On 12 October 2011 17:54, Jakub Jelinek wrote:
> Hi!
Hi,
>
> This patch allows vectorization of some loops that use
> bool (which is especially important now that we use bool more often
> even for stmts that weren't originally using bool in the sources),
> in particular (when bool is cast to an
Update expected output for
> architectures with multiple vector sizes.
>
> --
> Matthew Gretton-Dann
> Principal Engineer, PD Software - Tools, ARM Ltd[attachment "1110-
> fix-pr30858.txt" deleted by Ira Rosen/Haifa/IBM]
copy)
@@ -1,3 +1,21 @@
+2011-10-16 Ira Rosen
+
+ * tree-vect-stmts.c (vectorizable_load): For SLP without permutation
+ treat the first load of the node as the first element in its
+ interleaving chain.
+ * tree-vect-slp.c (vect_get_and_check_slp_defs): Swap the operan
Hi,
This patch fixes another occurrence of the same bug as in PR 50635.
Bootstrapped and tested on powerpc64-suse-linux.
Committed.
Ira
ChangeLog:
PR tree-optimization/50727
* tree-vect-patterns.c (vect_operation_fits_smaller_type): Add
DEF_STMT to the list of statement
Hi,
With Jakub's patch for bool types the 3 loops in gcc.dg/vect/vect-21.c
are now vectorizable on targets that support vector conditions.
Tested on powerpc64-suse-linux.
Committed.
Ira
testsuite/ChangeLog:
* gcc.dg/vect/vect-21.c: Expect the loops to get vectorized on
targets that suppo
Hi,
vect-114.c doesn't have misaligned accesses, so there is no need in
vect_hw_misalign.
Tested on powerpc64-suse-linux.
Committed.
Ira
testsuite/ChangeLog:
PR tree-optimization/50746
* gcc.dg/vect/vect-114.c: Remove vect_hw_misalign.
Index: testsuite/gcc.dg/vect/vect-114.c
=
On 18 October 2011 11:43, Jakub Jelinek wrote:
> On Tue, Oct 18, 2011 at 11:39:22AM +0200, Ira Rosen wrote:
>> On 2 October 2011 10:30, Ira Rosen wrote:
>> > On 29 September 2011 17:30, Ramana Radhakrishnan
>> > wrote:
>> >> On 19
On 20 October 2011 23:50, Jakub Jelinek wrote:
> Hi!
Hi,
>
> While looking at *.vect dumps from Polyhedron, I've noticed the lack
> of SLP vectorization of builtin calls.
>
> This patch is an attempt to handle at least 1 and 2 operand builtin calls
> (SLP doesn't handle ternary stmts either yet)
On 21 October 2011 14:52, Jakub Jelinek wrote:
> On Fri, Oct 21, 2011 at 02:37:06PM +0200, Ira Rosen wrote:
>> > @@ -1620,7 +1615,13 @@ vectorizable_call (gimple stmt, gimple_s
>> >
>> > gcc_assert (!gimple_vuse (stmt));
>> >
>> > - if (modifier
On 21 October 2011 16:25, Jakub Jelinek wrote:
> On Fri, Oct 21, 2011 at 03:44:11PM +0200, Ira Rosen wrote:
>> But it's OK to allow modifier != NONE if it's not SLP, so we need &&, no?
>
> Well, in my patch that check was guarded by the if (slp_node ...),
> s
gt; Support of loops with bool.
>
>
>
>
>
44c77
< other then reduction cycles in nested loops) (2009-06-16)
---
> other than reduction cycles in nested loops) (2009-06-16)
82c115,116
< to this project include Revital Eres, Rich
++.dg/vect/vect.exp: Set target dependent flags for slp-* tests.
* g++.dg/vect/slp-pr50819.cc: New test.
Index: ChangeLog
===
--- ChangeLog (revision 180333)
+++ ChangeLog (working copy)
@@ -1,3 +1,21 @@
+2011-10-23
Hi,
With this patch we are able to stop basic block analysis in case of
unsupported data-ref and still vectorize the first part of the basic
block.
Bootstrapped and tested on powerpc64-suse-linux.
Committed.
Ira
ChangeLog:
PR tree-optimization/50730
* tree-vect-data-refs.c (vec
On 28 October 2011 20:44, Jakub Jelinek wrote:
> Hi!
Hi,
>
> This patch implements what I've talked about, with this we can now
> with -mavx2 as well as -mxop vectorize long long/unsigned long long
> shifts by int or long long/unsigned long long shifts by long long
> (where the FE casts it to in
On 31 October 2011 11:53, Jakub Jelinek wrote:
> On Sun, Oct 30, 2011 at 12:38:32AM -0400, David Miller wrote:
>> gcc.dg/pr48616.c segfaults on sparc as of a day or two ago
>>
>> vectorizable_shift() crashes because op1_vectype is NULL and
>> we hit this code path:
>>
>> /* Vector shifted by vec
On 31 October 2011 13:23, Jakub Jelinek wrote:
> On Mon, Oct 31, 2011 at 01:14:25PM +0200, Ira Rosen wrote:
>> > --- gcc/tree-vect-stmts.c.jj 2011-10-28 16:21:06.0 +0200
>> > +++ gcc/tree-vect-stmts.c 2011-10-31 10:27:57.0 +0100
>
Hi,
With the recent patches for __restrict__, the outer loop in
gcc.dg/vect/no-scevccp-outer-6-global.c is now vectorizable, because
it doesn't require loop versioning for alias anymore. The comment in
the test is probably obsolete, and checking for widen-mult doesn't
make much sense, because the
on't understand how it happened. Thanks for noticing!
I'll unrevert it after testing on powerpc64-suse-linux.
Thanks,
Ira
>
>
> Jakub
>
Index: ChangeLog
===
--- ChangeLog (revision 180930)
+++ ChangeL
Hi,
This patch adds a support of conditions in SLP.
It also fixes a bug in pattern handling in SLP (we should put pattern
statements instead of original statements in the root), and allows
pattern def-stmts in SLP.
Bootstrapped on powerpc64-suse-linux and tested on
powerpc64-suse-linux and x86_64
On 7 November 2011 20:35, Jakub Jelinek wrote:
> Hi!
Hi,
>
> Here is an updated patch, which handles both modifier == NONE
> and modifier == NARROW for SLP, after all it wasn't that hard.
> Additionally it checks that the fndecls and various call flags
> match, and adds some testcases.
>
> Boots
On 8 November 2011 09:22, Jakub Jelinek wrote:
> First of all, whether copysignf, sqrtf and/or lrint are vectorized is
> very much target specific, should I guard the dg-final lines with
> { target { i?86-*-* x86_64-*-* } }
> resp.
> { target { { i?86-*-* x86_64-*-* } && !lp64 } }
> (the latter f
On 8 November 2011 11:32, Jakub Jelinek wrote:
> On Tue, Nov 08, 2011 at 10:03:23AM +0200, Ira Rosen wrote:
>> The second option would be nicer.
> ...
>
> Thanks. Here is an updated patch, will bootstrap/regtest it now.
> Ok for trunk if it passes?
Yes.
Thanks,
Ira
Hi,
Some of the recently added vectorizer pattern detection functions
create pattern def stmts, and set vectype for these statements. This
causes an assert failure in vect_determine_vectorization_factor, since
we only expect data-refs and pattern statements themselves to have the
vectype set. This
On 9 November 2011 23:32, Jakub Jelinek wrote:
> Hi!
>
> When a bool store gets a pattern stmt, we need to update
> DR_STMT (otherwise the original rather than replaced stmts
> are used e.g. for interleaving etc.).
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, testcase
> tested on pow
Hi,
This patch handles CALL_EXPRs in constant/invariant operand creation in SLP.
Bootstrapped and tested on powerpc64-suse-linux.
Committed.
Ira
ChangeLog:
PR tree-optimization/51058
* tree-vect-slp.c (vect_get_constant_vectors): Handle CALL_EXPR.
testsuite/ChangeLog:
On 10 November 2011 21:31, Jakub Jelinek wrote:
> Hi!
>
> This patch fixes some compiler memory leaks in SLP.
> For vect_free_oprnd_info I've removed the FREE_DEF_STMTS argument
> and am freeing the defs always, but set them to NULL when moving the vectors
> over elsewhere, because otherwise if ve
On 11 November 2011 17:32, Jakub Jelinek wrote:
> Hi!
Hi,
>
> Removing the scalar call in vectorizable_call for SLP vectorization
> is too early, when another SLP instance refers to the same scalar call,
> we'll ICE because that stmt doesn't have bb anymore or gsi_for_stmt
> doesn't succeed for
On 11 November 2011 19:06, Jakub Jelinek wrote:
> On Fri, Nov 11, 2011 at 06:57:58PM +0200, Ira Rosen wrote:
>> On 11 November 2011 17:32, Jakub Jelinek wrote:
>> > 2011-11-11 Jakub Jelinek
>> >
>> >PR tree-optimization/51058
>> >*
> + /* Now we have matched the statement pattern
> +
> +rhs1 = (T1)x;
> +rhs2 = (T1)y;
> +op_result = rhs1 OP rhs2;
> +lhs = (T2)op_result;
Just a note that the patch I proposed for the vectorizer (
http://gcc.gnu.org/ml/gcc-patches/2011-07/msg01472.html) also ha
Richard Guenther wrote on 02/08/2011 01:33:49 PM:
>
> On Tue, 2 Aug 2011, Ira Rosen wrote:
>
> >
> > > + /* Now we have matched the statement pattern
> > > +
> > > +rhs1 = (T1)x;
> > > +rhs2 = (T1)y;
> > > +
Richard Guenther wrote on 02/08/2011 04:25:58 PM:
>
> Thinking about it it probably makes sense to keep a variant of this
> in the vectorizer - after all it has quite specific requirements on
> operand sizes while VRP would probably demote as far as possible
> (maybe taking PROMOTE_MODE into a
On 19 July 2011 09:44, Ira Rosen wrote:
> Hi,
>
> This patch tries to reduce over-promotion of vector operations that
> could be done with narrower elements, e.g., for
>
> char a;
> int b, c;
> short d;
>
> b = (int) a;
> c = b << 2;
> d = (short) c;
Hi,
In case of multiple loop exit phis in vectorization of reduction we reduce
them to one vector. The result of this reduction is later ignored in case
we do the final value extraction with scalar code. This causes wrong code
generation for gfortran.dg/forall_7.f90 with -O3 -funroll-loops on Cel
Hi,
In vectorization of reduction we use a dummy def_type when getting a
copy of a vector operand. Therefore, instead of just using a constant
operand from a previous copy, we try to create a copy of it, causing
the failure. This patch adds a call to vect_is_simple_use () to get a
correct def_type
Hi,
This patch adds a check in vect_operation_fits_smaller_type () that a
widening statement has a stmt_vec_info, i.e., that it is a loop
statement.
Bootstrapped and tested on powerpc64-suse-linux.
Committed.
Ira
ChangeLog:
PR tree-optimization/50039
* tree-vect-patterns.c (vect_opera
Hi,
This patch changes the default vector size for auto-vectorization on
ARM NEON to 128 bits. This new version is a result of a discussion
with Richard and Ramana.
wwwdocs changes will follow shortly.
Bootstrapped and tested on arm-linux-gnueabi. The testsuite changes
were also checked on power
Hi,
The first part was already reviewed several months ago, but I am
resubmitting it along with -mvectorize-with-neon-double documentation.
OK to commit?
Thanks,
Ira
* htdocs/gcc-4.7/changes.html (targets): Document ARM NEON default
vector size change and -mvectorize-with-neon-double
Jakub Jelinek wrote on 22/08/2011 05:22:59 PM:
> Hi!
>
> The following testcase ICEs, because gsi_end_p (*gsi) and thus
> there is no stmt after it from which to copy over the location.
> As can be seen in the PR, we could do ugly hacks to retrieve locus
> from previous stmt (non-debug of cours
gcc-patches-ow...@gcc.gnu.org wrote on 24/08/2011 02:20:50 PM:
>
> This avoids the file/location clutter in front of each line
> in the vectorizer dump. While this is useful for people
> requesting -fvectorizer-verbose=N in dump files this makes
> you unable to compare dumps for testcases on a b
Richard Guenther wrote on 01/09/2011 10:33:23 AM:
> On Thu, 1 Sep 2011, Ira Rosen wrote:
>
> >
> >
> > gcc-patches-ow...@gcc.gnu.org wrote on 24/08/2011 02:20:50 PM:
> > >
> > > This avoids the file/location clutter in front of each line
> > >
Hi,
When vectorizing a function call we replace the original call with a
dummy statement to ensure that DCE later removes it. We also remove
its stmt_vec_info, which causes the segfault when we try to access it
through related pattern stmt. The following patch updates related
pattern stmt to be th
On 1 September 2011 11:16, Jakub Jelinek wrote:
> On Thu, Sep 01, 2011 at 10:14:29AM +0200, Richard Guenther wrote:
>> On Thu, Sep 1, 2011 at 10:12 AM, Ira Rosen wrote:
>> > When vectorizing a function call we replace the original call with a
>> > dummy statement to en
Richard Guenther wrote on 01/09/2011 11:13:29 AM:
> > > > IMO it's a bad idea. It's now impossible to find anything when
> > compiling a
> > > > big file. How about only removing the file name?
> > >
> > > How about, as Micha suggested, print the location of the loop
> > > we currently investig
Richard Guenther wrote on 01/09/2011 12:26:25 PM:
> > > Well, it seems to be different what everybody else does and it's
> > > highly redundant for a whole bunch of lines.
> > >
> > > But, it solves my diff issue and the overly long lines as well.
> > >
> > > Your patch changes both dump-file a
its stmt_vec_info.
Bootstrapped and tested on powerpc64-suse-linux.
Committed.
Ira
ChangeLog:
2011-09-04 Jakub Jelinek
Ira Rosen
PR tree-optimization/50208
* tree-vect-patterns.c (vect_handle_widen_mult_by_const):
Add an argument. Check that def_stmt is inside
On 17 August 2011 15:49, Richard Earnshaw wrote:
> On 16/08/11 10:28, Ira Rosen wrote:
>> Hi,
>>
>> This patch changes the default vector size for auto-vectorization on
>> ARM NEON to 128 bits. This new version is a result of a discussion
>> with Richard and R
Hi,
This patch makes data-refs analysis to not fail if simple_iv returns
false in basic block SLP.
Bootstrapped and tested on powerpc64-suse-linux.
OK for mainline?
Thanks,
Ira
ChangeLog:
* tree-data-ref.c (dr_analyze_innermost): Rename to...
(dr_analyze_innermost_1): ... this. Add
Bootstrapped and tested on powerpc64-suse-linux.
Committed to trunk.
Ira
ChangeLog:
* tree-vect-data-refs.c (vect_analyze_data_ref_dependence): Allow
read-after-read dependencies in basic block SLP.
testsuite/ChangeLog:
* gcc.dg/vect/bb-slp-25.c: New.
Index: tree-vect-dat
Hi,
Strided accesses of single element or with gaps may require creation
of epilogue loop. At the moment we don't support peeling for outer
loops, therefore, we should not allow such strided accesses in outer
loops.
Bootstrapped and tested on powerpc64-suse-linux.
Committed to trunk.
Now testing
Hi,
This patch adds a missing handling of MAX/MIN_EXPR in SLP reduction.
Boostrapped and tested on powerpc64-suse-linux.
Committed to trunk.
Ira
ChangeLog:
PR tree-optimization/50414
* tree-vect-slp.c (vect_get_constant_vectors): Handle MAX_EXPR and
MIN_EXPR.
testsuite
Hi,
This patch adds an if-statement to avoid loop vectorization and fixes
underscores around restrict in gcc.dg/vect/bb-slp-25.c.
Tested by Dominique on x86_64-apple-darwin10 and on x86_64-suse-linux.
Committed to trunk.
Ira
2011-09-18 Dominique d'Humieres
Ira
ls.
* g++.dg/vect/slp-pr50413.cc: New.
Index: ChangeLog
===
--- ChangeLog (revision 178967)
+++ ChangeLog (working copy)
@@ -1,3 +1,9 @@
+2011-09-19 Ira Rosen
+
+ PR tree-optimization/50413
+ * tree-v
Hi,
g++.dg/vect/slp-pr50413.cc should check that the basic block in shift
() doesn't get vectorized. This patch removes other functions that
contain basic blocks that can be vectorizable on some platforms.
Tested on x86_64-suse-linux.
Committed to trunk.
Ira
testsuite/ChangeLog:
* g++.
Hi,
This patch adds a missing support of constant operands in reduction in SLP.
Bootstrapped and tested on powerpc64-suse-linux.
Committed.
Ira
ChangeLog:
PR tree-optimization/50451
* tree-vect-slp.c (vect_get_constant_vectors): Don't fail for
constant operands in reduc
On 23 September 2011 23:09, Richard Guenther wrote:
> On Wed, Sep 14, 2011 at 2:01 PM, Ira Rosen wrote:
>> Hi,
>>
>> This patch makes data-refs analysis to not fail if simple_iv returns
>> false in basic block SLP.
>>
>> Bootstrapped and tested on powerpc64
Hi,
This patch supports an automatic choice of vector size in basic block
vectorization similar to the loop vectorization case.
I am not sure about the new keyword.
Bootstrapped on powerpc64-suse-linux, tested on powerpc64-suse-linux
and arm-linux-gnueabi.
Thanks,
Ira
ChangeLog:
* tre
On 25 September 2011 14:45, Richard Guenther wrote:
> On Sun, Sep 25, 2011 at 12:59 PM, Ira Rosen wrote:
>> Hi,
>>
>> This patch supports an automatic choice of vector size in basic block
>> vectorization similar to the loop vectorization case.
>>
>> I am
On 26 September 2011 17:12, Richard Guenther wrote:
> On Mon, Sep 19, 2011 at 9:54 AM, Ira Rosen wrote:
>> Hi,
>>
>> This patch adds a support of widening shift left. The following
>> pattern is detected:
>>
>> type a_t;
>> TYPE a_T, res_T;
>
geLog (working copy)
@@ -1,3 +1,18 @@
+2011-09-27 Ira Rosen
+
+ * tree-vect-stmts.c (vectorizable_type_demotion): Handle basic block
+ vectorization.
+ (vectorizable_type_promotion): Likewise.
+ (vect_analyze_stmt): Call vectorizable_type_demotio
On 29 September 2011 17:30, Ramana Radhakrishnan
wrote:
> On 19 September 2011 08:54, Ira Rosen wrote:
>
>>
>> Bootstrapped on powerpc64-suse-linux, tested on powerpc64-suse-linux
>> and arm-linux-gnueabi
>> OK for mainline?
>
> Sorry I missed this patch. Is
Hi,
Michael pointed out this problem in
check_effective_target_vect_multiple_sizes and
check_effective_target_vect64 that I added lately.
Tested on powerpc64-suse-linux.
Committed as obvious.
Thanks,
Ira
testsuite/ChangeLog:
* lib/target-supports.exp (check_effective_target_vect_multip
On 5 October 2011 20:06, Jakub Jelinek wrote:
> Hi!
>
> If vect_recog_func fails (or the other spot where vect_pattern_recog_1
> returns early), the vector allocated in the function isn't freed, leading
> to memory leak. But, more importantly, doing a VEC_alloc + VEC_free
> num_stmts_in_loop * NU
On 6 October 2011 18:17, Jakub Jelinek wrote:
> Hi!
>
> Since Richard's changes recently to allow different modes in vcond
> patterns (so far on i?86/x86_64 only I think) we can vectorize more
> COND_EXPRs than before, and this patch improves it a tiny bit more
> - even i?86/x86_64 support vconds
On 6 October 2011 18:19, Jakub Jelinek wrote:
> Hi!
>
> tree-vectorizer.h already has typedefs for the recog functions,
> and using that typedef we can make these two functions slightly more
> readable.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
Thanks,
Ira
>
>
On 6 October 2011 19:28, Jakub Jelinek wrote:
> On Thu, Oct 06, 2011 at 07:27:28PM +0200, Ira Rosen wrote:
>> > + i = 1;
>> > + if ((rhs_code == COND_EXPR || rhs_code == VEC_COND_EXPR)
>>
>> I don't understand why we need VEC_COND_EX
Hi,
In vectorizer pattern recognition when a pattern def_stmt already
exists, we need to mark it properly for the current pattern.
Another problem is that we don't really have to check that TYPE_OUT is
a vector type. It is set by the pattern detection procedures, and if
the type is invalid we fail
Hi,
This patch fixes a maybe-uninitialized error.
Bootstrapped and tested on powerpc64-suse-linux.
Committed.
Ira
ChangeLog:
PR bootstrap/51112
* tree-vect-stmts.c (vectorizable_condition): Initialize comp_vectype.
Index: tree-vect-stmts.c
=
gcc-patches-ow...@gcc.gnu.org wrote on 22/11/2011 03:31:22 PM:
> From: Ramana Radhakrishnan
> > gcc/testsuite/lib/
> > * target-supports.exp (check_effective_target_vect_intfloat_cvt):
True
> > for ARM NEON.
> > (check_effective_target_vect_uintfloat_cvt): Likewise.
> > (check
Hi,
In vectorizer's over-widening pattern recognition the last statement
is expected to be a type demotion, but the check for that was
incomplete. We now check that the resulting type is not bigger than
the original type of the computation.
Bootstrapped and tested on powerpc64-suse-linux, tested
On 1 December 2011 18:41, Jakub Jelinek wrote:
> Hi!
Hi,
>
> As mentioned in the PR, vect_pattern_recog_1 attempts to find out
> if the computed type_in and type_out are already vector types or not,
> and uses VECTOR_MODE_P (TYPE_MODE (type_in)) as the test. Unfortunately,
> get_vectype_for_sca
-pattern-2b.c: Ditto.
>* gcc.dg/vect/vect-shift-2.c: Ditto.
>* gcc.dg/vect/vect-strided-a-u8-i8-gap2.c: Ditto.
>* gcc.dg/vect/vect-strided-a-u8-i8-gap7.c: Ditto.
>* gcc.dg/vect/vect-strided-u8-i8-gap2.c: Ditto.
>* gcc.dg/vect/vect-strided-u8-i8-gap4.c: Ditto.
>* gcc.dg/vect/vect-strided-u8-i8-gap7.c: Ditto.
>
> --
> ---
> Best regards,
> Michael V. Zolotukhin,
> Software Engineer
> Intel Corporation.
> [attachment "vect_tests.patch" deleted by Ira Rosen/Haifa/IBM]
(attached).
This is ok with me.
Thanks,
Ira
>
> On 2 December 2011 20:49, Ira Rosen wrote:
> >
> > gcc-patches-ow...@gcc.gnu.org wrote on 02/12/2011 06:23:25 PM:
> >
> >> Hi,
> >>
> >> This patch increases array sizes in tests from vect.exp suite, th
Hi,
This patch adds a missing exit phi node for outer loop in
vectorization of double reduction.
Bootstrapped and tested on powerpc64-suse-linux.
Committed.
Ira
ChangeLog:
PR middle-end/51285
* tree-vect-loop.c (vect_create_epilog_for_reduction): Create exit
phi nodes f
gcc-patches-ow...@gcc.gnu.org wrote on 05/12/2011 10:39:07 AM:
> From: Michael Zolotukhin
> To: Richard Guenther
> Cc: gcc-patches@gcc.gnu.org, izamya...@gmail.com
> Date: 05/12/2011 10:39 AM
> Subject: Re: [Patch] Increase array sizes in vect-tests to enable
> 256-bit vectorization
> Sent by:
On 9 December 2011 19:08, Jakub Jelinek wrote:
> Hi!
>
> As mentioned in the PR, we ICE on the following testcase, because
> there are DRs in a GIMPLE_CALL stmt and when there is just one, we
> compute vectype for the call as if it were a load or store, but during
> computation of vectorization fa
gcc-patches-ow...@gcc.gnu.org wrote on 12/12/2011 01:00:52 PM:
> I changed xfails to target-checks - for now I use common
> vect_multiple_sizes (though it'll fail when wider vectors emerge).
> Also, I changed AVX-check to the version Uros suggested. Please check
> updated patch (attached).
>
> A
Michael Zolotukhin wrote on 12/12/2011
01:57:09 PM:
>
> By the way, how could we check if '-mprefer-avx128' was specified from
> target-supports.exp?
>
If I understand your question correctly, you can use check-flags (see
check_effective_target_arm_fp16_ok_nocache for example).
> Is there any
gcc-patches-ow...@gcc.gnu.org wrote on 13/12/2011 04:05:57 AM:
> On core2, unaligned vector load/store using movdqu is a very slow
operation.
> Experiments show it is six times slower than movdqa (aligned) and this is
> irrespective of whether the resulting data happens to be aligned or not.
> F
Jakub Jelinek wrote on 14/12/2011 02:25:13 PM:
>
> @@ -1573,6 +1576,211 @@ vect_recog_vector_vector_shift_pattern (
>return pattern_stmt;
> }
>
> +/* Detect a signed division by power of two constant that wouldn't be
> + otherwise vectorized:
> +
> + type a_t, b_t;
> +
> + S1 a_t =
Jakub Jelinek wrote on 15/12/2011 09:02:57 AM:
> On Thu, Dec 15, 2011 at 08:32:26AM +0200, Ira Rosen wrote:
> > > + cond = build2 (LT_EXPR, boolean_type_node, oprnd0, build_int_cst
> > > (itype, 0));
> > > + gsi = gsi_for_stmt (last_stmt);
> >
Uros Bizjak wrote on 15/12/2011 09:56:12 AM:
> On Thu, Dec 15, 2011 at 5:16 AM, Michael Zolotukhin
> wrote:
> > Thanks!
> > Fixed patch is attached.
> >
> > Any other comments?
> >
> > Changelog:
> > 2011-12-14 Michael Zolotukhin
> >
> > * gcc.dg/vect/no-section-anchors-vect-31.c: Ad
Jakub Jelinek wrote on 15/12/2011 12:54:29 PM:
> Perhaps it would be even cleaner to get rid of the pattern stmt and def
stmt
> seq distinction and just have pattern as whole be represented as
gimple_seq,
> but perhaps that cleanup can be deferred for later.
Sounds good.
> This patch also fix
Jakub Jelinek wrote on 15/12/2011 03:51:25 PM:
> On Thu, Dec 15, 2011 at 03:35:34PM +0200, Ira Rosen wrote:
> > > This patch also fixes
> > > a problem where vect_determine_vectorization_factor would iterate the
> > same
> > > stmt twice - for some reaso
Jakub Jelinek wrote on 15/12/2011 03:51:25 PM:
> On Thu, Dec 15, 2011 at 03:35:34PM +0200, Ira Rosen wrote:
> > > This patch also fixes
> > > a problem where vect_determine_vectorization_factor would iterate the
> > same
> > > stmt twice - for some reaso
Hi,
This patch fixes an attempt to access gsi of pattern statement.
Bootstrapped and tested on ia64-unknown-linux-gnu by Uros and on
powerpc64-suse-linux by me.
Committed.
Ira
ChangeLog:
PR tree-optimization/51684
* tree-vect-slp.c (vect_schedule_slp_instance): Get gsi of ori
e.
> * gcc.dg/vect/vect-widen-mult-half-u8.c: Likewise.
> * gcc.dg/vect/vect-widen-mult-half.c: Likewise.
> * gcc.dg/vect/vect-widen-mult-sum.c: Likewise.
> * gcc.dg/vect/vect-widen-mult-u16.c: Likewise.
> * gcc.dg/vect/wrapv-vect-reduc-dot-s8b.c: Likewise.
> [attachment "51097.patch" deleted by Ira Rosen/Haifa/IBM]
Igor Zamyatin wrote on 29/12/2011 02:04:45 PM:
> When compiler configured with, say corei7-avx, it outputs twice more
> diagnostics on integer tests since AVX deals mostly with floats. I.e.
> compiler tries to vectorize on AVX vector size, than fails and then
> vectorizes on smaller vector size
tor sizes. Why doesn't the vectorizer
fail during vect_determine_vectorization_factor?
Ira
>
> I understand that all this looks quite bulky but it's hard to create
> something which looks better without loosing generality
>
>
> On Thu, Dec 29, 2011 at 4:15 PM, Ira Rosen wr
-* with
corresponding flags.
(See attached file: patch.txt)Index: ChangeLog
===
--- ChangeLog (revision 182767)
+++ ChangeLog (working copy)
@@ -1,3 +1,10 @@
+2012-01-01 Ira Rosen
+
+ PR tree-optimization/51704
Hi,
As described in PR 51269, the vectorizer adjusts number of prologue loop
iterations according to cost model, but never uses the result. This happens
because the result is not returned from the function that computes it, and
is, therefore, ignored.
Bootstrapped and tested on powerpc64-suse-li
Hi,
In over-widening pattern we expect the last statement to be a type
demotion, but don't check this properly. The patch fixes the check, and
also updates vect-widen-shift-u8.c to expect additional widening shift
pattern instead of over-widening pattern.
Bootstrapped and tested on powerpc64-sus
Jakub Jelinek wrote on 01/02/2012 06:40:13 PM:
> Hi!
>
Hi,
> vinfo_for_stmt can't be used on stmts outside of the current loop.
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Yes.
Thanks,
Ira
>
> 2012-02-01 Jakub Jelinek
>
>PR tree-optimization/52073
>*
ChangeLog
===
--- ChangeLog (revision 183901)
+++ ChangeLog (working copy)
@@ -1,3 +1,31 @@
+2012-02-05 Ira Rosen
+
+ PR tree-optimization/52091
+ * tree-vectorizer.h (vect_is_simple_use): Add an argument.
+ (vect_is_simple_use_1): Likewise.
+
Hi,
This patch adds a support of pattern recognition in basic block SLP.
Bootstrapped and tested on powerpc64-suse-linux.
Ira
ChangeLog:
* tree-vectorizer.h (vect_pattern_recog): Add new argument.
* tree-vect-loop.c (vect_analyze_loop_2): Update call to
vect_pattern_re
Hi,
This patch enhances mixed condition pattern detection to work with
non-constant integral then and else clauses. It checks that 'then' and
'else' are results of type conversion from the comparison type to their
current type, and generates the whole cond_epxr in comparison type
(ignoring the co
Hi,
I am starting to work on a new project and won't be able to continue with
vectorizer maintenance.
I'd like to thank all the people I had a chance to work with for making my
GCC experience so enjoyable.
All the best,
Ira
2012-02-08 Ira Rosen
* MAINTAINERS (Various M
> >
> > No, we shouldn't arrive with a NULL use_stmt here.
>
> I think a proper fix will be to fail if there are no uses.
> I'll prepare a patch on Sunday.
>
Here is the patch. It bails out if LHS has no uses.
Bootstrapped and tested on powerpc64-suse-linux.
Committed.
Ira
ChangeLog:
Hi,
The vectorizer supports strided loads with gaps, e.g., when only a[4i]
and a[4i+2] are accessed, it generates a vector load a[4i:4i+3], i.e.,
creating an access to a[4i+3], which doesn't exist in the scalar code.
This access maybe invalid as described in the PR.
This patch creates an epilogue
Hi,
This patch uses MAP_ANON if MAP_ANONYMOUS is not defined fixing this
test's failure on x86_64-apple-darwin10.
Tested on x86_64-suse-linux and on x86_64-apple-darwin10 (by Dominique).
OK to apply?
Thanks,
Ira
testsuite/ChangeLog:
PR testsuite/49222
* gcc.dg/vect/pr49038.c: Use MAP
-tree-scev-cprop.
Index: ChangeLog
===
--- ChangeLog (revision 174424)
+++ ChangeLog (working copy)
@@ -1,3 +1,10 @@
+2011-05-30 Ira Rosen
+
+ PR tree-optimization/49199
+ * tree-vect-loop.c (vect_is_slp_reduction
1 - 100 of 182 matches
Mail list logo
