[VPlan] Skip cost assert if VPlan converted to single-scalar recipes.

Check if a VPlan transform converted recipes to single-scalar
VPReplicateRecipes (after 07c085a). If
that's the case, the legacy cost model incorrectly overestimates the cost.

Fixes #141237.

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -7082,6 +7082,11 @@ InstructionCost VPCostContext::getLegacyCost(Instruction *UI,
   return CM.getInstructionCost(UI, VF);
 }
 
+bool VPCostContext::isLegacyUniformAfterVectorization(Instruction *I,
+                                                      ElementCount VF) const {
+  return CM.isUniformAfterVectorization(I, VF);
+}
+
 bool VPCostContext::skipCostComputation(Instruction *UI, bool IsVector) const {
   return CM.ValuesToIgnore.contains(UI) ||
          (IsVector && CM.VecValuesToIgnore.contains(UI)) ||
@@ -7315,7 +7320,8 @@ InstructionCost LoopVectorizationPlanner::cost(VPlan &Plan,
 /// cost-model did not account for.
 static bool planContainsAdditionalSimplifications(VPlan &Plan,
                                                   VPCostContext &CostCtx,
-                                                  Loop *TheLoop) {
+                                                  Loop *TheLoop,
+                                                  ElementCount VF) {
   // First collect all instructions for the recipes in Plan.
   auto GetInstructionForCost = [](const VPRecipeBase *R) -> Instruction * {
     if (auto *S = dyn_cast<VPSingleDefRecipe>(R))
@@ -7352,6 +7358,16 @@ static bool planContainsAdditionalSimplifications(VPlan &Plan,
       // comparing against the legacy cost isn't desirable.
       if (isa<VPPartialReductionRecipe>(&R))
         return true;
+
+      /// If a VPlan transform folded a recipe to one producing a single-scalar,
+      /// but the original instruction wasn't uniform-after-vectorization in the
+      /// legacy cost model, the legacy cost overestimates the actual cost.
+      if (auto *RepR = dyn_cast<VPReplicateRecipe>(&R)) {
+        if (RepR->isSingleScalar() &&
+            !CostCtx.isLegacyUniformAfterVectorization(
+                RepR->getUnderlyingInstr(), VF))
+          return true;
+      }
       if (Instruction *UI = GetInstructionForCost(&R)) {
         // If we adjusted the predicate of the recipe, the cost in the legacy
         // cost model may be different.
@@ -7477,9 +7493,10 @@ VectorizationFactor LoopVectorizationPlanner::computeBestVF() {
   // legacy cost model doesn't properly model costs for such loops.
   assert((BestFactor.Width == LegacyVF.Width || BestPlan.hasEarlyExit() ||
           planContainsAdditionalSimplifications(getPlanFor(BestFactor.Width),
-                                                CostCtx, OrigLoop) ||
-          planContainsAdditionalSimplifications(getPlanFor(LegacyVF.Width),
-                                                CostCtx, OrigLoop)) &&
+                                                CostCtx, OrigLoop,
+                                                BestFactor.Width) ||
+          planContainsAdditionalSimplifications(
+              getPlanFor(LegacyVF.Width), CostCtx, OrigLoop, LegacyVF.Width)) &&
          " VPlan cost model and legacy cost model disagreed");
   assert((BestFactor.Width.isScalar() || BestFactor.ScalarCost > 0) &&
          "when vectorizing, the scalar cost must be computed.");

llvm/lib/Transforms/Vectorize/VPlanHelpers.h

@@ -364,6 +364,11 @@ struct VPCostContext {
 
   /// Returns the OperandInfo for \p V, if it is a live-in.
   TargetTransformInfo::OperandValueInfo getOperandInfo(VPValue *V) const;
+
+  /// Return true if \p I is considered uniform-after-vectorization in the
+  /// legacy cost model for \p VF. Only used to check for additional VPlan
+  /// simplifications.
+  bool isLegacyUniformAfterVectorization(Instruction *I, ElementCount VF) const;
 };
 
 /// This class can be used to assign names to VPValues. For VPValues without

llvm/test/Transforms/LoopVectorize/X86/uniform_load.ll

@@ -1,4 +1,10 @@
-; RUN: opt -passes=loop-vectorize -S -mcpu=core-avx2 < %s | FileCheck %s
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --filter-out-after "middle.block:" --version 5
+; RUN: opt -passes=loop-vectorize -S %s | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@inc = global float 5.000000e-01, align 4
 
 ;float inc = 0.5;
 ;void foo(ptr A, unsigned N) {
@@ -8,40 +14,116 @@
 ;  }
 ;}
 
-; CHECK-LABEL: foo
-; CHECK: vector.body
-; CHECK: load <8 x float>
-; CHECK: fadd <8 x float>
-; CHECK: store <8 x float>
+define void @foo(ptr nocapture noalias %A, i64 %N) #0 {
+; CHECK-LABEL: define void @foo(
+; CHECK-SAME: ptr noalias captures(none) [[A:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br i1 false, [[SCALAR_PH:label %.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr @inc, align 4
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x float> poison, float [[TMP1]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x float> [[BROADCAST_SPLATINSERT]], <8 x float> poison, <8 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds float, ptr [[A]], i64 0
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 0
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 8
+; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 16
+; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 24
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <8 x float>, ptr [[TMP3]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <8 x float>, ptr [[TMP4]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD3:%.*]] = load <8 x float>, ptr [[TMP5]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD4:%.*]] = load <8 x float>, ptr [[TMP6]], align 4
+; CHECK-NEXT:    [[TMP7:%.*]] = fadd <8 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP8:%.*]] = fadd <8 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD2]]
+; CHECK-NEXT:    [[TMP9:%.*]] = fadd <8 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD3]]
+; CHECK-NEXT:    [[TMP10:%.*]] = fadd <8 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD4]]
+; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 0
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 8
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 16
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 24
+; CHECK-NEXT:    store <8 x float> [[TMP7]], ptr [[TMP14]], align 4
+; CHECK-NEXT:    store <8 x float> [[TMP8]], ptr [[TMP11]], align 4
+; CHECK-NEXT:    store <8 x float> [[TMP9]], ptr [[TMP12]], align 4
+; CHECK-NEXT:    store <8 x float> [[TMP10]], ptr [[TMP13]], align 4
+; CHECK-NEXT:    br label %[[MIDDLE_BLOCK:.*]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+;
+entry:
+  br label %loop
 
-target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
-target triple = "x86_64-unknown-linux-gnu"
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %l.inc = load float, ptr @inc, align 4
+  %gep.A = getelementptr inbounds float, ptr %A, i64 %iv
+  %l.A = load float, ptr %gep.A, align 4
+  %add = fadd float %l.inc, %l.A
+  store float %add, ptr %gep.A, align 4
+  %iv.next = add nuw nsw i64 %iv, 1
+  %ec = icmp eq i64 %iv.next, 32
+  br i1 %ec, label %exit, label %loop
 
-@inc = global float 5.000000e-01, align 4
+exit:
+  ret void
+}
 
-define void @foo(ptr nocapture %A, i32 %N) #0 {
+define void @uniform_load_can_fold_users(ptr noalias %src, ptr %dst, i64 %start, double %d) {
+; CHECK-LABEL: define void @uniform_load_can_fold_users(
+; CHECK-SAME: ptr noalias [[SRC:%.*]], ptr [[DST:%.*]], i64 [[START:%.*]], double [[D:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[START]], 1
+; CHECK-NEXT:    [[SMIN:%.*]] = call i64 @llvm.smin.i64(i64 [[START]], i64 0)
+; CHECK-NEXT:    [[TMP1:%.*]] = sub i64 [[TMP0]], [[SMIN]]
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP1]], 2
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], [[SCALAR_PH:label %.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP1]], 2
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP1]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[TMP2:%.*]] = sub i64 [[START]], [[N_VEC]]
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 0
+; CHECK-NEXT:    [[TMP4:%.*]] = add i64 [[INDEX]], 1
+; CHECK-NEXT:    [[TMP5:%.*]] = load double, ptr [[SRC]], align 8
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x double> poison, double [[TMP5]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x double> [[BROADCAST_SPLATINSERT]], <2 x double> poison, <2 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP6:%.*]] = fmul <2 x double> [[BROADCAST_SPLAT]], splat (double 9.000000e+00)
+; CHECK-NEXT:    [[TMP7:%.*]] = extractelement <2 x double> [[TMP6]], i32 0
+; CHECK-NEXT:    [[TMP8:%.*]] = fdiv double [[TMP7]], [[D]]
+; CHECK-NEXT:    [[TMP9:%.*]] = sub i64 [[TMP3]], 1
+; CHECK-NEXT:    [[TMP10:%.*]] = sub i64 [[TMP4]], 1
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr double, ptr [[DST]], i64 [[TMP3]]
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr double, ptr [[DST]], i64 [[TMP4]]
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr double, ptr [[TMP11]], i64 [[TMP9]]
+; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr double, ptr [[TMP12]], i64 [[TMP10]]
+; CHECK-NEXT:    store double [[TMP8]], ptr [[TMP13]], align 8
+; CHECK-NEXT:    store double [[TMP8]], ptr [[TMP14]], align 8
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; CHECK-NEXT:    [[TMP15:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP15]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+;
 entry:
-  %cmp3 = icmp eq i32 %N, 0
-  br i1 %cmp3, label %for.end, label %for.body.preheader
-
-for.body.preheader:                               ; preds = %entry
-  br label %for.body
-
-for.body:                                         ; preds = %for.body.preheader, %for.body
-  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
-  %0 = load float, ptr @inc, align 4
-  %arrayidx = getelementptr inbounds float, ptr %A, i64 %indvars.iv
-  %1 = load float, ptr %arrayidx, align 4
-  %add = fadd float %0, %1
-  store float %add, ptr %arrayidx, align 4
-  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
-  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
-  %exitcond = icmp eq i32 %lftr.wideiv, %N
-  br i1 %exitcond, label %for.end.loopexit, label %for.body
-
-for.end.loopexit:                                 ; preds = %for.body
-  br label %for.end
-
-for.end:                                          ; preds = %for.end.loopexit, %entry
+  br label %loop
+
+loop:
+  %iv.1 = phi i64 [ 0, %entry ], [ %iv.1.next, %loop ]
+  %iv.2 = phi i64 [ %start, %entry ], [ %iv.2.next, %loop ]
+  %l = load double, ptr %src, align 8
+  %m = fmul double %l, 9.0
+  %div = fdiv double %m, %d
+  %sub = sub i64 %iv.1, 1
+  %gep.1 = getelementptr double, ptr %dst, i64 %iv.1
+  %gep.2 = getelementptr double, ptr %gep.1, i64 %sub
+  store double %div, ptr %gep.2, align 8
+  %iv.1.next = add i64 %iv.1, 1
+  %iv.2.next = add i64 %iv.2, -1
+  %ec = icmp sgt i64 %iv.2, 0
+  br i1 %ec , label %loop, label %exit
+
+exit:
   ret void
 }
+
+attributes #0 = { "target-cpu"="core-avx2" }