[LV] Fix tryToWiden logic for binop simplification

[artagnon]

Fix the logic in the GetConstantViaSCEV routine in tryToWiden to match intent: we don't need to check if the VPValue has a defining recipe, and can simply check the underlying value. This allows us to better vectorize certain cases of binary ops.

[llvmbot]

@llvm/pr-subscribers-vectorizers

@llvm/pr-subscribers-llvm-transforms

Author: Ramkumar Ramachandra (artagnon)

Changes

Fix the logic in the GetConstantViaSCEV routine in tryToWiden to match intent: we don't need to check if the VPValue has a defining recipe, and can simply check the underlying value. This allows us to better vectorize certain cases of binary ops.

---

Full diff: https://github.com/llvm/llvm-project/pull/128886.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/LoopVectorize.cpp (+7-10)
• (modified) llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll (+27-14)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index e8a5db28ea0a4..9e7175ed529dc 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -8621,17 +8621,14 @@ VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I,
       // The legacy cost model uses SCEV to check if some of the operands are
       // constants. To match the legacy cost model's behavior, use SCEV to try
       // to replace operands with constants.
-      ScalarEvolution &SE = *PSE.getSE();
-      auto GetConstantViaSCEV = [this, &SE](VPValue *Op) {
-        if (!Op->isLiveIn())
+      auto GetConstantViaSCEV = [this](VPValue *Op) {
+        ScalarEvolution &SE = *PSE.getSE();
+        Value *V = dyn_cast_if_present<Value>(Op->getUnderlyingValue());
+        if (!V || !SE.isSCEVable(V->getType()))
           return Op;
-        Value *V = Op->getUnderlyingValue();
-        if (isa<Constant>(V) || !SE.isSCEVable(V->getType()))
-          return Op;
-        auto *C = dyn_cast<SCEVConstant>(SE.getSCEV(V));
-        if (!C)
-          return Op;
-        return Plan.getOrAddLiveIn(C->getValue());
+        if (auto *C = dyn_cast<SCEVConstant>(SE.getSCEV(V)))
+          return Plan.getOrAddLiveIn(C->getValue());
+        return Op;
       };
       // For Mul, the legacy cost model checks both operands.
       if (I->getOpcode() == Instruction::Mul)
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll b/llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll
index a8d44421a3c37..6b55f5291efd8 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll
@@ -7,10 +7,24 @@ target triple = "arm64-apple-macosx"
 define i64 @mul_select_operand_known_1_via_scev() {
 ; CHECK-LABEL: define i64 @mul_select_operand_known_1_via_scev() {
 ; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    br i1 false, label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <2 x i64> [ <i64 12, i64 1>, %[[VECTOR_PH]] ], [ [[VEC_PHI]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
+; CHECK-NEXT:    br i1 true, label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vector.reduce.mul.v2i64(<2 x i64> [[VEC_PHI]])
+; CHECK-NEXT:    br i1 true, label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; CHECK:       [[SCALAR_PH]]:
+; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i64 [ [[TMP0]], %[[MIDDLE_BLOCK]] ], [ 12, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 2, %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
 ; CHECK-NEXT:    br label %[[LOOP:.*]]
 ; CHECK:       [[LOOP]]:
-; CHECK-NEXT:    [[RED:%.*]] = phi i64 [ 12, %[[ENTRY]] ], [ [[RED_NEXT:%.*]], %[[LOOP]] ]
-; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[RED:%.*]] = phi i64 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[RED_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
 ; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[IV]], 1
 ; CHECK-NEXT:    [[CMP1_I:%.*]] = icmp eq i32 [[TMP1]], 0
 ; CHECK-NEXT:    [[NARROW_I:%.*]] = select i1 [[CMP1_I]], i32 1, i32 [[IV]]
@@ -18,9 +32,9 @@ define i64 @mul_select_operand_known_1_via_scev() {
 ; CHECK-NEXT:    [[RED_NEXT]] = mul nsw i64 [[RED]], [[MUL]]
 ; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1
 ; CHECK-NEXT:    [[EC:%.*]] = icmp eq i32 [[IV]], 1
-; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]]
+; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
 ; CHECK:       [[EXIT]]:
-; CHECK-NEXT:    [[RES:%.*]] = phi i64 [ [[RED_NEXT]], %[[LOOP]] ]
+; CHECK-NEXT:    [[RES:%.*]] = phi i64 [ [[RED_NEXT]], %[[LOOP]] ], [ [[TMP0]], %[[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i64 [[RES]]
 ;
 entry:
@@ -51,20 +65,17 @@ define i32 @add_reduction_select_operand_constant_but_non_uniform() {
 ; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
 ; CHECK:       [[VECTOR_BODY]]:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ <i32 42, i32 0, i32 0, i32 0>, %[[VECTOR_PH]] ], [ [[TMP2:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP1:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[TMP2]] = add <4 x i32> zeroinitializer, [[VEC_PHI]]
-; CHECK-NEXT:    [[TMP1]] = add <4 x i32> zeroinitializer, [[VEC_PHI1]]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ <i32 42, i32 0, i32 0, i32 0>, %[[VECTOR_PH]] ], [ splat (i32 42), %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ splat (i32 42), %[[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
 ; CHECK-NEXT:    [[TMP0:%.*]] = icmp eq i32 [[INDEX_NEXT]], 64
-; CHECK-NEXT:    br i1 [[TMP0]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK-NEXT:    br i1 [[TMP0]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
-; CHECK-NEXT:    [[BIN_RDX:%.*]] = add <4 x i32> [[TMP1]], [[TMP2]]
-; CHECK-NEXT:    [[TMP3:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX]])
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> splat (i32 84))
 ; CHECK-NEXT:    br i1 true, label %[[EXIT:.*]], label %[[SCALAR_PH]]
 ; CHECK:       [[SCALAR_PH]]:
 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 64, %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
-; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP3]], %[[MIDDLE_BLOCK]] ], [ 42, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP1]], %[[MIDDLE_BLOCK]] ], [ 42, %[[ENTRY]] ]
 ; CHECK-NEXT:    br label %[[LOOP:.*]]
 ; CHECK:       [[LOOP]]:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[ADD2_REASS:%.*]], %[[LOOP]] ]
@@ -72,9 +83,9 @@ define i32 @add_reduction_select_operand_constant_but_non_uniform() {
 ; CHECK-NEXT:    [[ADD2_REASS]] = add i32 [[IV]], 1
 ; CHECK-NEXT:    [[RDX_NEXT]] = add i32 0, [[RDX]]
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[ADD2_REASS]], 64
-; CHECK-NEXT:    br i1 [[CMP]], label %[[LOOP]], label %[[EXIT]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK-NEXT:    br i1 [[CMP]], label %[[LOOP]], label %[[EXIT]], !llvm.loop [[LOOP5:![0-9]+]]
 ; CHECK:       [[EXIT]]:
-; CHECK-NEXT:    [[ADD_LCSSA:%.*]] = phi i32 [ [[RDX_NEXT]], %[[LOOP]] ], [ [[TMP3]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[ADD_LCSSA:%.*]] = phi i32 [ [[RDX_NEXT]], %[[LOOP]] ], [ [[TMP1]], %[[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[ADD_LCSSA]]
 ;
 entry:
@@ -98,4 +109,6 @@ exit:
 ; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
 ; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
 ; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
+; CHECK: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}
+; CHECK: [[LOOP5]] = distinct !{[[LOOP5]], [[META2]], [[META1]]}
 ;.

[artagnon]

Per what @fhahn said in #125365, this is supposed to be incorrect, but I can't explain the improvements then. Can someone have a look?

[fhahn]

This is miscompiled with the patch, it now reduces <84, 84, 84 ,84> instead of <84, 0, 0, 0>.

See #125436 for more context

[artagnon]

Thanks, and sorry for not checking the history of the changed lines.