[llvm] [AArch64] Tweak fixed-length loop.dependence.mask costs (PR #175538)

[Benjamin Maxwell via llvm-commits]

https://github.com/MacDue created https://github.com/llvm/llvm-project/pull/175538

It's not free (MOV + XTN) to convert from the predicate result of whilewr/rw to a fixed-length mask, so the cost should be slightly higher.

>From 81a15a4e39102ab1b81d51462db39fd8c67e1b54 Mon Sep 17 00:00:00 2001
From: Benjamin Maxwell <benjamin.maxwell at arm.com>
Date: Mon, 12 Jan 2026 13:37:22 +0000
Subject: [PATCH] [AArch64] Tweak fixed-length loop.dependence.mask costs

It's not free (MOV + XTN) to convert from the predicate result of
whilewr/rw to a fixed-length mask, so the cost should be slightly
higher.
---
 .../AArch64/AArch64TargetTransformInfo.cpp       | 12 +++++++++---
 .../CostModel/AArch64/loop_dependence_mask.ll    | 16 ++++++++--------
 2 files changed, 17 insertions(+), 11 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 85be8db9d3ae2..59f782b986a4f 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -1071,9 +1071,15 @@ AArch64TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
       EVT VecVT = getTLI()->getValueType(DL, RetTy);
       unsigned EltSizeInBytes =
           cast<ConstantInt>(ICA.getArgs()[2])->getZExtValue();
-      if (is_contained({1u, 2u, 4u, 8u}, EltSizeInBytes) &&
-          VecVT.getVectorMinNumElements() == (16 / EltSizeInBytes))
-        return 1;
+      if (!is_contained({1u, 2u, 4u, 8u}, EltSizeInBytes) ||
+          VecVT.getVectorMinNumElements() != (16 / EltSizeInBytes))
+        break;
+      InstructionCost Cost = 1;
+      // For fixed-vector types at least a MOV and XTN are needed to convert
+      // from the predicate to a fixed-length mask.
+      if (isa<FixedVectorType>(RetTy))
+        Cost += 2;
+      return Cost;
     }
     break;
   }
diff --git a/llvm/test/Analysis/CostModel/AArch64/loop_dependence_mask.ll b/llvm/test/Analysis/CostModel/AArch64/loop_dependence_mask.ll
index 5b3070fcf347e..7acd776a91b3c 100644
--- a/llvm/test/Analysis/CostModel/AArch64/loop_dependence_mask.ll
+++ b/llvm/test/Analysis/CostModel/AArch64/loop_dependence_mask.ll
@@ -17,10 +17,10 @@ define void @loop_dependence_war_mask(ptr %a, ptr %b) {
 ; CHECK-EXPANDED-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
 ; CHECK-LABEL: 'loop_dependence_war_mask'
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res1 = call <16 x i1> @llvm.loop.dependence.war.mask.v16i1(ptr %a, ptr %b, i64 1)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res2 = call <8 x i1> @llvm.loop.dependence.war.mask.v8i1(ptr %a, ptr %b, i64 2)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res3 = call <4 x i1> @llvm.loop.dependence.war.mask.v4i1(ptr %a, ptr %b, i64 4)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res4 = call <2 x i1> @llvm.loop.dependence.war.mask.v2i1(ptr %a, ptr %b, i64 8)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %res1 = call <16 x i1> @llvm.loop.dependence.war.mask.v16i1(ptr %a, ptr %b, i64 1)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %res2 = call <8 x i1> @llvm.loop.dependence.war.mask.v8i1(ptr %a, ptr %b, i64 2)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %res3 = call <4 x i1> @llvm.loop.dependence.war.mask.v4i1(ptr %a, ptr %b, i64 4)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %res4 = call <2 x i1> @llvm.loop.dependence.war.mask.v2i1(ptr %a, ptr %b, i64 8)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res5 = call <vscale x 16 x i1> @llvm.loop.dependence.war.mask.nxv16i1(ptr %a, ptr %b, i64 1)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res6 = call <vscale x 8 x i1> @llvm.loop.dependence.war.mask.nxv8i1(ptr %a, ptr %b, i64 2)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res7 = call <vscale x 4 x i1> @llvm.loop.dependence.war.mask.nxv4i1(ptr %a, ptr %b, i64 4)
@@ -54,10 +54,10 @@ define void @loop_dependence_raw_mask(ptr %a, ptr %b) {
 ; CHECK-EXPANDED-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret void
 ;
 ; CHECK-LABEL: 'loop_dependence_raw_mask'
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res1 = call <16 x i1> @llvm.loop.dependence.raw.mask.v16i1(ptr %a, ptr %b, i64 1)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res2 = call <8 x i1> @llvm.loop.dependence.raw.mask.v8i1(ptr %a, ptr %b, i64 2)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res3 = call <4 x i1> @llvm.loop.dependence.raw.mask.v4i1(ptr %a, ptr %b, i64 4)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res4 = call <2 x i1> @llvm.loop.dependence.raw.mask.v2i1(ptr %a, ptr %b, i64 8)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %res1 = call <16 x i1> @llvm.loop.dependence.raw.mask.v16i1(ptr %a, ptr %b, i64 1)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %res2 = call <8 x i1> @llvm.loop.dependence.raw.mask.v8i1(ptr %a, ptr %b, i64 2)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %res3 = call <4 x i1> @llvm.loop.dependence.raw.mask.v4i1(ptr %a, ptr %b, i64 4)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %res4 = call <2 x i1> @llvm.loop.dependence.raw.mask.v2i1(ptr %a, ptr %b, i64 8)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res5 = call <vscale x 16 x i1> @llvm.loop.dependence.raw.mask.nxv16i1(ptr %a, ptr %b, i64 1)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res6 = call <vscale x 8 x i1> @llvm.loop.dependence.raw.mask.nxv8i1(ptr %a, ptr %b, i64 2)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %res7 = call <vscale x 4 x i1> @llvm.loop.dependence.raw.mask.nxv4i1(ptr %a, ptr %b, i64 4)