[clang][CodeGen] Fix metadata when vectorization is disabled by pragma #135163

kasuga-fj · 2025-04-10T11:35:06Z

There are two ways to disable vectorization with pragma, vectorize(disable) and vectorize_width(1). The document (https://clang.llvm.org/docs/LanguageExtensions.html#vectorization-interleaving-and-predication) states:

Specifying a width/count of 1 disables the optimization, and is equivalent to vectorize(disable) or interleave(disable).

So the behavior should be the same when using either is used (under the assumption that scalable is not explicitly specified). However, the current implementation doesn't satisfy this. When vectorize(disable) is specified, it is converted internally to the same as vectorize_width(1), and the metadata is generated as if vectorization is not explicitly specified as enabled or disabled. This can cause a problem when other transformations are also specified by pragma. For example, if unroll_count(8) is specified, the loop should have a metadata that contains the information directly, like:

!loop0 = !{!"loop0", !vectorize_disable, !unroll}
!vectorize_disable = {...}
!unroll = !{!"llvm.loop.unroll_count", i32 8}

But now it's attached into followup metadata for vectorization.

!loop0 = !{!"loop0", !vectorize_width, !followup}
!vectorize_width = !{!"llvm.loop.vectorize.width", i32 1}
!followup = !{!"llvm.loop.vectorize.followup_all", !unroll}
!unroll = !{!"llvm.loop.unroll_count", i32 8}

As a result, the unroll attributes are ignored because the vectorization is not applied.

This patch fixes this issue by generating metadata that disables vectorization when vectorize(disable), vectorize_width(1) or vectorize_width(1, fixed) are specified. Note that if scalable is explicitly enabled (like vectorize_width(1, scalable)), vectorization is still processed.

The document also says:

If the transformation is disabled (e.g. vectorize(disable)), that takes precedence over transformations option pragmas implying that transformation.

Therefore, if vectorization is disabled, all other vectorize options like vectorize_predicate should be ignored. This patch also omits to generate attributes for vectorization when it is disabled.

Fix #75839

There are two ways to disable vectorization with pragma, `vectorize(disable)` and `vectorize_width(1)`. The document (https://clang.llvm.org/docs/LanguageExtensions.html#vectorization-interleaving-and-predication) states: > Specifying a width/count of 1 disables the optimization, and is equivalent to `vectorize(disable)` or `interleave(disable)`. So the behavior should be the same when using either is used. However, the current implementation doesn't satisfy this. When `vectorize(disable)` is specified, it is converted internally to the same as `vectorize_width(1)`, and the metadata is generated as if vectorization is not explicitly specified as enabled or disabled. This can cause a problem when other transformations are also specified by pragma. For example, if `unroll_count(8)` is specified, the loop should have a metadata that contains the information directly, like: ``` !loop0 = !{!"loop0", !vectorize_disable, !unroll} !vectorize_disable = {...} !unroll = !{!"llvm.loop.unroll_count", i32 8} ``` But now it's attached into followup metadata for vectorization. ``` !loop0 = !{!"loop0", !vectorize_width, !followup} !vectorize_width = !{!"llvm.loop.vectorize.width", i32 1} !followup = !{!"llvm.loop.vectorize.followup_all", !unroll} !unroll = !{!"llvm.loop.unroll_count", i32 8} ``` As a result, the unroll attributes are ignored because the vectorization is not applied. This patch fixes this issue by generating metadata that disables vectorization when `vectorize(disable)` or `vectorize_width(1)` are specified. The document also says: > If the transformation is disabled (e.g. vectorize(disable)), that takes precedence over transformations option pragmas implying that transformation. Therefore, if vectorization is disabled, all other vectorize options like `vectorize_predicate` should be ignored. This patch also omits to generate attributes for vectorization when it is disabled. Fix llvm#75839

llvmbot · 2025-04-10T11:35:38Z

@llvm/pr-subscribers-clang-codegen

@llvm/pr-subscribers-clang

Author: Ryotaro Kasuga (kasuga-fj)

Changes

There are two ways to disable vectorization with pragma, vectorize(disable) and vectorize_width(1). The document (https://clang.llvm.org/docs/LanguageExtensions.html#vectorization-interleaving-and-predication) states:

> Specifying a width/count of 1 disables the optimization, and is equivalent to vectorize(disable) or interleave(disable).

So the behavior should be the same when using either is used. However, the current implementation doesn't satisfy this. When vectorize(disable) is specified, it is converted internally to the same as vectorize_width(1), and the metadata is generated as if vectorization is not explicitly specified as enabled or disabled. This can cause a problem when other transformations are also specified by pragma. For example, if unroll_count(8) is specified, the loop should have a metadata that contains the information directly, like:

!loop0 = !{!"loop0", !vectorize_disable, !unroll}
!vectorize_disable = {...}
!unroll = !{!"llvm.loop.unroll_count", i32 8}

But now it's attached into followup metadata for vectorization.

!loop0 = !{!"loop0", !vectorize_width, !followup}
!vectorize_width = !{!"llvm.loop.vectorize.width", i32 1}
!followup = !{!"llvm.loop.vectorize.followup_all", !unroll}
!unroll = !{!"llvm.loop.unroll_count", i32 8}

As a result, the unroll attributes are ignored because the vectorization is not applied.

This patch fixes this issue by generating metadata that disables vectorization when vectorize(disable) or vectorize_width(1) are specified.

The document also says:

> If the transformation is disabled (e.g. vectorize(disable)), that
takes precedence over transformations option pragmas implying that transformation.

Therefore, if vectorization is disabled, all other vectorize options like vectorize_predicate should be ignored. This patch also omits to generate attributes for vectorization when it is disabled.

Fix #75839

Full diff: https://github.com/llvm/llvm-project/pull/135163.diff

5 Files Affected:

(modified) clang/lib/CodeGen/CGLoopInfo.cpp (+3-4)
(modified) clang/test/CodeGenCXX/pragma-loop-pr27643.cpp (+6-6)
(modified) clang/test/CodeGenCXX/pragma-loop-predicate.cpp (+11-11)
(modified) clang/test/CodeGenCXX/pragma-loop-safety.cpp (+2-2)
(modified) clang/test/CodeGenCXX/pragma-loop.cpp (+25-4)

diff --git a/clang/lib/CodeGen/CGLoopInfo.cpp b/clang/lib/CodeGen/CGLoopInfo.cpp
index 2b7d7881ab990..e9d622d78bf6d 100644
--- a/clang/lib/CodeGen/CGLoopInfo.cpp
+++ b/clang/lib/CodeGen/CGLoopInfo.cpp
@@ -200,7 +200,8 @@ LoopInfo::createLoopVectorizeMetadata(const LoopAttributes &Attrs,
   LLVMContext &Ctx = Header->getContext();
 
   std::optional<bool> Enabled;
-  if (Attrs.VectorizeEnable == LoopAttributes::Disable)
+  if (Attrs.VectorizeEnable == LoopAttributes::Disable ||
+      Attrs.VectorizeWidth == 1)
     Enabled = false;
   else if (Attrs.VectorizeEnable != LoopAttributes::Unspecified ||
            Attrs.VectorizePredicateEnable != LoopAttributes::Unspecified ||
@@ -643,9 +644,7 @@ 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);
-        setVectorizeScalable(LoopAttributes::Unspecified);
+        setVectorizeEnable(false);
         break;
       case LoopHintAttr::Interleave:
         // Disable interleaving by speciyfing a count of 1.
diff --git a/clang/test/CodeGenCXX/pragma-loop-pr27643.cpp b/clang/test/CodeGenCXX/pragma-loop-pr27643.cpp
index 37e29776264b6..b4614c4d39be4 100644
--- a/clang/test/CodeGenCXX/pragma-loop-pr27643.cpp
+++ b/clang/test/CodeGenCXX/pragma-loop-pr27643.cpp
@@ -40,14 +40,14 @@ void loop4(int *List, int Length) {
     List[i] = i * 2;
 }
 
-// CHECK: ![[LOOP1]] = distinct !{![[LOOP1]], [[MP:![0-9]+]], ![[VEC_WIDTH_1:.*]], ![[FIXED_WIDTH:.*]], ![[VEC_ENABLE:.*]]}
-// CHECK: ![[VEC_WIDTH_1]] = !{!"llvm.loop.vectorize.width", i32 1}
-// CHECK: ![[FIXED_WIDTH]] = !{!"llvm.loop.vectorize.scalable.enable", i1 false}
-// CHECK: ![[VEC_ENABLE]] = !{!"llvm.loop.vectorize.enable", i1 true}
+// CHECK: ![[LOOP1]] = distinct !{![[LOOP1]], [[MP:![0-9]+]], ![[VEC_DISABLE:[0-9]+]]}
+// CHECK: ![[VEC_DISABLE]] = !{!"llvm.loop.vectorize.enable", i1 false}
 
-// CHECK: ![[LOOP2]] = distinct !{![[LOOP2]], [[MP]], ![[VEC_WIDTH_2:.*]], ![[FIXED_WIDTH:.*]], ![[VEC_ENABLE]]}
+// CHECK: ![[LOOP2]] = distinct !{![[LOOP2]], [[MP]], ![[VEC_WIDTH_2:.*]], ![[FIXED_WIDTH:[0-9]+]], ![[VEC_ENABLE:[0-9]+]]}
 // CHECK: ![[VEC_WIDTH_2]] = !{!"llvm.loop.vectorize.width", i32 2}
+// CHECK: ![[FIXED_WIDTH]] = !{!"llvm.loop.vectorize.scalable.enable", i1 false}
+// CHECK: ![[VEC_ENABLE]] = !{!"llvm.loop.vectorize.enable", i1 true}
 
-// CHECK: ![[LOOP3]] = distinct !{![[LOOP3]], [[MP]], ![[VEC_WIDTH_1]], ![[FIXED_WIDTH]]}
+// CHECK: ![[LOOP3]] = distinct !{![[LOOP3]], [[MP]], ![[VEC_DISABLE]]}
 
 // CHECK: ![[LOOP4]] = distinct !{![[LOOP4]], [[MP]], ![[VEC_WIDTH_2]], ![[FIXED_WIDTH]], ![[VEC_ENABLE]]}
diff --git a/clang/test/CodeGenCXX/pragma-loop-predicate.cpp b/clang/test/CodeGenCXX/pragma-loop-predicate.cpp
index 8a25ed8de6239..4877733f6ea2a 100644
--- a/clang/test/CodeGenCXX/pragma-loop-predicate.cpp
+++ b/clang/test/CodeGenCXX/pragma-loop-predicate.cpp
@@ -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_predicate is
+// ignored.
 void test4(int *List, int Length) {
 // CHECK-LABEL: @{{.*}}test4{{.*}}(
 // CHECK: br label {{.*}}, !llvm.loop ![[LOOP4:.*]]
@@ -48,7 +48,7 @@ void test4(int *List, int Length) {
     List[i] = i * 2;
 }
 
-// Check that vectorize and vectorize_predicate are disabled.
+// Check that vectorize is disabled and vectorize_predicate is ignored.
 void test5(int *List, int Length) {
 // CHECK-LABEL: @{{.*}}test5{{.*}}(
 // CHECK: br label {{.*}}, !llvm.loop ![[LOOP5:.*]]
@@ -114,16 +114,16 @@ void test9(int *List, int Length) {
 // CHECK-NEXT: ![[LOOP3]] = distinct !{![[LOOP3]], [[MP]], [[GEN6]], [[GEN3]]}
 
 // CHECK-NEXT: ![[LOOP4]] = distinct !{![[LOOP4]], [[MP]], [[GEN10:![0-9]+]]}
-// CHECK-NEXT: [[GEN10]] = !{!"llvm.loop.vectorize.width", i32 1}
+// CHECK-NEXT: [[GEN10]] = !{!"llvm.loop.vectorize.enable", i1 false}
 
-// CHECK-NEXT: ![[LOOP5]] = distinct !{![[LOOP5]], [[MP]], [[GEN6]], [[GEN10]]}
+// CHECK-NEXT: ![[LOOP5]] = distinct !{![[LOOP5]], [[MP]], [[GEN10]]}
 
-// CHECK-NEXT: ![[LOOP6]] = distinct !{![[LOOP6]], [[MP]], [[GEN8]], [[GEN10]], [[GEN11:![0-9]+]]}
-// CHECK-NEXT: [[GEN11]] = !{!"llvm.loop.vectorize.scalable.enable", i1 false}
+// CHECK-NEXT: ![[LOOP6]] = distinct !{![[LOOP6]], [[MP]], [[GEN10]]}
 
-// CHECK-NEXT: ![[LOOP7]] = distinct !{![[LOOP7]], [[MP]], [[GEN8]], [[GEN12:![0-9]+]], [[GEN11]], [[GEN3]]}
-// CHECK-NEXT: [[GEN12]] = !{!"llvm.loop.vectorize.width", i32 4}
+// CHECK-NEXT: ![[LOOP7]] = distinct !{![[LOOP7]], [[MP]], [[GEN8]], [[GEN11:![0-9]+]], [[GEN12:![0-9]+]], [[GEN3]]}
+// CHECK-NEXT: [[GEN11]] = !{!"llvm.loop.vectorize.width", i32 4}
+// CHECK-NEXT: [[GEN12]] = !{!"llvm.loop.vectorize.scalable.enable", i1 false}
 
-// CHECK-NEXT: ![[LOOP8]] = distinct !{![[LOOP8]], [[MP]], [[GEN6]], [[GEN10]], [[GEN11]]}
+// CHECK-NEXT: ![[LOOP8]] = distinct !{![[LOOP8]], [[MP]], [[GEN10]]}
 
-// CHECK-NEXT: ![[LOOP9]] = distinct !{![[LOOP9]], [[MP]], [[GEN6]], [[GEN12]], [[GEN11]], [[GEN3]]}
+// CHECK-NEXT: ![[LOOP9]] = distinct !{![[LOOP9]], [[MP]], [[GEN6]], [[GEN11]], [[GEN12]], [[GEN3]]}
diff --git a/clang/test/CodeGenCXX/pragma-loop-safety.cpp b/clang/test/CodeGenCXX/pragma-loop-safety.cpp
index db1b6769e0514..771570f62057b 100644
--- a/clang/test/CodeGenCXX/pragma-loop-safety.cpp
+++ b/clang/test/CodeGenCXX/pragma-loop-safety.cpp
@@ -53,6 +53,6 @@ void interleave_test(int *List, int Length) {
 // CHECK: ![[INTERLEAVE_1]] = !{!"llvm.loop.interleave.count", i32 1}
 // CHECK: ![[INTENABLE_1]] = !{!"llvm.loop.vectorize.enable", i1 true}
 // CHECK: ![[ACCESS_GROUP_8]] = distinct !{}
-// CHECK: ![[LOOP2_HINTS]] = distinct !{![[LOOP2_HINTS]], [[MP]], ![[PARALLEL_ACCESSES_11:[0-9]+]], ![[UNROLL_DISABLE]], ![[WIDTH_1:[0-9]+]], ![[INTENABLE_1]]}
+// CHECK: ![[LOOP2_HINTS]] = distinct !{![[LOOP2_HINTS]], [[MP]], ![[PARALLEL_ACCESSES_11:[0-9]+]], ![[UNROLL_DISABLE]], ![[VECTORIZE_DISABLED:[0-9]+]]}
 // CHECK: ![[PARALLEL_ACCESSES_11]] = !{!"llvm.loop.parallel_accesses", ![[ACCESS_GROUP_8]]}
-// CHECK: ![[WIDTH_1]] = !{!"llvm.loop.vectorize.width", i32 1}
+// CHECK: ![[VECTORIZE_DISABLED]] = !{!"llvm.loop.vectorize.enable", i1 false}
diff --git a/clang/test/CodeGenCXX/pragma-loop.cpp b/clang/test/CodeGenCXX/pragma-loop.cpp
index 4857299f1c037..c08c42c384c04 100644
--- a/clang/test/CodeGenCXX/pragma-loop.cpp
+++ b/clang/test/CodeGenCXX/pragma-loop.cpp
@@ -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
 void for_test_scalable_1(int *List, int Length) {
 #pragma clang loop vectorize_width(1, scalable) interleave_count(4) unroll(disable) distribute(disable)
   for (int i = 0; i < Length; i++) {
@@ -203,6 +203,24 @@ void for_test_scalable_1(int *List, int Length) {
   }
 }
 
+// Verify for loop is performing scalable vectorization
+void for_test_scalable_2(int *List, int Length) {
+#pragma clang loop vectorize_width(2, scalable) interleave_count(4) unroll(disable) distribute(disable)
+  for (int i = 0; i < Length; i++) {
+    // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_20:.*]]
+    List[i] = i * 2;
+  }
+}
+
+// Verify unroll attributes are directly attached to the loop metadata
+void for_test_vectorize_disable_unroll(int *List, int Length) {
+#pragma clang loop vectorize(disable) unroll_count(8)
+  for (int i = 0; i < Length; i++) {
+    // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_21:.*]]
+    List[i] = i * 2;
+  }
+}
+
 // CHECK-DAG: ![[MP:[0-9]+]] = !{!"llvm.loop.mustprogress"}
 
 // CHECK-DAG: ![[UNROLL_DISABLE:[0-9]+]] = !{!"llvm.loop.unroll.disable"}
@@ -220,9 +238,9 @@ void for_test_scalable_1(int *List, int Length) {
 // CHECK-DAG: ![[INTERLEAVE_16:[0-9]+]] = !{!"llvm.loop.interleave.count", i32 16}
 
 // CHECK-DAG: ![[VECTORIZE_ENABLE:[0-9]+]] = !{!"llvm.loop.vectorize.enable", i1 true}
+// CHECK-DAG: ![[VECTORIZE_DISABLE:[0-9]+]] = !{!"llvm.loop.vectorize.enable", i1 false}
 // CHECK-DAG: ![[FIXED_VEC:[0-9]+]] = !{!"llvm.loop.vectorize.scalable.enable", i1 false}
 // CHECK-DAG: ![[SCALABLE_VEC:[0-9]+]] = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
-// CHECK-DAG: ![[WIDTH_1:[0-9]+]] = !{!"llvm.loop.vectorize.width", i32 1}
 // CHECK-DAG: ![[WIDTH_2:[0-9]+]] = !{!"llvm.loop.vectorize.width", i32 2}
 // CHECK-DAG: ![[WIDTH_5:[0-9]+]] = !{!"llvm.loop.vectorize.width", i32 5}
 // CHECK-DAG: ![[WIDTH_6:[0-9]+]] = !{!"llvm.loop.vectorize.width", i32 6}
@@ -241,7 +259,7 @@ void for_test_scalable_1(int *List, int Length) {
 
 // CHECK-DAG: ![[LOOP_4]] = distinct !{![[LOOP_4]], ![[WIDTH_2]], ![[FIXED_VEC]], ![[INTERLEAVE_2]], ![[VECTORIZE_ENABLE]]}
 
-// CHECK-DAG: ![[LOOP_5]] = distinct !{![[LOOP_5]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[WIDTH_1]]}
+// CHECK-DAG: ![[LOOP_5]] = distinct !{![[LOOP_5]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[VECTORIZE_DISABLE]]}
 
 // CHECK-DAG: ![[LOOP_6]] = distinct !{![[LOOP_6]], ![[MP]], ![[WIDTH_2]], ![[FIXED_VEC]], ![[INTERLEAVE_2]], ![[VECTORIZE_ENABLE]], ![[FOLLOWUP_VECTOR_3]]}
 
@@ -269,4 +287,7 @@ void for_test_scalable_1(int *List, int Length) {
 
 // CHECK-DAG: ![[LOOP_17]] = distinct !{![[LOOP_17]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[FIXED_VEC]], ![[INTERLEAVE_4]], ![[VECTORIZE_ENABLE]]}
 // CHECK-DAG: ![[LOOP_18]] = distinct !{![[LOOP_18]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[SCALABLE_VEC]], ![[INTERLEAVE_4]], ![[VECTORIZE_ENABLE]]}
-// CHECK-DAG: ![[LOOP_19]] = distinct !{![[LOOP_19]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[WIDTH_1]], ![[SCALABLE_VEC]], ![[INTERLEAVE_4]], ![[VECTORIZE_ENABLE]]}
+// CHECK-DAG: ![[LOOP_19]] = distinct !{![[LOOP_19]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[VECTORIZE_DISABLE]]}
+// CHECK-DAG: ![[LOOP_20]] = distinct !{![[LOOP_20]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[WIDTH_2]], ![[SCALABLE_VEC]], ![[INTERLEAVE_4]], ![[VECTORIZE_ENABLE]]}
+
+// CHECK-DAG: ![[LOOP_21]] = distinct !{![[LOOP_21]], ![[MP]], ![[VECTORIZE_DISABLE]], ![[UNROLL_8]]}

kasuga-fj · 2025-04-10T11:38:54Z

clang/lib/CodeGen/CGLoopInfo.cpp

@@ -200,7 +200,8 @@ LoopInfo::createLoopVectorizeMetadata(const LoopAttributes &Attrs,
  LLVMContext &Ctx = Header->getContext();

  std::optional<bool> Enabled;
-  if (Attrs.VectorizeEnable == LoopAttributes::Disable)


Perhaps this condition has never been met.

kasuga-fj · 2025-04-10T11:44:04Z

I think the llvm.loop.vectorize.enable metadata should be appended here unconditionally.

llvm-project/clang/lib/CodeGen/CGLoopInfo.cpp

Lines 282 to 301 in 79186af

    
           // vectorize.enable is set if: 
        
           // 1) loop hint vectorize.enable is set, or 
        
           // 2) it is implied when vectorize.predicate is set, or 
        
           // 3) it is implied when vectorize.width is set to a value > 1 
        
           // 4) it is implied when vectorize.scalable.enable is true 
        
           // 5) it is implied when vectorize.width is unset (0) and the user 
        
           //    explicitly requested fixed-width vectorization, i.e. 
        
           //    vectorize.scalable.enable is false. 
        
           if (Attrs.VectorizeEnable != LoopAttributes::Unspecified || 
        
               (IsVectorPredicateEnabled && Attrs.VectorizeWidth != 1) || 
        
               Attrs.VectorizeWidth > 1 || 
        
               Attrs.VectorizeScalable == LoopAttributes::Enable || 
        
               (Attrs.VectorizeScalable == LoopAttributes::Disable && 
        
                Attrs.VectorizeWidth != 1)) { 
        
             bool AttrVal = Attrs.VectorizeEnable != LoopAttributes::Disable; 
        
             Args.push_back( 
        
                 MDNode::get(Ctx, {MDString::get(Ctx, "llvm.loop.vectorize.enable"), 
        
                                   ConstantAsMetadata::get(ConstantInt::get( 
        
                                       llvm::Type::getInt1Ty(Ctx), AttrVal))})); 
        
           }

This conditional expression should be evaluated at the beginning of createVectorizeMetadata to determine the value of Enabled.

david-arm · 2025-04-10T12:32:17Z

clang/test/CodeGenCXX/pragma-loop.cpp

@@ -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


This isn't true because VF=vscale x 1 vectorize_width(1, scalable) is a valid vectorisation factor and RISC-V targets in particular make use of this. The value of vscale at runtime could be > 1, and so you're still processing > 1 element.

You are right, I misunderstood the condition, thanks.

So what if the width is 1 and fixed/scalable is not explicitly specified? Should vectorization run? If taking the intent of the original implementation, then it appears to not vectorize the loop.

llvm-project/clang/lib/CodeGen/CGLoopInfo.cpp

Lines 646 to 648 in a03b225

// Disable vectorization by specifying a width of 1.

setVectorizeWidth(1);

setVectorizeScalable(LoopAttributes::Unspecified);

If taking the intent of the original implementation, then it appears to not vectorize the loop.

Yeah I think that makes sense, i.e. we should disable vectorization if (Attrs.VectorizeEnable == LoopAttributes::Disable || (Attrs.VectorizeWidth == 1 && Attrs.VectorizeScalable != LoopAttributes:: Enable)).

We should probably also update this bit in the docs to be explicit about it:

Specifying a non-scalable width/count of 1 disables the optimization, and is equivalent to vectorize(disable) or interleave(disable).

Fixed both code and docs. Thanks.

david-arm · 2025-04-10T12:33:59Z

clang/test/CodeGenCXX/pragma-loop.cpp

@@ -269,4 +287,7 @@ void for_test_scalable_1(int *List, int Length) {

 // CHECK-DAG: ![[LOOP_17]] = distinct !{![[LOOP_17]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[FIXED_VEC]], ![[INTERLEAVE_4]], ![[VECTORIZE_ENABLE]]}
 // CHECK-DAG: ![[LOOP_18]] = distinct !{![[LOOP_18]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[SCALABLE_VEC]], ![[INTERLEAVE_4]], ![[VECTORIZE_ENABLE]]}
-// CHECK-DAG: ![[LOOP_19]] = distinct !{![[LOOP_19]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[WIDTH_1]], ![[SCALABLE_VEC]], ![[INTERLEAVE_4]], ![[VECTORIZE_ENABLE]]}
+// CHECK-DAG: ![[LOOP_19]] = distinct !{![[LOOP_19]], ![[MP]], ![[UNROLL_DISABLE]], ![[DISTRIBUTE_DISABLE]], ![[VECTORIZE_DISABLE]]}


Again, I think this is broken. We should still be vectorising with VF=vscale x 1

Undid this, thanks.

lukel97

Thanks, LGTM. Though this is a part of the code that I'm not familiar with so probably best to wait for another LGTM

david-arm · 2025-04-15T08:51:24Z

clang/test/CodeGenCXX/pragma-loop-predicate.cpp

@@ -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_predicate is


nit: The loop doesn't have a vectorize_predicate pragma so I think this should just be // Check that vectorization is disabled

You are right, thanks!

llvmbot added clang Clang issues not falling into any other category clang:codegen IR generation bugs: mangling, exceptions, etc. labels Apr 10, 2025

kasuga-fj commented Apr 10, 2025

View reviewed changes

kasuga-fj requested review from Meinersbur, sjoerdmeijer and david-arm April 10, 2025 11:45

david-arm requested a review from fhahn April 10, 2025 12:26

david-arm reviewed Apr 10, 2025

View reviewed changes

david-arm requested a review from lukel97 April 10, 2025 12:34

kasuga-fj added 3 commits April 11, 2025 09:44

Fix OpenMP test

aed6f10

Fix behavior when scalable is specified

98bee49

Update the docs

0e60403

llvmbot added the clang:openmp OpenMP related changes to Clang label Apr 11, 2025

lukel97 approved these changes Apr 11, 2025

View reviewed changes

david-arm reviewed Apr 15, 2025

View reviewed changes

Fix the comment, thanks!

16b6d51

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[clang][CodeGen] Fix metadata when vectorization is disabled by pragma #135163

[clang][CodeGen] Fix metadata when vectorization is disabled by pragma #135163

kasuga-fj commented Apr 10, 2025 •

edited

Loading

llvmbot commented Apr 10, 2025 •

edited

Loading

kasuga-fj Apr 10, 2025

kasuga-fj commented Apr 10, 2025

david-arm Apr 10, 2025

kasuga-fj Apr 10, 2025

lukel97 Apr 10, 2025

kasuga-fj Apr 11, 2025

david-arm Apr 10, 2025

kasuga-fj Apr 11, 2025

lukel97 left a comment

david-arm Apr 15, 2025

kasuga-fj Apr 15, 2025

	// Disable vectorization by specifying a width of 1.
	setVectorizeWidth(1);
	setVectorizeScalable(LoopAttributes::Unspecified);

[clang][CodeGen] Fix metadata when vectorization is disabled by pragma #135163

Are you sure you want to change the base?

[clang][CodeGen] Fix metadata when vectorization is disabled by pragma #135163

Conversation

kasuga-fj commented Apr 10, 2025 • edited Loading

llvmbot commented Apr 10, 2025 • edited Loading

Choose a reason for hiding this comment

kasuga-fj commented Apr 10, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

lukel97 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kasuga-fj commented Apr 10, 2025 •

edited

Loading

llvmbot commented Apr 10, 2025 •

edited

Loading