[llvm] cf976bf - [LoopInterchange] Add tests for the vectorization profitability (NFC) (#133665)

[via llvm-commits]

Author: Ryotaro Kasuga
Date: 2025-04-02T21:02:30+09:00
New Revision: cf976bfdeba21405e1674c6dd09940773b2275ef

URL: https://github.com/llvm/llvm-project/commit/cf976bfdeba21405e1674c6dd09940773b2275ef
DIFF: https://github.com/llvm/llvm-project/commit/cf976bfdeba21405e1674c6dd09940773b2275ef.diff

LOG: [LoopInterchange] Add tests for the vectorization profitability (NFC) (#133665)

There is a problem with the current profitability check for
vectorization in LoopInterchange. There are both false positives and
false negatives. The former means that the heuristic may say that "an
exchange is necessary to vectorize the innermost loop" even though it's
already possible. The latter means that the heuristic may miss a case
where an exchange is necessary to vectorize the innermost loop. Note
that this is not a dependency analysis problem. This is caused by
incorrect handling of the dependency matrix in the profitability check,
so these problems can occur even if the analysis is accurate (no
overestimation).

This patch adds tests to clarify the cases that should be fixed. The
root cause of these cases is that the heuristic doesn't handle the
direction of a dependency correctly.

Added: 
    llvm/test/Transforms/LoopInterchange/profitability-vectorization-heuristic.ll

Modified: 
    

Removed: 
    


################################################################################
diff  --git a/llvm/test/Transforms/LoopInterchange/profitability-vectorization-heuristic.ll b/llvm/test/Transforms/LoopInterchange/profitability-vectorization-heuristic.ll
new file mode 100644
index 0000000000000..efd2a9c09e7cf
--- /dev/null
+++ b/llvm/test/Transforms/LoopInterchange/profitability-vectorization-heuristic.ll
@@ -0,0 +1,108 @@
+; RUN: opt < %s -passes=loop-interchange -cache-line-size=64 \
+; RUN:     -pass-remarks-output=%t -disable-output -loop-interchange-profitabilities=vectorize
+; RUN: FileCheck -input-file %t %s
+
+ at A = dso_local global [256 x [256 x float]] zeroinitializer
+ at B = dso_local global [256 x [256 x float]] zeroinitializer
+ at C = dso_local global [256 x [256 x float]] zeroinitializer
+
+; Check that the below loops are exchanged for vectorization.
+;
+; for (int i = 0; i < 256; i++) {
+;   for (int j = 1; j < 256; j++) {
+;     A[i][j] = A[i][j-1] + B[i][j];
+;     C[i][j] += 1;
+;   }
+; }
+;
+; FIXME: These loops are not exchanged at this time due to the problem in
+; profitability heuristic calculation for vectorization.
+
+; CHECK:      --- !Missed
+; CHECK-NEXT: Pass:            loop-interchange
+; CHECK-NEXT: Name:            InterchangeNotProfitable
+; CHECK-NEXT: Function:        interchange_necessary_for_vectorization
+; CHECK-NEXT: Args:
+; CHECK-NEXT:   - String:          Interchanging loops is not considered to improve cache locality nor vectorization.
+; CHECK-NEXT: ...
+define void @interchange_necessary_for_vectorization() {
+entry:
+  br label %for.i.header
+
+for.i.header:
+  %i = phi i64 [ 1, %entry ], [ %i.next, %for.i.inc ]
+  br label %for.j.body
+
+for.j.body:
+  %j = phi i64 [ 1, %for.i.header ], [ %j.next, %for.j.body ]
+  %j.dec = add nsw i64 %j, -1
+  %a.load.index = getelementptr nuw inbounds [256 x [256 x float]], ptr @A, i64 %i, i64 %j.dec
+  %b.index = getelementptr nuw inbounds [256 x [256 x float]], ptr @B, i64 %i, i64 %j
+  %c.index = getelementptr nuw inbounds [256 x [256 x float]], ptr @C, i64 %i, i64 %j
+  %a = load float, ptr %a.load.index, align 4
+  %b = load float, ptr %b.index, align 4
+  %c = load float, ptr %c.index, align 4
+  %add.0 = fadd float %a, %b
+  %a.store.index = getelementptr nuw inbounds [256 x [256 x float]], ptr @A, i64 %i, i64 %j
+  store float %add.0, ptr %a.store.index, align 4
+  %add.1 = fadd float %c, 1.0
+  store float %add.1, ptr %c.index, align 4
+  %j.next = add nuw nsw i64 %j, 1
+  %cmp.j = icmp eq i64 %j.next, 256
+  br i1 %cmp.j, label %for.i.inc, label %for.j.body
+
+for.i.inc:
+  %i.next = add nuw nsw i64 %i, 1
+  %cmp.i = icmp eq i64 %i.next, 256
+  br i1 %cmp.i, label %exit, label %for.i.header
+
+exit:
+  ret void
+}
+
+; Check that the following innermost loop can be vectorized so that
+; interchanging is unnecessary.
+;
+; for (int i = 0; i < 256; i++)
+;   for (int j = 1; j < 256; j++)
+;     A[i][j-1] = A[i][j] + B[i][j];
+;
+; FIXME: These loops are exchanged at this time due to the problem in
+; profitability heuristic calculation for vectorization.
+
+; CHECK:      --- !Passed
+; CHECK-NEXT: Pass:            loop-interchange
+; CHECK-NEXT: Name:            Interchanged
+; CHECK-NEXT: Function:        interchange_unnecesasry_for_vectorization
+; CHECK-NEXT: Args:
+; CHECK-NEXT:   - String:          Loop interchanged with enclosing loop.
+define void @interchange_unnecesasry_for_vectorization() {
+entry:
+  br label %for.i.header
+
+for.i.header:
+  %i = phi i64 [ 1, %entry ], [ %i.next, %for.i.inc ]
+  br label %for.j.body
+
+for.j.body:
+  %j = phi i64 [ 1, %for.i.header ], [ %j.next, %for.j.body ]
+  %j.dec = add nsw i64 %j, -1
+  %a.load.index = getelementptr nuw inbounds [256 x [256 x float]], ptr @A, i64 %i, i64 %j
+  %b.index = getelementptr nuw inbounds [256 x [256 x float]], ptr @B, i64 %i, i64 %j
+  %a = load float, ptr %a.load.index, align 4
+  %b = load float, ptr %b.index, align 4
+  %add = fadd float %a, %b
+  %a.store.index = getelementptr nuw inbounds [256 x [256 x float]], ptr @A, i64 %i, i64 %j.dec
+  store float %add, ptr %a.store.index, align 4
+  %j.next = add nuw nsw i64 %j, 1
+  %cmp.j = icmp eq i64 %j.next, 256
+  br i1 %cmp.j, label %for.i.inc, label %for.j.body
+
+for.i.inc:
+  %i.next = add nuw nsw i64 %i, 1
+  %cmp.i = icmp eq i64 %i.next, 256
+  br i1 %cmp.i, label %exit, label %for.i.header
+
+exit:
+  ret void
+}