[VPlan] Only use SCEV for live-ins in tryToWiden.

[fhahn]

Replacing a recipe with a live-in may not be correct in all cases,
e.g. when replacing recipes involving header-phi recipes, like
reductions.

For now, only use SCEV to simplify live-ins.

More powerful input simplification can be built in top of
#124432 in the future.

Fixes #119173.
Fixes #125374.

[llvmbot]

@llvm/pr-subscribers-vectorizers

@llvm/pr-subscribers-llvm-transforms

Author: Florian Hahn (fhahn)

Changes

Fixes #119173.

---

Full diff: https://github.com/llvm/llvm-project/pull/125436.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/LoopVectorize.cpp (+6-2)
• (modified) llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll (+14-27)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index f88c13525421c8..1361737248284d 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -6626,8 +6626,10 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I,
     // fold away.  We can generalize this for all operations using the notion
     // of neutral elements.  (TODO)
     if (I->getOpcode() == Instruction::Mul &&
-        (PSE.getSCEV(I->getOperand(0))->isOne() ||
-         PSE.getSCEV(I->getOperand(1))->isOne()))
+        ((TheLoop->isLoopInvariant(I->getOperand(0)) &&
+          PSE.getSCEV(I->getOperand(0))->isOne()) ||
+         (TheLoop->isLoopInvariant(I->getOperand(1)) &&
+          PSE.getSCEV(I->getOperand(1))->isOne())))
       return 0;
 
     // Detect reduction patterns
@@ -8577,6 +8579,8 @@ VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I,
       // to replace operands with constants.
       ScalarEvolution &SE = *PSE.getSE();
       auto GetConstantViaSCEV = [this, &SE](VPValue *Op) {
+        if (!Op->isLiveIn())
+          return Op;
         Value *V = Op->getUnderlyingValue();
         if (isa<Constant>(V) || !SE.isSCEVable(V->getType()))
           return Op;
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll b/llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll
index 6b55f5291efd8c..a8d44421a3c37b 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/mul-simplification.ll
@@ -7,24 +7,10 @@ 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 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[RED_NEXT:%.*]], %[[LOOP]] ]
-; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[RED:%.*]] = phi i64 [ 12, %[[ENTRY]] ], [ [[RED_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[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]]
@@ -32,9 +18,9 @@ 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]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]]
 ; CHECK:       [[EXIT]]:
-; CHECK-NEXT:    [[RES:%.*]] = phi i64 [ [[RED_NEXT]], %[[LOOP]] ], [ [[TMP0]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[RES:%.*]] = phi i64 [ [[RED_NEXT]], %[[LOOP]] ]
 ; CHECK-NEXT:    ret i64 [[RES]]
 ;
 entry:
@@ -65,17 +51,20 @@ 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]] ], [ splat (i32 42), %[[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ splat (i32 42), %[[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:    [[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 [[LOOP4:![0-9]+]]
+; CHECK-NEXT:    br i1 [[TMP0]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
-; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> splat (i32 84))
+; 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:    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 [ [[TMP1]], %[[MIDDLE_BLOCK]] ], [ 42, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP3]], %[[MIDDLE_BLOCK]] ], [ 42, %[[ENTRY]] ]
 ; CHECK-NEXT:    br label %[[LOOP:.*]]
 ; CHECK:       [[LOOP]]:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[ADD2_REASS:%.*]], %[[LOOP]] ]
@@ -83,9 +72,9 @@ 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 [[LOOP5:![0-9]+]]
+; CHECK-NEXT:    br i1 [[CMP]], label %[[LOOP]], label %[[EXIT]], !llvm.loop [[LOOP3:![0-9]+]]
 ; CHECK:       [[EXIT]]:
-; CHECK-NEXT:    [[ADD_LCSSA:%.*]] = phi i32 [ [[RDX_NEXT]], %[[LOOP]] ], [ [[TMP1]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[ADD_LCSSA:%.*]] = phi i32 [ [[RDX_NEXT]], %[[LOOP]] ], [ [[TMP3]], %[[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[ADD_LCSSA]]
 ;
 entry:
@@ -109,6 +98,4 @@ 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]]}
 ;.