:
On 01/12/15 12:59, Jakub Jelinek wrote:
Hi!
As mentioned in the PR, giving up for all vector mode extensions
is unnecessary, but unlike scalar integer extensions, where the low part
of the extended value is the original value, for vectors this is not true,
thus the old value is lost. Which means we can perform REE, but only if
all uses of the definition are the same (code+mode) extension.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2015-01-12 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/64286
* ree.c (add_removable_extension): Don't add vector mode
extensions if all uses of the source register aren't the same
vector extensions.
* gcc.target/i386/avx2-pr64286.c: New test.
Does it make sense to remove your change for 59754 in combine_reaching_defs?
Shouldn't this patch handle that case as well?
You're right, this patch handles that too.
New patch, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?
2015-01-13 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/64286
* ree.c (combine_reaching_defs): Move part of comment earlier,
remove !SCALAR_INT_MODE_P check.
(add_removable_extension): Don't add vector mode
extensions if all uses of the source register aren't the same
vector extensions.
* gcc.target/i386/avx2-pr64286.c: New test.