From: Andrew Pinski
Previously the idea was gimple_simplify_phiopt would call
resimplify with a NULL sequence but that sometimes fails
even if there was only one statement produced. The cases
where it fails is when there are two simplifications happen.
In the case of the min/max production, the
From: Andrew Pinski
Instead of waiting for DCE to remove the unused statement,
and maybe optimize another conditional, it is better if
we don't move the statement and have the statement
removed.
gcc/ChangeLog:
* tree-ssa-phiopt.c (used_in_seq): New function.
From: Andrew Pinski
Just two misc improvements to simplify and match of phiopt.
I decided to submit this two before I finish up the min/max
movement to match.pd as that will take some time.
Both showed up while looking into the movement of min/max
though.
Andrew Pinski (2):
Improve early
On 7/2/2021 10:13 AM, Jeff Law wrote:
This is a minor missed optimization we found with our internal port.
Given this code:
typedef struct {short a; short b;} T;
extern void g1();
void f(T s)
{
if (s.a < 0)
g1();
}
"s" is passed in a register, but it's still a
Currently, doloop.xx variable is using the type as niter which may shorter
than word size. For some cases, it may be better to use word size type.
For example, on some 64bit system, to access 32bit niter, subreg maybe used.
Then using 64bit type would not need to use subreg if the value can be
Hi,
Ping.
— Jojo
在 2021年6月29日 +0800 PM4:11，Jojo R ，写道：
> T-HEAD extends some customized ISAs for Cores.
> The patches support ldr/str insns, it likes arm's LDR insn, the
> memory model is a base register indexed by (optionally scaled) register.
On 7/8/21 6:28 PM, Segher Boessenkool wrote:
> It needs testing on BE.
Will do.
>> +static bool consecutive_mem_locations (rtx, rtx);
>
> Please don't; just move functions to somewhere earlier than where they
> are used.
Will do.
>> @@ -16841,8 +16843,35 @@ rs6000_split_multireg_move
On 6/25/2021 8:40 AM, Richard Sandiford wrote:
Xi Ruoyao via Gcc-patches writes:
On Fri, 2021-06-25 at 01:02 +0800, Xi Ruoyao wrote:
On Thu, 2021-06-24 at 10:48 -0600, Jeff Law wrote:
I'd like to know a bit more here. mips.exp shouldn't care about the
options passed to the compiler and
On 6/30/2021 2:33 PM, Indu Bhagat via Gcc-patches wrote:
PR testsuite/101269 - new test case gcc.dg/debug/btf/btf-datasec-1.c
fails with its introduction in r12-1852
BTF datasec records for .rodata/.data are expected for now for all targets.
For powerpc based targets, use -msdata=none when
On 7/1/2021 7:02 PM, Martin Sebor via Gcc-patches wrote:
-Wvla-parameter relies on operand_equal_p() with OEP_LEXICOGRAPHIC
set to compare VLA bounds for equality. But operand_equal_p()
doesn't consider decl names, and so nontrivial expressions that
refer to the same function parameter are
On Linux/x86_64,
a110855667782dac7b674d3e328b253b3b3c919b is the first bad commit
commit a110855667782dac7b674d3e328b253b3b3c919b
Author: Martin Sebor
Date: Wed Jul 7 14:05:25 2021 -0600
Correct handling of variable offset minus constant in -Warray-bounds
[PR100137]
caused
FAIL:
Hi!
On Thu, Jul 08, 2021 at 05:01:05PM -0500, Peter Bergner wrote:
> The MMA build built-ins currently use individual lxv instructions to
> load up the registers of a __vector_pair or __vector_quad. If the
> memory addresses of the built-in operands are to adjacent locations,
> then we could use
On 7/5/2021 8:04 PM, Paul Hua wrote:
Looks good to me, but I have no right to approve.
But your opinions are well respected :-)
I'll go ahead and ACK, though in general I'm stepping away from
reviewing target specific work.
jeff
To unblock bootstrap this morning that was failing due to stricter
array bounds checking, I suppressed two -Warray-bounds instances
in cp/modules.cc without analyzing them, tracking the to-do in
pr101372. Now that I understand what's going on -- the warning
is behaving as designed, flagging
Hi Gaius,
On Thu, 8 Jul 2021, Gaius Mulley wrote:
> Here are two proposed patches to wwwdocs:
thank you for thinking of updating the web pages, too!
> diff --git a/htdocs/frontends.html b/htdocs/frontends.html
:
> http://www.nongnu.org/gm2/;>GNU Modula-2 implements
> the PIM2, PIM3, PIM4 and
The test xfailed for ILP32 has been apparently passing for some time.
I've removed the xfail after confirming in with -m32 on x86_64 and
powerpc64.
Martin
commit 68b938fada4c728c0b850b44125d9a173c01c8fb
Author: Martin Sebor
Date: Thu Jul 8 16:22:25 2021 -0600
testsuite: Remove an xfail.
Hello,
It was brought up when discussing PR debug/101283 (Several tests fail on
Darwin with -gctf/gbtf) that it will be good to provide means for targets to
opt out of CTF/BTF support.
By and large, it seems to me that CTF/BTF debug formats can be safely enabled
for all ELF-based targets by
Hello Gerald,
Here are two proposed patches to wwwdocs:
htdocs/frontends.html: Update the description of GNU Modula-2.
htdocs/git.html: Document the new devel/modula-2 branch.
regards,
Gaius
=
diff --git a/htdocs/frontends.html b/htdocs/frontends.html
index bec33b7b..60f08aa4
I have committed the attached change adjusting the expected test
output to the difference between LP64 and ILP32.
Tested in both modes on x86_64 and with a powerpc64 cross-compiler.
Martin
Adjust expected output for LP32 [PR100451].
gcc/testsuite/ChangeLog:
PR testsuite/100451
*
I have committed the attached change to ease testing with bare
bones cross-compilers with no libstdc++ headers.
Tested on x86_64 and with a powerpc64 cross-compiler.
Martin
commit c68cac900ab4ccaf6b1a31168bc9a302ebc46428
Author: Martin Sebor
Date: Thu Jul 8 16:02:01 2021 -0600
Avoid
The MMA build built-ins currently use individual lxv instructions to
load up the registers of a __vector_pair or __vector_quad. If the
memory addresses of the built-in operands are to adjacent locations,
then we could use an lxvp in some cases to load up two registers at once.
The patch below
On 7/8/21 11:28 AM, Patrick Palka wrote:
Here we're crashing ultimately because the mechanism for delaying
substitution into a requires-expression (or constexpr if) doesn't
expect to see dependent args. But we end up capturing dependent
args here when substituting into the default template
On Thu, Jul 08, 2021 at 05:34:24PM -0400, Jason Merrill wrote:
> On 7/8/21 4:26 PM, Marek Polacek wrote:
> > Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
>
> OK for trunk and 11, at least. I lean toward putting it on older release
> branches as well, but it doesn't seem urgent.
On 7/8/21 4:26 PM, Marek Polacek wrote:
On Thu, Jul 08, 2021 at 09:35:02AM -0400, Marek Polacek wrote:
On Thu, Jul 08, 2021 at 09:30:27AM -0400, Jason Merrill wrote:
On 7/7/21 9:40 PM, Marek Polacek wrote:
It sounds plausible that this assert
int f();
This patch resolves the failure of powerpc64 long double complex divide
in native ibm long double format after the patch "Practical improvement
to libgcc complex divide".
See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101104
The new code uses the following macros which are intended to be mapped
And another installment in optimizing a dead architecture. This builds
on prior patches to improve compare/test elimination for shifts.
Specifically for the older chips in the H8 family we have to handle
variable shifts with a loop.
Right now the splitter generates (set (pc) (if_then_else
(Resend this email since the previous one didn’t quote, I changed one setting
in my mail client, hopefully that can fix this issue).
Hi, Martin,
Thank you for the review and comment.
> On Jul 8, 2021, at 8:29 AM, Martin Jambor wrote:
>> diff --git a/gcc/tree-sra.c b/gcc/tree-sra.c
>> index
Hi Martin,
> Yesterday's enhancement to -Warray-bounds has exposed a couple of
> issues in libgo where the code writes into an invalid constant
> address that the warning is designed to flag.
>
> On the assumption that those invalid addresses are deliberate,
> the attached patch suppresses these
On Thu, Jul 08, 2021 at 09:35:02AM -0400, Marek Polacek wrote:
> On Thu, Jul 08, 2021 at 09:30:27AM -0400, Jason Merrill wrote:
> > On 7/7/21 9:40 PM, Marek Polacek wrote:
> > > It sounds plausible that this assert
> > >
> > >int f();
> > >static_assert(noexcept(sizeof(f(;
> > >
> >
On 7/5/2021 5:17 AM, Richard Biener via Gcc-patches wrote:
On Fri, Jul 2, 2021 at 6:13 PM Jeff Law wrote:
This is a minor missed optimization we found with our internal port.
Given this code:
typedef struct {short a; short b;} T;
extern void g1();
void f(T s)
{
if (s.a < 0)
On 7/8/21 4:41 AM, Andreas Schwab wrote:
On Jul 07 2021, Marek Polacek via Gcc-patches wrote:
On Wed, Jul 07, 2021 at 02:38:11PM -0600, Martin Sebor via Gcc-patches wrote:
I certainly will. Pushed in r12-2132.
I think this patch breaks bootstrap on x86_64:
It also breaks bootstrap on
Hi Ian,
Yesterday's enhancement to -Warray-bounds has exposed a couple of
issues in libgo where the code writes into an invalid constant
address that the warning is designed to flag.
On the assumption that those invalid addresses are deliberate,
the attached patch suppresses these instances by
I have committed the attached patch to unblock the bootstrap errors
due to the tightening up of the -Warray-bounds checking in r12-213.
I have also temporarily disabled a couple of instances of the warning
in gcc/cp/module.cc. They don't appear to be caused by the same
tighter checking but I
Richard Biener writes:
> On Thu, Jul 8, 2021 at 2:46 PM Richard Sandiford via Gcc-patches
> wrote:
>>
>> This patch passes the reduc_info to get_initial_defs_for_reduction,
>> so that the function can get general information from there rather
>> than from the first SLP statement. This isn't a
Here we're crashing ultimately because the mechanism for delaying
substitution into a requires-expression (or constexpr if) doesn't
expect to see dependent args. But we end up capturing dependent
args here when substituting into the default template argument during
coerce_template_parms for the
Hi, Martin,
Thank you for the review and comment.
On Jul 8, 2021, at 8:29 AM, Martin Jambor
mailto:mjam...@suse.cz>> wrote:
diff --git a/gcc/tree-sra.c b/gcc/tree-sra.c
index c05d22f3e8f1..35051d7c6b96 100644
--- a/gcc/tree-sra.c
+++ b/gcc/tree-sra.c
@@ -384,6 +384,13 @@ static struct
/*
Remove the Unknown_ type representation attribute predicates from
Einfo.Utils. "not Known_Alignment (...)" is at least as readable as
"Unknown_Alignment (...)" -- we don't need a bunch of functions that
just do a "not".
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
*
Prevent AST climbing from going outside of the current program unit;
tune style; add comments. Also, only set the Current_Assignment global
variable when needed and clear it once the analysis of an assignment
statement is done.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
Ada_2022 introduces the notion of Target_Name, written @, to be used in
assignment statements, where it denotes the value of the left-hand side
prior to the assignment. This patch diagnoses illegal uses of the target
name outside of its legal context.
Tested on x86_64-pc-linux-gnu, committed on
This disables the last special encoding done in Get_Encoded_Name, except
when -fgnat-encodings=all is passed on the command line.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* exp_dbug.adb (Get_Encoded_Name): Do not encode names of discrete
types with custom
Misc cleanups found while working on transient scopes.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* comperr.adb (Compiler_Abort): Call Sinput.Unlock, because if
this is called late, then Source_Dump would crash otherwise.
* debug.adb: Correct
The .inc extension isn't recognized by gprconfig. The original
motivation for using this extension was to match the convention
of putting code in .inc ala unwind.inc. However it's easier in this
situation to just rename it to a .h file.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
GNATcoverage possibly relies on the presence of the duplicate D lines in ALI
files for its Source Coverage Obligation tables among different instantiations
of a same generic. Mention this in comments.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* lib-writ.ads: Mention
In the presence of style switch -gnatyO, the compiler emits a spurious
style violation message naming an inherited operation that does not come
from an explicit subprogram declaration.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* style.adb (Missing_Overriding): Do not
Add syntax and semantic support for this new Ada 2022 feature.
Support for proper accessibility levels to be investigated in a second step.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* par-ch5.adb (P_Iterator_Specification): Add support for access
definition in
In the case of complex generic instantiations, the warning on component
not being present can be spurious (corresponding to dead code for the
given instance), so we disable it.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* sem_util.ads, sem_util.adb
The perfect hash function generated by the compiler to speed up the Value
attribute of an enumeration type contains an implicit loop and, therefore,
violates the No_Implicit_Loops restriction when it is active.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* exp_imgv.adb:
Simplify "Present (L) and then not Is_Empty_List (L)" into "not
Is_Empty_List (L)", since Is_Empty_List can be called on No_List
and returns True.
Code cleanup; semantics is unaffected.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* sem_ch12.adb, sem_ch6.adb, sem_ch9.adb,
There is only one call to Unit_Display and it is guarded by the
List_Units global variable. There is no need to retest this variable
inside the Unit_Display routine.
Code cleanup; semantics is unaffected.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* par-ch10.adb
When loading of renamed child unit failed, we didn't properly restore
the value of a global Parsing_Main_Extended_Source variable.
This is primarily a cleanup change; behaviour is not affected (perhaps
except for errors reported on complicated code that is illegal anyway).
Tested on
Code cleanup; semantics is unaffected.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* lib-writ.adb (Ensure_System_Dependency): Simplify condition.diff --git a/gcc/ada/lib-writ.adb b/gcc/ada/lib-writ.adb
--- a/gcc/ada/lib-writ.adb
+++ b/gcc/ada/lib-writ.adb
@@ -147,7 +147,7
The comment in Par.Load says "... or we are in big trouble, and abandon
the compilation", but the code merely emitted errors and kept going. Now
it emits errors, flags the problem in the unit table and gives up. Also,
it was wrong for this routine to remove the unit, because the callers
who add
Only style fixes; comments and code themselves are unchanged.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* lib-load.adb (Load_Unit): Fix style in comment.
* par-load.adb (Load): Likewise.
* scng.adb (Initialize_Scanner): Fix whitespace.diff --git
Single question marks are deprecated.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* exp_prag.adb (Expand_Pragma_Inspection_Point): Fix error
message.diff --git a/gcc/ada/exp_prag.adb b/gcc/ada/exp_prag.adb
--- a/gcc/ada/exp_prag.adb
+++ b/gcc/ada/exp_prag.adb
@@
After a syntax error, if the code is compiled with -gnatq, semantic
analysis should still proceed without internal errors if possible. Add
special case to recognize ill-formed array type.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* layout.adb (Layout_Type): Do not call
Before this patch, the following program would make GNAT crash:
procedure P is
Unused_Var : Integer with Shared => False;
pragma Inspection_Point;
begin
null;
end tmp;
This was because the Shared aspect resulted in a freeze node being
inserted after the Inspection_Point pragma. This
In case of compilation error, the low and high bounds of the array
type might have been replaced by an error node. Deal with this case
by checking that the bounds are known at compile time.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* layout.adb (Layout_Type): Add guard
To help experiment with this new model.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* libgnat/a-cbdlli.adb, libgnat/a-cbhama.adb,
libgnat/a-cbhase.adb, libgnat/a-cbmutr.adb,
libgnat/a-cborma.adb, libgnat/a-cborse.adb,
libgnat/a-cobove.adb,
For a packed constrained array type with a Component_Size clause, it
may be possible to compute both its RM_Size and Esize. Do this as it
benefits GNATprove for checking validity of overlays.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* layout.adb (Layout_Type): Special
If an exception declaration occurs in a nonstatic scope (for example,
within the body of a task type),
System.Exception_Table.Register_Exception is to be called the first (and
*only* the first) time the declaration is elaborated. A library-level
"this exception has been registered" Boolean flag
Make it explicit that tasking is not used in the compiler, which also
allows generating simpler and more efficient code.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* gcc-interface/system.ads: Add No_Tasking restriction.diff --git a/gcc/ada/gcc-interface/system.ads
Cursors for Hashed maps and hashed sets include a component that speeds
up iteration over these containers. However, in the presence of multiple
insertions into the corresponding hash-tables, this component may become
unreliable when a cursor obtained before an iteration is compared with a
cursor
As part of experimenting with No_Dynamic_Accessibility_Checks, it seems
that reverting the meaning of -gnatd_b is a better default for this
experiment.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* debug.adb, sem_util.adb: Revert meaning of -gnatd_b.
*
To help experiment with this new model.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* make.adb, osint.adb: Make code compatible with
No_Dynamic_Accessibility_Checks restriction.diff --git a/gcc/ada/make.adb b/gcc/ada/make.adb
--- a/gcc/ada/make.adb
+++
Replace a linear search with a hash table query.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* lib-writ.adb (Ensure_System_Dependency): Replace search in
Lib.Units with a search in Lib.Unit_Names.diff --git a/gcc/ada/lib-writ.adb b/gcc/ada/lib-writ.adb
---
Code cleanup; semantics is unaffected.
Tested on x86_64-pc-linux-gnu, committed on trunk
gcc/ada/
* sinput-l.adb (Load_File): Simplify foreword manipulation with
concatenation; similar for filename with preprocessed output.diff --git a/gcc/ada/sinput-l.adb b/gcc/ada/sinput-l.adb
On Thu, 8 Jul 2021 at 12:42, Andreas Schwab wrote:
>
> On Jul 07 2021, Marek Polacek via Gcc-patches wrote:
>
> > On Wed, Jul 07, 2021 at 02:38:11PM -0600, Martin Sebor via Gcc-patches
> > wrote:
> >> I certainly will. Pushed in r12-2132.
> >
> > I think this patch breaks bootstrap on x86_64:
>
On Thu, Jul 08, 2021 at 09:30:27AM -0400, Jason Merrill wrote:
> On 7/7/21 9:40 PM, Marek Polacek wrote:
> > It sounds plausible that this assert
> >
> >int f();
> >static_assert(noexcept(sizeof(f(;
> >
> > should pass: sizeof produces a std::size_t and its operand is not
> >
On 7/7/21 9:40 PM, Marek Polacek wrote:
It sounds plausible that this assert
int f();
static_assert(noexcept(sizeof(f(;
should pass: sizeof produces a std::size_t and its operand is not
evaluated, so it can't throw. noexcept should only evaluate to
false for potentially evaluated
Hi,
On Wed, Jul 07 2021, Qing Zhao via Gcc-patches wrote:
> Hi,
>
> This is the 4th version of the patch for the new security feature for GCC.
I have been following the threads about this feature only very lightly,
so please accept my apologies if my comments are about something which
has been
Hi,
> 2021-06-16 Martin Jambor
>
> PR ipa/101066
> * ipa-sra.c (class isra_call_summary): New member
> m_before_any_store, initialize it in the constructor.
> (isra_call_summary::dump): Dump the new field.
> (ipa_sra_call_summaries::duplicate): Copy it.
>
On Thu, Jul 8, 2021 at 2:49 PM Richard Sandiford via Gcc-patches
wrote:
>
> After previous patches, we can now easily provide the neutral op
> as an argument to get_initial_def_for_reduction. This in turn
> allows the adjustment calculation to be moved outside of
> get_initial_def_for_reduction,
On Thu, Jul 8, 2021 at 2:48 PM Richard Sandiford via Gcc-patches
wrote:
>
> This patch generalises the interface to neutral_op_for_slp_reduction
> so that it can be used for non-SLP reductions too. This isn't much
> of a win on its own, but it helps later patches.
I guess that makes sense - OK.
Richard Biener writes:
> On Thu, Jul 8, 2021 at 2:45 PM Richard Sandiford via Gcc-patches
> wrote:
>>
>> This patch adds a helper function called vect_phi_initial_value
>> for returning the incoming value of a given loop phi. The main
>> reason for adding it is to ensure that the right
On Thu, Jul 8, 2021 at 2:46 PM Richard Sandiford via Gcc-patches
wrote:
>
> This patch passes the reduc_info to get_initial_defs_for_reduction,
> so that the function can get general information from there rather
> than from the first SLP statement. This isn't a win on its own,
> but it becomes
> -Original Message-
> From: Gcc-patches bounces+kyrylo.tkachov=arm@gcc.gnu.org> On Behalf Of Christophe
> Lyon via Gcc-patches
> Sent: 07 July 2021 13:24
> To: Christophe LYON
> Cc: gcc Patches
> Subject: Re: [PATCH] testsuite: Add arm_arch_v7a_ok effective-target to
> pr57351.c
On Thu, Jul 8, 2021 at 2:45 PM Richard Sandiford via Gcc-patches
wrote:
>
> This patch adds a helper function called vect_phi_initial_value
> for returning the incoming value of a given loop phi. The main
> reason for adding it is to ensure that the right preheader edge
> is used when
On Thu, Jul 8, 2021 at 2:44 PM Richard Sandiford via Gcc-patches
wrote:
>
> Vector reduction accumulators can differ in signedness from the
> final scalar result. The conversions to handle that case were
> distributed through vect_create_epilog_for_reduction; this patch
> does the conversion
On Thu, Jul 8, 2021 at 2:43 PM Richard Sandiford via Gcc-patches
wrote:
>
> vect_create_epilog_for_reduction had a variable called new_phis.
> It collected the statements that produce the exit block definitions
> of the vector reduction accumulators. Although those statements
> are indeed phis
On Thu, Jul 8, 2021 at 2:42 PM Richard Sandiford via Gcc-patches
wrote:
>
> This patch constructs an array_slice of the scalar statements that
> produce live-out reduction results in the original unvectorised loop.
> There are three cases:
>
> - SLP reduction chains: the final SLP stmt is
On Thu, Jul 8, 2021 at 2:41 PM Richard Sandiford via Gcc-patches
wrote:
>
> vect_create_epilog_for_reduction only handles two cases: single-loop
> reductions and double reductions. “nested cycles” (i.e. reductions
> in the inner loop when vectorising an outer loop) are handled elsewhere
> and
On Thu, Jul 8, 2021 at 10:25 AM Roger Sayle wrote:
>
>
> This patch tweaks the way GCC handles 32-bit integer division on
> x86_64, when the numerator is constant. Currently the function
>
> int foo (int x) {
> return 100/x;
> }
>
> generates the code:
> foo:movl$100, %eax
>
> This patch has been tested on x86_64-pc-linux-gnu with a "make
> bootstrap" and "make -k check" with no new failures. This should
> be relatively safe, as there are no changes in behaviour unless
> the user explicitly specifies -fnon-call-exceptions, when the C
> compiler then behaves more like
This patch adds support for reusing a main loop's reduction accumulator
in an epilogue loop. This in turn lets the loops share a single piece
of vector->scalar reduction code.
The patch has the following restrictions:
(1) The epilogue reduction can only operate on a single vector
(e.g.
After previous patches, we can now easily provide the neutral op
as an argument to get_initial_def_for_reduction. This in turn
allows the adjustment calculation to be moved outside of
get_initial_def_for_reduction, which is the main motivation
of the patch.
gcc/
* tree-vect-loop.c
This patch generalises the interface to neutral_op_for_slp_reduction
so that it can be used for non-SLP reductions too. This isn't much
of a win on its own, but it helps later patches.
gcc/
* tree-vect-loop.c (neutral_op_for_slp_reduction): Replace with...
Similarly to the previous patch, this one passes the reduc_info
to get_initial_def_for_reduction, rather than a stmt_vec_info that
lacks the metadata. This again becomes useful later.
gcc/
* tree-vect-loop.c (get_initial_def_for_reduction): Take the
reduc_info instead of the
This patch passes the reduc_info to get_initial_defs_for_reduction,
so that the function can get general information from there rather
than from the first SLP statement. This isn't a win on its own,
but it becomes important with later patches.
gcc/
* tree-vect-loop.c
This patch adds a helper function called vect_phi_initial_value
for returning the incoming value of a given loop phi. The main
reason for adding it is to ensure that the right preheader edge
is used when vectorising nested loops. (PHI_ARG_DEF_FROM_EDGE
itself doesn't assert that the given edge
Vector reduction accumulators can differ in signedness from the
final scalar result. The conversions to handle that case were
distributed through vect_create_epilog_for_reduction; this patch
does the conversion up-front instead.
gcc/
* tree-vect-loop.c (vect_create_epilog_for_reduction):
vect_create_epilog_for_reduction had a variable called new_phis.
It collected the statements that produce the exit block definitions
of the vector reduction accumulators. Although those statements
are indeed phis initially, they are often replaced with normal
statements later, leading to puzzling
This patch constructs an array_slice of the scalar statements that
produce live-out reduction results in the original unvectorised loop.
There are three cases:
- SLP reduction chains: the final SLP stmt is live-out
- full SLP reductions: all SLP stmts are live-out
- non-SLP reductions: the single
vect_create_epilog_for_reduction only handles two cases: single-loop
reductions and double reductions. “nested cycles” (i.e. reductions
in the inner loop when vectorising an outer loop) are handled elsewhere
and don't need a vector->scalar reduction.
The function had variables called
Quoting from the final patch in the series:
This patch adds support for reusing a main loop's reduction accumulator
in an epilogue loop. This in turn lets the loops share a single piece
of vector->scalar reduction code.
Richard Biener via Gcc-patches writes:
> On Thu, Jul 8, 2021 at 1:52 PM Richard Sandiford via Gcc-patches
> wrote:
>>
>> match.pd has a rule to simplify an extension, operation and truncation
>> back to the original type:
>>
>> (simplify
>>(convert (op:s@0 (convert1?@3 @1) (convert2?@4
On Thu, Jul 8, 2021 at 2:04 PM Richard Sandiford via Gcc-patches
wrote:
>
> -msve-vector-bits=128 causes the AArch64 port to list 128-bit Advanced
> SIMD as the first-choice mode for vectorisation, with SVE being used for
> things that Advanced SIMD can't handle as easily. However, ifcvt would
>
On Thu, Jul 8, 2021 at 1:52 PM Richard Sandiford via Gcc-patches
wrote:
>
> match.pd has a rule to simplify an extension, operation and truncation
> back to the original type:
>
> (simplify
>(convert (op:s@0 (convert1?@3 @1) (convert2?@4 @2)))
>
> Currently it handles cases in which @2 is an
-msve-vector-bits=128 causes the AArch64 port to list 128-bit Advanced
SIMD as the first-choice mode for vectorisation, with SVE being used for
things that Advanced SIMD can't handle as easily. However, ifcvt would
not then try to use SVE's predicated FP arithmetic, leading to tests
like TSVC
vectorizable_reduction had code guarded by:
if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def
|| STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def)
But that's always true after:
if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def
&& STMT_VINFO_DEF_TYPE
match.pd has a rule to simplify an extension, operation and truncation
back to the original type:
(simplify
(convert (op:s@0 (convert1?@3 @1) (convert2?@4 @2)))
Currently it handles cases in which @2 is an INTEGER_CST, but it
also works for POLY_INT_CSTs.[*]
For INTEGER_CST it doesn't
1 - 100 of 110 matches
Mail list logo
