@@ -204,6 +205,10 @@ static cl::opt
EnableLoopInterchange("enable-loopinterchange", cl::init(false),
cl::Hidden,
cl::desc("Enable the LoopInterchange Pass"));
+static cl::opt EnableLoopFusion("enable-loopfusion", cl::init(false),
kasuga-fj wrote:
I’ll leave the revert up to the author.
https://github.com/llvm/llvm-project/pull/140182
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
kasuga-fj wrote:
> To be honest, I am very surprised that this PR was submitted, let alone
> landed. We have **explicitly not approved** enabling LoopInterchange by
> default in LLVM due to outstanding issues. So instead you go ahead and enable
> it directly in Flang instead, without even noti
kasuga-fj wrote:
> This is wrong, we fixed all outstanding bugs reported against DA at the time
> when we enabled interchange in flang.
I’m referring to the bugs that existed at that time but were not reported. The
issues that were raised do not represent all the defects.
> If we don't enable
kasuga-fj wrote:
I'm not trying to emphasize the number of issues raised on GitHub, the real
concern is their substance. What I pointed out in the issues represents only a
small portion of the overall problem. There are fundamental flaws in DA's
implementation, particularly in how it handles w
kasuga-fj wrote:
It is only enabled in flang, and not in clang at the moment.
https://github.com/llvm/llvm-project/pull/140182
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
kasuga-fj wrote:
> Flang's policy is not really different from Clang's, but I can raise it in
> the next community call
Thanks. I think this should be removed from LLVM 21.
https://github.com/llvm/llvm-project/pull/140182
___
cfe-commits mailing list
kasuga-fj wrote:
Just in case: it's not "just after LoopInterchange", it's "before
LoopDeletion". If you want to run it just after LoopInterchange, please add
LoopFuse after adding `LPM` to `OptimizePM`. It might be better adding some
test to check the position.
Also, you may want to avoid re
@@ -204,6 +205,10 @@ static cl::opt
EnableLoopInterchange("enable-loopinterchange", cl::init(false),
cl::Hidden,
cl::desc("Enable the LoopInterchange Pass"));
+static cl::opt EnableLoopFusion("enable-loopfusion", cl::init(false),
kasuga-fj wrote:
> In that case. it may be better to move it to
> `buildModuleOptimizationPipeline`. This is undoubtedly an optimization, and a
> non-trivial one at that. There are still open questions about the exact
> placement of the pass within the optimization pipeline, but moving it into
@@ -204,6 +205,10 @@ static cl::opt
EnableLoopInterchange("enable-loopinterchange", cl::init(false),
cl::Hidden,
cl::desc("Enable the LoopInterchange Pass"));
+static cl::opt EnableLoopFusion("enable-loopfusion", cl::init(false),
+
@@ -204,6 +205,10 @@ static cl::opt
EnableLoopInterchange("enable-loopinterchange", cl::init(false),
cl::Hidden,
cl::desc("Enable the LoopInterchange Pass"));
+static cl::opt EnableLoopFusion("enable-loopfusion", cl::init(false),
kasuga-fj wrote:
> What do you think is the right place in pass pipeline?
I don't know, but just a gut feeling, somewhere in
`buildModuleOptimizationPipeline` looks reasonable to me rather than
`buildFunctionSimplificationPipeline`. However, again, it should be determined
by what you'd like t
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/142686
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
@@ -518,6 +524,9 @@
PassBuilder::buildO1FunctionSimplificationPipeline(OptimizationLevel Level,
invokeLoopOptimizerEndEPCallbacks(LPM2, Level);
+ if (PTO.LoopFuse)
+FPM.addPass(LoopFusePass());
+
kasuga-fj wrote:
> LoopFusion is a function pass
I do
@@ -32,6 +32,9 @@ page](https://llvm.org/releases/).
## New Compiler Flags
+* -floop-interchange is now recognized by flang.
kasuga-fj wrote:
Seems a mistake while resolving the conflict.
https://github.com/llvm/llvm-project/pull/142686
https://github.com/kasuga-fj commented:
FWIW, using merge instead of rebase might be a better. That way, you might
avoid resolving the same conflicts repeatedly.
https://github.com/llvm/llvm-project/pull/142686
___
cfe-commits mailing list
cfe-commits
kasuga-fj wrote:
You probably made a mistake with the rebase...
https://github.com/llvm/llvm-project/pull/142686
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
@@ -3157,7 +3157,7 @@ void tools::handleVectorizeSLPArgs(const ArgList &Args,
void tools::handleInterchangeLoopsArgs(const ArgList &Args,
ArgStringList &CmdArgs) {
- // FIXME: instead of relying on shouldEnableVectorizerAtOLevel, we may
@@ -518,6 +524,9 @@
PassBuilder::buildO1FunctionSimplificationPipeline(OptimizationLevel Level,
invokeLoopOptimizerEndEPCallbacks(LPM2, Level);
+ if (PTO.LoopFuse)
+FPM.addPass(LoopFusePass());
+
kasuga-fj wrote:
Is this an appropriate place? I think
@@ -2112,7 +2124,6 @@ PassBuilder::buildLTODefaultPipeline(OptimizationLevel
Level,
LPM.addPass(LoopFlattenPass());
LPM.addPass(IndVarSimplifyPass());
LPM.addPass(LoopDeletionPass());
- // FIXME: Add loop interchange.
kasuga-fj wrote:
I don't think t
kasuga-fj wrote:
So, why wouldn't adding a `cl::opt` flag work in the end? In fact, you've added
`enable-loopfuse` flag in this patch. If it's enough, I personally think it
might be better to avoid introducing additional flags unnecessarily.
https://github.com/llvm/llvm-project/pull/142686
___
kasuga-fj wrote:
Thanks for the reviews!
https://github.com/llvm/llvm-project/pull/135163
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj closed
https://github.com/llvm/llvm-project/pull/135163
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/135163
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
@@ -643,9 +650,7 @@ void LoopInfoStack::push(BasicBlock *Header,
clang::ASTContext &Ctx,
case LoopHintAttr::Disable:
switch (Option) {
case LoopHintAttr::Vectorize:
-// Disable vectorization by specifying a width of 1.
-setVectorizeWidth(1);
-
@@ -5237,8 +5237,8 @@ width/count of the set of target architectures supported
by your application.
...
}
-Specifying a width/count of 1 disables the optimization, and is equivalent to
-``vectorize(disable)`` or ``interleave(disable)``.
+Specifying a *non-scalable* widt
https://github.com/kasuga-fj updated
https://github.com/llvm/llvm-project/pull/135163
>From 79186af3df57357d40ec3b3e4eaf9da978b31ef6 Mon Sep 17 00:00:00 2001
From: Ryotaro Kasuga
Date: Fri, 4 Apr 2025 12:32:57 +
Subject: [PATCH 1/8] [clang][CodeGen] Fix metadata when vectorization is
disab
https://github.com/kasuga-fj updated
https://github.com/llvm/llvm-project/pull/135163
>From 79186af3df57357d40ec3b3e4eaf9da978b31ef6 Mon Sep 17 00:00:00 2001
From: Ryotaro Kasuga
Date: Fri, 4 Apr 2025 12:32:57 +
Subject: [PATCH 1/7] [clang][CodeGen] Fix metadata when vectorization is
disab
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/135163
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
@@ -643,9 +650,7 @@ void LoopInfoStack::push(BasicBlock *Header,
clang::ASTContext &Ctx,
case LoopHintAttr::Disable:
switch (Option) {
case LoopHintAttr::Vectorize:
-// Disable vectorization by specifying a width of 1.
-setVectorizeWidth(1);
-
@@ -5237,8 +5237,8 @@ width/count of the set of target architectures supported
by your application.
...
}
-Specifying a width/count of 1 disables the optimization, and is equivalent to
-``vectorize(disable)`` or ``interleave(disable)``.
kasuga-fj wrote:
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/135163
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
kasuga-fj wrote:
I somehow missed this until now, but it turns out that specifying
`vectorize(disable) vectorize_width(scalable)` causes an error due to
"incompatible directives" (godbolt: https://godbolt.org/z/vv1vKqdTn). I had
been concerned that the ambiguity here might make implementation
kasuga-fj wrote:
As far as I can tell, it doesn't seem like a link to UsersManual is there, but
you can reach both from [the top page](https://clang.llvm.org/docs/index.html).
https://github.com/llvm/llvm-project/pull/147014
___
cfe-commits mailing li
https://github.com/kasuga-fj commented:
It is already documented in [Clang command line argument
reference](https://clang.llvm.org/docs/ClangCommandLineReference.html#cmdoption-clang-floop-interchange)
(perhaps this is automatically generated based on
[`clang/Driver/Options.td`](https://github
kasuga-fj wrote:
Gentle ping
https://github.com/llvm/llvm-project/pull/135163
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
kasuga-fj wrote:
There is still a correctness issue with LoopInterchange, as I reported in
#140238. This problem is not specific to C/C++. The following code demonstrates
a case where illegal loop-interchange occurs.
```fortran
program main
implicit none
real, save :: A(5, 5)
real,
kasuga-fj wrote:
Ping.
I think there are two possible interpretations of `vectorize(disable)
vectorize_width(scalable)`, as follows:
- It is equivalent to `vectorize(disable)` so vectorization is disabled, as the
document says:
> Pragmas setting transformation options imply the transformati
kasuga-fj wrote:
Printing an error message looks reasonable to me, but I think we first need to
clarify what codes are "unsupported".
https://github.com/llvm/llvm-project/pull/131985
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://list
kasuga-fj wrote:
> For this reason OpenMP 5.1 added new pragma:
>
> ```c
> #pragma omp simd reduction(+:v0)
> #pragma omp unroll partial(4)
> ```
FWIW, I tried it and the following error occurred.
```
:10:1: error: OpenMP constructs may not be nested inside a simd region
except for ordered si
kasuga-fj wrote:
Thanks for sharing. I haven't looked at the details yet, but I simplified it as
follows.
```c
#include
void g(float);
void f(int m) {
#pragma omp parallel
{
float v0 = 0.0;
#pragma omp simd reduction(+:v0)
#pragma unroll(4)
for(int i=0; i < m; i++)
{
v0 +
@@ -5237,8 +5237,8 @@ width/count of the set of target architectures supported
by your application.
...
}
-Specifying a width/count of 1 disables the optimization, and is equivalent to
-``vectorize(disable)`` or ``interleave(disable)``.
kasuga-fj wrote:
kasuga-fj wrote:
> Unfortuantely IIRC there is a semantic difference between pragma and metadata:
>
>- llvm.loop.vectorize.enable==false disables vectorization AND
> interleaving (by switching off the entire LoopVectorize pass which does both
> of thise things)
>- llvm.loop.vectorize.wi
@@ -5237,8 +5237,8 @@ width/count of the set of target architectures supported
by your application.
...
}
-Specifying a width/count of 1 disables the optimization, and is equivalent to
-``vectorize(disable)`` or ``interleave(disable)``.
+Specifying a *non-scalable* widt
@@ -37,8 +37,8 @@ void test3(int *List, int Length) {
List[i] = i * 2;
}
-// Check that disabling vectorization means a vectorization width of 1, and
-// also that vectorization_predicate isn't enabled.
+// Check that vectorize is disabled, and also that vectorization_pred
https://github.com/kasuga-fj updated
https://github.com/llvm/llvm-project/pull/135163
>From 79186af3df57357d40ec3b3e4eaf9da978b31ef6 Mon Sep 17 00:00:00 2001
From: Ryotaro Kasuga
Date: Fri, 4 Apr 2025 12:32:57 +
Subject: [PATCH 1/5] [clang][CodeGen] Fix metadata when vectorization is
disab
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/135163
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
@@ -194,7 +194,7 @@ void for_test_scalable(int *List, int Length) {
}
}
-// Verify for loop is performing scalable vectorization
+// Verify for loop is NOT performing vectorization because the width is 1
kasuga-fj wrote:
Fixed both code and docs. Thanks.
h
https://github.com/kasuga-fj updated
https://github.com/llvm/llvm-project/pull/135163
>From 79186af3df57357d40ec3b3e4eaf9da978b31ef6 Mon Sep 17 00:00:00 2001
From: Ryotaro Kasuga
Date: Fri, 4 Apr 2025 12:32:57 +
Subject: [PATCH 1/4] [clang][CodeGen] Fix metadata when vectorization is
disab
@@ -194,7 +194,7 @@ void for_test_scalable(int *List, int Length) {
}
}
-// Verify for loop is performing scalable vectorization
+// Verify for loop is NOT performing vectorization because the width is 1
kasuga-fj wrote:
You are right, I misunderstood the c
https://github.com/kasuga-fj closed
https://github.com/llvm/llvm-project/pull/133707
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/131985
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/133707
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj created
https://github.com/llvm/llvm-project/pull/133707
pragma-loop.cpp contains tests for loop metadata generated via pragma
directives. These tests were not working as (perhaps) expected. This is because
the regex `.*` consumes multiple elements in metadata. For
@@ -44,23 +44,19 @@ MDNode *LoopInfo::createPipeliningMetadata(const
LoopAttributes &Attrs,
else if (Attrs.PipelineInitiationInterval != 0)
Enabled = true;
+ SmallVector Args;
+ Args.append(LoopProperties.begin(), LoopProperties.end());
+
if (Enabled != true) {
-
@@ -44,23 +44,19 @@ MDNode *LoopInfo::createPipeliningMetadata(const
LoopAttributes &Attrs,
else if (Attrs.PipelineInitiationInterval != 0)
Enabled = true;
+ SmallVector Args;
+ Args.append(LoopProperties.begin(), LoopProperties.end());
+
if (Enabled != true) {
-
@@ -44,23 +44,19 @@ MDNode *LoopInfo::createPipeliningMetadata(const
LoopAttributes &Attrs,
else if (Attrs.PipelineInitiationInterval != 0)
Enabled = true;
+ SmallVector Args;
+ Args.append(LoopProperties.begin(), LoopProperties.end());
+
if (Enabled != true) {
-
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/131985
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj closed
https://github.com/llvm/llvm-project/pull/131985
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj updated
https://github.com/llvm/llvm-project/pull/131985
>From 63dec28732e6a90f5f3449441fa8472b5b60a9f4 Mon Sep 17 00:00:00 2001
From: Ryotaro Kasuga
Date: Wed, 26 Mar 2025 07:29:18 +
Subject: [PATCH] [clang][CodeGen] Generate follow-up metadata for loops in
co
https://github.com/kasuga-fj commented:
Thanks for your review!
https://github.com/llvm/llvm-project/pull/131985
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
@@ -44,23 +44,19 @@ MDNode *LoopInfo::createPipeliningMetadata(const
LoopAttributes &Attrs,
else if (Attrs.PipelineInitiationInterval != 0)
Enabled = true;
+ SmallVector Args;
+ Args.append(LoopProperties.begin(), LoopProperties.end());
+
if (Enabled != true) {
-
@@ -44,23 +44,19 @@ MDNode *LoopInfo::createPipeliningMetadata(const
LoopAttributes &Attrs,
else if (Attrs.PipelineInitiationInterval != 0)
Enabled = true;
+ SmallVector Args;
+ Args.append(LoopProperties.begin(), LoopProperties.end());
+
if (Enabled != true) {
-
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/131985
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
@@ -227,36 +226,26 @@ void for_test_scalable_1(int *List, int Length) {
// CHECK: ![[LOOP_5]] = distinct !{![[LOOP_5]], ![[UNROLL_DISABLE:.*]],
![[DISTRIBUTE_DISABLE:.*]], ![[WIDTH_1:.*]]}
// CHECK: ![[WIDTH_1]] = !{!"llvm.loop.vectorize.width", i32 1}
-// CHECK: ![[LOOP_6]]
@@ -170,11 +161,10 @@ LoopInfo::createUnrollAndJamMetadata(const LoopAttributes
&Attrs,
MDNode::get(Ctx, MDString::get(Ctx,
"llvm.loop.unroll_and_jam.disable")));
bool FollowupHasTransforms = false;
- MDNode *Followup = createPartialUnrollMetadata(Attrs, FollowupLoo
@@ -44,23 +44,19 @@ MDNode *LoopInfo::createPipeliningMetadata(const
LoopAttributes &Attrs,
else if (Attrs.PipelineInitiationInterval != 0)
Enabled = true;
+ SmallVector Args;
+ Args.append(LoopProperties.begin(), LoopProperties.end());
+
if (Enabled != true) {
-
https://github.com/kasuga-fj commented:
> Possibly the `create*Metadata` should only return a list of properties
> instead of an MDNode, with the MDNode only to be created when applying the
> properties to a Loop. However, the easier fix would probably be to extract
> the properties from the
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/131985
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj updated
https://github.com/llvm/llvm-project/pull/131985
>From 889f40c5570af8a02e301c2bf3c6382f69210140 Mon Sep 17 00:00:00 2001
From: Ryotaro Kasuga
Date: Mon, 17 Mar 2025 11:24:47 +
Subject: [PATCH 1/3] [LoopUtils] Fix metadata generated by makeFollowupLoopID
kasuga-fj wrote:
Thanks for the comments. I'll try to work on it in the near future.
https://github.com/llvm/llvm-project/pull/125830
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj approved this pull request.
https://github.com/llvm/llvm-project/pull/125830
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj edited
https://github.com/llvm/llvm-project/pull/112501
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/kasuga-fj created
https://github.com/llvm/llvm-project/pull/112501
Previously `#pragma clang loop pipeline` only accepted `disable`. This patch
adds `enable` as a valid argument for this pragma. This allows Software
Pipelining optimization to be applied to some loops instead
75 matches
Mail list logo
