[AArch64] Return Invalid partial reduction cost for i128 accumulator.

[sdesmalen-arm]

PR #158641 introduced an issue where i128 accumulator types resulted
in a valid cost, because for a <2 x i128> type the code that
checks for unsupported type legalization would see a type action
of 'TypeSplitVector' which is supported, even though the legalised
type of <1 x i128> would require further scalarization.

This fixes #162009

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-backend-aarch64

Author: Sander de Smalen (sdesmalen-arm)

Changes

PR #158641 introduced an issue where i128 accumulator types resulted
in a valid cost, because for a <2 x i128> type the code that
checks for unsupported type legalization would see a type action
of 'TypeSplitVector' which is supported, even though the legalised
type of <1 x i128> would require further scalarization.

This fixes #162009

---

Full diff: https://github.com/llvm/llvm-project/pull/162066.diff

2 Files Affected:

• (modified) llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp (+7-4)
• (added) llvm/test/Transforms/LoopVectorize/AArch64/pr162009.ll (+79)

diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 50a8754d74075..bb3b9d5bf2c74 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -5666,18 +5666,21 @@ InstructionCost AArch64TTIImpl::getPartialReductionCost(
   VectorType *AccumVectorType =
       VectorType::get(AccumType, VF.divideCoefficientBy(Ratio));
   // We don't yet support all kinds of legalization.
-  auto TA = TLI->getTypeAction(AccumVectorType->getContext(),
-                               EVT::getEVT(AccumVectorType));
-  switch (TA) {
+  auto TC = TLI->getTypeConversion(AccumVectorType->getContext(),
+                             EVT::getEVT(AccumVectorType));
+  switch (TC.first) {
   default:
     return Invalid;
   case TargetLowering::TypeLegal:
   case TargetLowering::TypePromoteInteger:
   case TargetLowering::TypeSplitVector:
+    // The legalised type (e.g. after splitting) must be legal too.
+    if (TLI->getTypeAction(AccumVectorType->getContext(), TC.second) !=
+        TargetLowering::TypeLegal)
+      return Invalid;
     break;
   }
 
-  // Check what kind of type-legalisation happens.
   std::pair<InstructionCost, MVT> AccumLT =
       getTypeLegalizationCost(AccumVectorType);
   std::pair<InstructionCost, MVT> InputLT =
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/pr162009.ll b/llvm/test/Transforms/LoopVectorize/AArch64/pr162009.ll
new file mode 100644
index 0000000000000..6095b246ccdad
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/pr162009.ll
@@ -0,0 +1,79 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -enable-epilogue-vectorization=false -S < %s | FileCheck %s --check-prefixes=CHECK-NO-PARTIAL-REDUCTION
+
+target triple = "aarch64"
+
+define i128 @add_reduc_i32_i128_unsupported(ptr %a, ptr %b) "target-features"="+dotprod" {
+; CHECK-NO-PARTIAL-REDUCTION-LABEL: define i128 @add_reduc_i32_i128_unsupported(
+; CHECK-NO-PARTIAL-REDUCTION-SAME: ptr [[A:%.*]], ptr [[B:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:  [[ENTRY:.*:]]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    br label %[[VECTOR_PH:.*]]
+; CHECK-NO-PARTIAL-REDUCTION:       [[VECTOR_PH]]:
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK-NO-PARTIAL-REDUCTION:       [[VECTOR_BODY]]:
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i128> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP7:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[TMP0:%.*]] = getelementptr i32, ptr [[A]], i64 [[INDEX]]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP0]], align 1
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[TMP1:%.*]] = zext <4 x i32> [[WIDE_LOAD]] to <4 x i64>
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[TMP2:%.*]] = getelementptr i32, ptr [[B]], i64 [[INDEX]]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, ptr [[TMP2]], align 1
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[TMP3:%.*]] = zext <4 x i32> [[WIDE_LOAD1]] to <4 x i64>
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[TMP4:%.*]] = mul nuw <4 x i64> [[TMP1]], [[TMP3]]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[TMP5:%.*]] = zext <4 x i64> [[TMP4]] to <4 x i128>
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[TMP7]] = add <4 x i128> [[VEC_PHI]], [[TMP5]]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4024
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    br i1 [[TMP6]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK-NO-PARTIAL-REDUCTION:       [[MIDDLE_BLOCK]]:
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[TMP8:%.*]] = call i128 @llvm.vector.reduce.add.v4i128(<4 x i128> [[TMP7]])
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    br label %[[SCALAR_PH:.*]]
+; CHECK-NO-PARTIAL-REDUCTION:       [[SCALAR_PH]]:
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    br label %[[FOR_BODY:.*]]
+; CHECK-NO-PARTIAL-REDUCTION:       [[FOR_BODY]]:
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[IV:%.*]] = phi i64 [ 4024, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[FOR_BODY]] ]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[ACCUM:%.*]] = phi i128 [ [[TMP8]], %[[SCALAR_PH]] ], [ [[ADD:%.*]], %[[FOR_BODY]] ]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[GEP_A:%.*]] = getelementptr i32, ptr [[A]], i64 [[IV]]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[LOAD_A:%.*]] = load i32, ptr [[GEP_A]], align 1
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[EXT_A:%.*]] = zext i32 [[LOAD_A]] to i64
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[GEP_B:%.*]] = getelementptr i32, ptr [[B]], i64 [[IV]]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[LOAD_B:%.*]] = load i32, ptr [[GEP_B]], align 1
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[EXT_B:%.*]] = zext i32 [[LOAD_B]] to i64
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[MUL:%.*]] = mul nuw i64 [[EXT_A]], [[EXT_B]]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[MUL_ZEXT:%.*]] = zext i64 [[MUL]] to i128
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[ADD]] = add i128 [[ACCUM]], [[MUL_ZEXT]]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[IV_NEXT]] = add i64 [[IV]], 1
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], 4025
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    br i1 [[EXITCOND_NOT]], label %[[FOR_EXIT:.*]], label %[[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK-NO-PARTIAL-REDUCTION:       [[FOR_EXIT]]:
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    [[ADD_LCSSA:%.*]] = phi i128 [ [[ADD]], %[[FOR_BODY]] ]
+; CHECK-NO-PARTIAL-REDUCTION-NEXT:    ret i128 [[ADD_LCSSA]]
+;
+entry:
+  br label %for.body
+
+for.body:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
+  %accum = phi i128 [ 0, %entry ], [ %add, %for.body ]
+  %gep.a = getelementptr i32, ptr %a, i64 %iv
+  %load.a = load i32, ptr %gep.a, align 1
+  %ext.a = zext i32 %load.a to i64
+  %gep.b = getelementptr i32, ptr %b, i64 %iv
+  %load.b = load i32, ptr %gep.b, align 1
+  %ext.b = zext i32 %load.b to i64
+  %mul = mul nuw i64 %ext.a, %ext.b
+  %mul.zext = zext i64 %mul to i128
+  %add = add i128 %accum, %mul.zext
+  %iv.next = add i64 %iv, 1
+  %exitcond.not = icmp eq i64 %iv.next, 4025
+  br i1 %exitcond.not, label %for.exit, label %for.body
+
+for.exit:
+  ret i128 %add
+}
+;.
+; CHECK-NO-PARTIAL-REDUCTION: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
+; CHECK-NO-PARTIAL-REDUCTION: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK-NO-PARTIAL-REDUCTION: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
+; CHECK-NO-PARTIAL-REDUCTION: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
+;.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[david-arm]

LGTM!