* [Bug rtl-optimization/40209] ICE in iv_analyze_def caused by stale REG_UNUSED note
2009-05-20 18:47 [Bug rtl-optimization/40209] New: ICE in iv opts caused by stale REG_UNUSED note nvachhar at google dot com
@ 2009-07-24 4:01 ` ian at gcc dot gnu dot org
2009-07-24 6:25 ` ubizjak at gmail dot com
` (4 subsequent siblings)
5 siblings, 0 replies; 9+ messages in thread
From: ian at gcc dot gnu dot org @ 2009-07-24 4:01 UTC (permalink / raw)
To: gcc-bugs
------- Comment #2 from ian at gcc dot gnu dot org 2009-07-24 04:01 -------
Subject: Bug 40209
Author: ian
Date: Fri Jul 24 04:01:13 2009
New Revision: 150038
URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=150038
Log:
PR rtl-optimization/40209
* loop-iv.c (iv_analysis_loop_init): Call df_note_add_problem.
Modified:
trunk/gcc/ChangeLog
trunk/gcc/loop-iv.c
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40209
^ permalink raw reply [flat|nested] 9+ messages in thread* [Bug rtl-optimization/40209] ICE in iv_analyze_def caused by stale REG_UNUSED note
2009-05-20 18:47 [Bug rtl-optimization/40209] New: ICE in iv opts caused by stale REG_UNUSED note nvachhar at google dot com
2009-07-24 4:01 ` [Bug rtl-optimization/40209] ICE in iv_analyze_def " ian at gcc dot gnu dot org
@ 2009-07-24 6:25 ` ubizjak at gmail dot com
2009-07-24 6:27 ` steven at gcc dot gnu dot org
` (3 subsequent siblings)
5 siblings, 0 replies; 9+ messages in thread
From: ubizjak at gmail dot com @ 2009-07-24 6:25 UTC (permalink / raw)
To: gcc-bugs
------- Comment #3 from ubizjak at gmail dot com 2009-07-24 06:25 -------
Please also add the testcase from Comment #1 to gcc testsuite.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40209
^ permalink raw reply [flat|nested] 9+ messages in thread* [Bug rtl-optimization/40209] ICE in iv_analyze_def caused by stale REG_UNUSED note
2009-05-20 18:47 [Bug rtl-optimization/40209] New: ICE in iv opts caused by stale REG_UNUSED note nvachhar at google dot com
2009-07-24 4:01 ` [Bug rtl-optimization/40209] ICE in iv_analyze_def " ian at gcc dot gnu dot org
2009-07-24 6:25 ` ubizjak at gmail dot com
@ 2009-07-24 6:27 ` steven at gcc dot gnu dot org
2009-07-24 15:39 ` nvachhar at google dot com
` (2 subsequent siblings)
5 siblings, 0 replies; 9+ messages in thread
From: steven at gcc dot gnu dot org @ 2009-07-24 6:27 UTC (permalink / raw)
To: gcc-bugs
------- Comment #4 from steven at gcc dot gnu dot org 2009-07-24 06:27 -------
A hint, please, about why the patch of comment #2 would be the correct fix. As
far as I can tell, loop-iv doesn't need the notes and shouldn't have to clean
up other passes' mess. This patch also introduces a pass ordering restriction
that shouldn't be there. And at the very least, if this bug gets papered over
like this, then there should be a comment in loop-iv that the problem is only
added to fix bugs elsewhere...
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40209
^ permalink raw reply [flat|nested] 9+ messages in thread* [Bug rtl-optimization/40209] ICE in iv_analyze_def caused by stale REG_UNUSED note
2009-05-20 18:47 [Bug rtl-optimization/40209] New: ICE in iv opts caused by stale REG_UNUSED note nvachhar at google dot com
` (2 preceding siblings ...)
2009-07-24 6:27 ` steven at gcc dot gnu dot org
@ 2009-07-24 15:39 ` nvachhar at google dot com
2009-07-24 17:05 ` steven at gcc dot gnu dot org
2009-07-24 17:47 ` nvachhar at google dot com
5 siblings, 0 replies; 9+ messages in thread
From: nvachhar at google dot com @ 2009-07-24 15:39 UTC (permalink / raw)
To: gcc-bugs
[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 1352 bytes --]
------- Comment #5 from nvachhar at google dot com 2009-07-24 15:39 -------
Subject: Re: ICE in iv_analyze_def caused by
stale REG_UNUSED note
loop-iv does need notes, albeit indirectly through the single_set
function. single_set looks at the REG_UNUSED note, and if all but one
set have that note on it, then it asserts that a parallel set is a
single set. I do agree that this problem is pervasive through gcc.
Any pass that uses single_set is implicitly a user of reg ntoes.
On Thu, Jul 23, 2009 at 11:27 PM, steven at gcc dot gnu dot
org<gcc-bugzilla@gcc.gnu.org> wrote:
>
>
> ------- Comment #4 from steven at gcc dot gnu dot org 2009-07-24 06:27 -------
> A hint, please, about why the patch of comment #2 would be the correct fix. As
> far as I can tell, loop-iv doesn't need the notes and shouldn't have to clean
> up other passes' mess. This patch also introduces a pass ordering restriction
> that shouldn't be there. And at the very least, if this bug gets papered over
> like this, then there should be a comment in loop-iv that the problem is only
> added to fix bugs elsewhere...
>
>
> --
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40209
>
> ------- You are receiving this mail because: -------
> You reported the bug, or are watching the reporter.
>
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40209
^ permalink raw reply [flat|nested] 9+ messages in thread* [Bug rtl-optimization/40209] ICE in iv_analyze_def caused by stale REG_UNUSED note
2009-05-20 18:47 [Bug rtl-optimization/40209] New: ICE in iv opts caused by stale REG_UNUSED note nvachhar at google dot com
` (3 preceding siblings ...)
2009-07-24 15:39 ` nvachhar at google dot com
@ 2009-07-24 17:05 ` steven at gcc dot gnu dot org
2009-07-24 17:47 ` nvachhar at google dot com
5 siblings, 0 replies; 9+ messages in thread
From: steven at gcc dot gnu dot org @ 2009-07-24 17:05 UTC (permalink / raw)
To: gcc-bugs
------- Comment #6 from steven at gcc dot gnu dot org 2009-07-24 17:04 -------
Then we should write a new function, something like this in df.h perhaps:
/* Given an INSN, return a SET expression if the insn has only one
SET whose SET_DEST is used. If SET_DEST is memory, then the SET is
assumed to be used somewhere, always.
This means that:
* If an insn has multiple SETs to only registers, but only one of those
SETs is used, return that one set.
* If an insn has multiple SETs to only registers, but none one of those
SETs is used, return NULL_RTX.
* If an insn has multiple SETs to registers and/or memory, but the SETs to
registers are dead, return the one memory SET, or NULL_RTX if there are
multiple sets to memory.
Only accept INSNs. single_set() can take a pattern, but that is
"old skool" style that should be abolished.
May miss some single_set SETs that single_set() can find, because this
function just looks for the number of uses of a register to determine
whether a set is used (which may fail for hard registers and for unusual
situations with pseudos that somehow are set twice but never used) whereas
single_set() uses liveness information through REG_UNUSED notes. Still,
I expect that, in practice, this function works just as well as
single_set().
One more caveat: The DF user must use incremental scanning, or the
DF_REG_USE_COUNT may be wrong. The result of df_single_set would be
unconservative for new registers since the last scan (DF_REG_USE_COUNT
will still be 0, indicating the new reg is unused and its SET can be
ignored). */
rtx
df_single_set (rtx insn)
{
rtx pat;
gcc_assert (INSN_P (insn));
/* ??? Could even just fall back if the current DF mode is not incremental
scanning. */
pat = PATTERN (insn);
if (GET_CODE (pat) == SET)
return pat;
else if (GET_CODE (pat) == PARALLEL)
{
int i;
rtx set = NULL_RTX;
for (i = 0; i < XVECLEN (pat, 0); i++)
{
rtx sub = XVECEXP (pat, 0, i);
switch (GET_CODE (sub))
{
case USE:
case CLOBBER:
break;
case SET:
if (! set)
{
rtx dest = SET_DEST (set);
if (GET_CODE (dest) == SUBREG)
dest = SUBREG_REG (dest);
if (! REG_P (dest)
|| DF_REG_USE_COUNT (dest) > 0)
set = sub;
}
else
return NULL_RTX;
break;
default:
return NULL_RTX;
}
}
}
else
return NULL_RTX;
}
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40209
^ permalink raw reply [flat|nested] 9+ messages in thread* [Bug rtl-optimization/40209] ICE in iv_analyze_def caused by stale REG_UNUSED note
2009-05-20 18:47 [Bug rtl-optimization/40209] New: ICE in iv opts caused by stale REG_UNUSED note nvachhar at google dot com
` (4 preceding siblings ...)
2009-07-24 17:05 ` steven at gcc dot gnu dot org
@ 2009-07-24 17:47 ` nvachhar at google dot com
5 siblings, 0 replies; 9+ messages in thread
From: nvachhar at google dot com @ 2009-07-24 17:47 UTC (permalink / raw)
To: gcc-bugs
[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 3707 bytes --]
------- Comment #7 from nvachhar at google dot com 2009-07-24 17:47 -------
Subject: Re: ICE in iv_analyze_def caused by
stale REG_UNUSED note
This looks reasonable to me. Given the caveats you mention in the
comment, this would need to be tested for correctness (esp. w.r.t. the
newly added registers) and obviously would need to be tested for
performance regressions.
On Fri, Jul 24, 2009 at 5:04 PM, steven at gcc dot gnu dot
org<gcc-bugzilla@gcc.gnu.org> wrote:
>
>
> ------- Comment #6 from steven at gcc dot gnu dot org 2009-07-24 17:04 -------
> Then we should write a new function, something like this in df.h perhaps:
>
> /* Given an INSN, return a SET expression if the insn has only one
> SET whose SET_DEST is used. If SET_DEST is memory, then the SET is
> assumed to be used somewhere, always.
>
> This means that:
> * If an insn has multiple SETs to only registers, but only one of those
> SETs is used, return that one set.
> * If an insn has multiple SETs to only registers, but none one of those
> SETs is used, return NULL_RTX.
> * If an insn has multiple SETs to registers and/or memory, but the SETs to
> registers are dead, return the one memory SET, or NULL_RTX if there are
> multiple sets to memory.
>
> Only accept INSNs. single_set() can take a pattern, but that is
> "old skool" style that should be abolished.
>
> May miss some single_set SETs that single_set() can find, because this
> function just looks for the number of uses of a register to determine
> whether a set is used (which may fail for hard registers and for unusual
> situations with pseudos that somehow are set twice but never used) whereas
> single_set() uses liveness information through REG_UNUSED notes. Still,
> I expect that, in practice, this function works just as well as
> single_set().
>
> One more caveat: The DF user must use incremental scanning, or the
> DF_REG_USE_COUNT may be wrong. The result of df_single_set would be
> unconservative for new registers since the last scan (DF_REG_USE_COUNT
> will still be 0, indicating the new reg is unused and its SET can be
> ignored). */
>
> rtx
> df_single_set (rtx insn)
> {
> rtx pat;
>
> gcc_assert (INSN_P (insn));
>
> /* ??? Could even just fall back if the current DF mode is not incremental
> scanning. */
> pat = PATTERN (insn);
> if (GET_CODE (pat) == SET)
> return pat;
> else if (GET_CODE (pat) == PARALLEL)
> {
> int i;
> rtx set = NULL_RTX;
> for (i = 0; i < XVECLEN (pat, 0); i++)
> {
> rtx sub = XVECEXP (pat, 0, i);
> switch (GET_CODE (sub))
> {
> case USE:
> case CLOBBER:
> break;
>
> case SET:
> if (! set)
> {
> rtx dest = SET_DEST (set);
> if (GET_CODE (dest) == SUBREG)
> dest = SUBREG_REG (dest);
> if (! REG_P (dest)
> || DF_REG_USE_COUNT (dest) > 0)
> set = sub;
> }
> else
> return NULL_RTX;
> break;
>
> default:
> return NULL_RTX;
> }
> }
> }
> else
> return NULL_RTX;
> }
>
>
> --
>
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40209
>
> ------- You are receiving this mail because: -------
> You reported the bug, or are watching the reporter.
>
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40209
^ permalink raw reply [flat|nested] 9+ messages in thread