[llvm] 31711c9 - [VPlan] Only apply forced cost to recipes with underlying values. (#168372)

[via llvm-commits]

Author: Florian Hahn
Date: 2025-11-21T14:21:16Z
New Revision: 31711c908fbd391601122c3457e71d0714fe4117

URL: https://github.com/llvm/llvm-project/commit/31711c908fbd391601122c3457e71d0714fe4117
DIFF: https://github.com/llvm/llvm-project/commit/31711c908fbd391601122c3457e71d0714fe4117.diff

LOG: [VPlan] Only apply forced cost to recipes with underlying values. (#168372)

Only apply forced instruction costs to recipes with underlying values to
match the legacy cost model. A VPlan may have a number of additional
VPInstructions without underlying values that are not considered for its
cost, and assigning forced costs to them would incorrectly inflate its
cost.

This fixes a cost divergence between legacy and VPlan-based cost models
with forced instruction costs.

PR: https://github.com/llvm/llvm-project/pull/168372

Added: 
    

Modified: 
    llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
    llvm/test/Transforms/LoopVectorize/AArch64/force-target-instruction-cost.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index e8e1a8d4eedde..535c77b048a7b 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -277,9 +277,14 @@ InstructionCost VPRecipeBase::cost(ElementCount VF, VPCostContext &Ctx) {
     RecipeCost = 0;
   } else {
     RecipeCost = computeCost(VF, Ctx);
-    if (UI && ForceTargetInstructionCost.getNumOccurrences() > 0 &&
-        RecipeCost.isValid())
-      RecipeCost = InstructionCost(ForceTargetInstructionCost);
+    RecipeCost = computeCost(VF, Ctx);
+    if (ForceTargetInstructionCost.getNumOccurrences() > 0 &&
+        RecipeCost.isValid()) {
+      if (UI)
+        RecipeCost = InstructionCost(ForceTargetInstructionCost);
+      else
+        RecipeCost = InstructionCost(0);
+    }
   }
 
   LLVM_DEBUG({

diff  --git a/llvm/test/Transforms/LoopVectorize/AArch64/force-target-instruction-cost.ll b/llvm/test/Transforms/LoopVectorize/AArch64/force-target-instruction-cost.ll
index 29bbd015eed1f..56edee44fe3b1 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/force-target-instruction-cost.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/force-target-instruction-cost.ll
@@ -380,6 +380,59 @@ for.end:
   ret void
 }
 
+define void @loop_with_freeze_and_conditional_srem(ptr %dst, ptr %keyinfo, ptr %invariant.ptr, i32 %divisor) #1 {
+; COMMON-LABEL: define void @loop_with_freeze_and_conditional_srem(
+; COMMON-SAME: ptr [[DST:%.*]], ptr [[KEYINFO:%.*]], ptr [[INVARIANT_PTR:%.*]], i32 [[DIVISOR:%.*]]) {
+; COMMON-NEXT:  [[ENTRY:.*]]:
+; COMMON-NEXT:    br label %[[LOOP:.*]]
+; COMMON:       [[LOOP]]:
+; COMMON-NEXT:    [[INDEX_NEXT:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ]
+; COMMON-NEXT:    [[LOADED:%.*]] = load i32, ptr [[INVARIANT_PTR]], align 4
+; COMMON-NEXT:    [[FROZEN:%.*]] = freeze i32 [[LOADED]]
+; COMMON-NEXT:    [[CMP:%.*]] = icmp eq i32 [[FROZEN]], 0
+; COMMON-NEXT:    br i1 [[CMP]], label %[[IF_ZERO:.*]], label %[[IF_NONZERO:.*]]
+; COMMON:       [[IF_ZERO]]:
+; COMMON-NEXT:    store i32 0, ptr [[KEYINFO]], align 4
+; COMMON-NEXT:    br label %[[LOOP_LATCH]]
+; COMMON:       [[IF_NONZERO]]:
+; COMMON-NEXT:    [[TMP11:%.*]] = srem i32 1, [[DIVISOR]]
+; COMMON-NEXT:    store i32 [[TMP11]], ptr [[DST]], align 4
+; COMMON-NEXT:    br label %[[LOOP_LATCH]]
+; COMMON:       [[LOOP_LATCH]]:
+; COMMON-NEXT:    [[IV_NEXT]] = add i64 [[INDEX_NEXT]], 1
+; COMMON-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], 32
+; COMMON-NEXT:    br i1 [[TMP16]], label %[[EXIT:.*]], label %[[LOOP]]
+; COMMON:       [[EXIT]]:
+; COMMON-NEXT:    ret void
+;
+entry:
+  br label %loop
+
+loop:                                             ; preds = %loop.latch, %entry
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.latch ]
+  %loaded = load i32, ptr %invariant.ptr, align 4
+  %frozen = freeze i32 %loaded
+  %cmp = icmp eq i32 %frozen, 0
+  br i1 %cmp, label %if.zero, label %if.nonzero
+
+if.zero:                                          ; preds = %loop
+  store i32 0, ptr %keyinfo, align 4
+  br label %loop.latch
+
+if.nonzero:                                       ; preds = %loop
+  %rem = srem i32 1, %divisor
+  store i32 %rem, ptr %dst, align 4
+  br label %loop.latch
+
+loop.latch:                                       ; preds = %if.nonzero, %if.zero
+  %iv.next = add i64 %iv, 1
+  %exitcond = icmp eq i64 %iv, 32
+  br i1 %exitcond, label %exit, label %loop
+
+exit:                                             ; preds = %loop.latch
+  ret void
+}
+
 attributes #0 = { "target-features"="+neon,+sve" vscale_range(1,16) }
 
 declare void @llvm.assume(i1 noundef)