--- Comment #16 from dje at gcc dot gnu dot org 2006-06-07 21:39 ---
I checked with the IBM XLC team and they speculatively increase the alignment
of variables that could be auto-vectorized, so that gives another vote for that
method. They did mention that whole-program IPA allows them
--- Comment #7 from rsandifo at gcc dot gnu dot org 2006-06-06 08:54
---
Based on David's descripion, a reduced testcase appears to be:
static short f[100];
int
bar (void)
{
return f[0];
}
void
foo (void)
{
int i;
for (i = 0; i 100; i++)
f[i]++;
}
Looking at the assembly
--- Comment #8 from pinskia at gcc dot gnu dot org 2006-06-06 14:26 ---
What about instead of absolute numbers doing label subtraction for section
anchors and then we can defer the decision for the layout of the section until
after all functions are done compiling?
--
pinskia at gcc
--- Comment #9 from richard at codesourcery dot com 2006-06-06 15:02
---
Subject: Re: [4.2 Regression] wrong code in spec tests for -ftree-vectorize
-maltivec
pinskia at gcc dot gnu dot org [EMAIL PROTECTED] writes:
What about instead of absolute numbers doing label subtraction for
--- Comment #10 from dje at gcc dot gnu dot org 2006-06-06 15:10 ---
The auto-vectorizer is a Tree-SSA pass. The section anchors are an RTL pass.
I do not understand why the alignment of the vectorized variables is not known
at section anchor creation time.
--
--- Comment #11 from richard at codesourcery dot com 2006-06-06 15:16
---
Subject: Re: [4.2 Regression] wrong code in spec tests for -ftree-vectorize
-maltivec
dje at gcc dot gnu dot org [EMAIL PROTECTED] writes:
The auto-vectorizer is a Tree-SSA pass. The section anchors are an
--- Comment #12 from pinskia at gcc dot gnu dot org 2006-06-06 15:18
---
(In reply to comment #10)
The auto-vectorizer is a Tree-SSA pass. The section anchors are an RTL pass.
I do not understand why the alignment of the vectorized variables is not known
at section anchor creation
--- Comment #13 from dje at gcc dot gnu dot org 2006-06-06 15:22 ---
We're performing the auto-vectorization in unit-at-a-time-mode, so maybe we
need to recompile the other functions. It seems that we're going to encounter
more problems along these lines with LTO.
--
--- Comment #14 from richard at codesourcery dot com 2006-06-06 15:53
---
Subject: Re: [4.2 Regression] wrong code in spec tests for -ftree-vectorize
-maltivec
dje at gcc dot gnu dot org [EMAIL PROTECTED] writes:
We're performing the auto-vectorization in unit-at-a-time-mode, so
--- Comment #15 from rguenth at gcc dot gnu dot org 2006-06-06 16:35
---
For other reasons it would be nice to be able to place sync points in the
pass schedule where we re-start with going over all functions for the remaining
passes. Per function SSA form is requires for this, though,
--- Comment #4 from steven at gcc dot gnu dot org 2006-06-05 10:29 ---
Just like other bugs, this one will need a test case.
--
steven at gcc dot gnu dot org changed:
What|Removed |Added
--- Comment #5 from janis at gcc dot gnu dot org 2006-06-05 20:58 ---
I haven't yet had time to continue trying to come up with a minimized testcase
but hope to get to that soon. I had a vague hope that someone who understands
the section anchor support and has access to SPEC CPU2000
--- Comment #6 from dje at gcc dot gnu dot org 2006-06-05 21:25 ---
I still think this looks like an alignment problem. Without section anchors
GCC generates:
.lcomm domain_array,13916,16
.type domain_array, @object
.lcomm N_domains,4,4
.type
--
mmitchel at gcc dot gnu dot org changed:
What|Removed |Added
Priority|P3 |P1
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27770
--
pinskia at gcc dot gnu dot org changed:
What|Removed |Added
Summary|wrong code in spec tests for|[4.2 Regression] wrong code
|-ftree-vectorize
15 matches
Mail list logo
