[llvm] [DSE] Add predicated vector length store support for masked store elimination (PR #134175)

[Michael Berg via llvm-commits]

https://github.com/mcberg2021 updated https://github.com/llvm/llvm-project/pull/134175

>From 5d1d22acb4abf5b08585689b667410acbfc9fff6 Mon Sep 17 00:00:00 2001
From: Michael Berg <michael.berg at sifive.com>
Date: Mon, 31 Mar 2025 14:40:55 -0700
Subject: [PATCH] [DSE] Add predicated vector length store support for masked
 store elimination

In isMaskedStoreOverwrite we process two stores that fully overwrite one
another, here we add support for predicated vector length stores so that
DSE will eliminate this variant of masked stores.
---
 .../Scalar/DeadStoreElimination.cpp           | 35 +++++++++++++++++++
 .../DeadStoreElimination/dead-vp.store.ll     | 34 ++++++++++++++++++
 2 files changed, 69 insertions(+)
 create mode 100644 llvm/test/Transforms/DeadStoreElimination/dead-vp.store.ll

diff --git a/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp b/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp
index 935f21fd484f3..0c5cc2a5487d2 100644
--- a/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp
+++ b/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp
@@ -269,6 +269,41 @@ static OverwriteResult isMaskedStoreOverwrite(const Instruction *KillingI,
       return OW_Unknown;
     return OW_Complete;
   }
+  if (KillingII->getIntrinsicID() == Intrinsic::vp_store) {
+    // Operands {0        , 1     , 2   , 3 }
+    //          {StoredVal, VecPtr, Mask, VL}
+    // Types.
+    VectorType *KillingTy =
+        cast<VectorType>(KillingII->getArgOperand(0)->getType());
+    VectorType *DeadTy = cast<VectorType>(DeadII->getArgOperand(0)->getType());
+    if (KillingTy->getScalarSizeInBits() != DeadTy->getScalarSizeInBits())
+      return OW_Unknown;
+    // Element count.
+    if (KillingTy->getElementCount() != DeadTy->getElementCount())
+      return OW_Unknown;
+    // Pointers.
+    Value *KillingPtr = KillingII->getArgOperand(1)->stripPointerCasts();
+    Value *DeadPtr = DeadII->getArgOperand(1)->stripPointerCasts();
+    if (KillingPtr != DeadPtr && !AA.isMustAlias(KillingPtr, DeadPtr))
+      return OW_Unknown;
+    // Masks.
+    // TODO: check that KillingII's mask is a superset of the DeadII's mask.
+    if (KillingII->getArgOperand(2) != DeadII->getArgOperand(2))
+      return OW_Unknown;
+    // Lengths.
+    if (KillingII->getArgOperand(3) != DeadII->getArgOperand(3))
+      return OW_Unknown;
+    AAMDNodes KillingAA = KillingII->getAAMetadata();
+    AAMDNodes DeadAA = DeadII->getAAMetadata();
+    // There must be scoped noalias metadata on both stores.
+    if (!KillingAA.Scope || !DeadAA.Scope || !KillingAA.NoAlias ||
+        !DeadAA.NoAlias)
+      return OW_Unknown;
+    // Check that both stores have the same scope and noalias metadata.
+    if (KillingAA.Scope != DeadAA.Scope || KillingAA.NoAlias != DeadAA.NoAlias)
+      return OW_Unknown;
+    return OW_Complete;
+  }
   return OW_Unknown;
 }
 
diff --git a/llvm/test/Transforms/DeadStoreElimination/dead-vp.store.ll b/llvm/test/Transforms/DeadStoreElimination/dead-vp.store.ll
new file mode 100644
index 0000000000000..825acd16a9d2a
--- /dev/null
+++ b/llvm/test/Transforms/DeadStoreElimination/dead-vp.store.ll
@@ -0,0 +1,34 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -passes=dse -S < %s | FileCheck %s
+target triple = "riscv64-unknown-linux-gnu"
+
+; Test predicated vector length masked stores for elimination
+
+define void @foo(ptr %a, i32 %vl, <vscale x 8 x i32> %v1, <vscale x 8 x i32> %v2) {
+;
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:    [[VP_OP:%.*]] = call <vscale x 8 x i32> @llvm.vp.add.nxv8i32(<vscale x 8 x i32> [[V1:%.*]], <vscale x 8 x i32> [[V2:%.*]], <vscale x 8 x i1> splat (i1 true), i32 [[VL:%.*]])
+; CHECK-NEXT:    call void @llvm.vp.store.nxv8i32.p0(<vscale x 8 x i32> [[VP_OP]], ptr nonnull [[A:%.*]], <vscale x 8 x i1> splat (i1 true), i32 [[VL]]), !alias.scope [[META0:![0-9]+]], !noalias [[META5:![0-9]+]]
+; CHECK-NEXT:    ret void
+;
+  call void @llvm.vp.store.nxv8i32.p0(<vscale x 8 x i32> %v1, ptr nonnull %a, <vscale x 8 x i1> shufflevector (<vscale x 8 x i1> insertelement (<vscale x 8 x i1> poison, i1 true, i32 0), <vscale x 8 x i1> poison, <vscale x 8 x i32> zeroinitializer), i32 %vl), !tbaa !16, !alias.scope !34, !noalias !37
+  %vp.op = call <vscale x 8 x i32> @llvm.vp.add.nxv8i32(<vscale x 8 x i32> %v1, <vscale x 8 x i32> %v2, <vscale x 8 x i1> shufflevector (<vscale x 8 x i1> insertelement (<vscale x 8 x i1> poison, i1 true, i32 0), <vscale x 8 x i1> poison, <vscale x 8 x i32> zeroinitializer), i32 %vl)
+  call void @llvm.vp.store.nxv8i32.p0(<vscale x 8 x i32> %vp.op, ptr nonnull %a, <vscale x 8 x i1> shufflevector (<vscale x 8 x i1> insertelement (<vscale x 8 x i1> poison, i1 true, i32 0), <vscale x 8 x i1> poison, <vscale x 8 x i32> zeroinitializer), i32 %vl), !alias.scope !34, !noalias !37
+  ret void
+}
+
+declare <vscale x 8 x i32> @llvm.vp.add.nxv8i32(<vscale x 8 x i32>, <vscale x 8 x i32>, <vscale x 8 x i1>, i32)
+declare void @llvm.vp.store.nxv8i32.p0(<vscale x 8 x i32>, ptr nocapture, <vscale x 8 x i1>, i32)
+
+!11 = !{!"omnipotent char", !12, i64 0}
+!12 = !{!"Simple C/C++ TBAA"}
+!13 = !{!"int", !11, i64 0}
+!16 = !{!13, !13, i64 0}
+!28 = distinct !{!28, !"LVerDomain"}
+!30 = distinct !{!30, !"LVerDomain"}
+!34 = !{!35, !36}
+!35 = distinct !{!35, !28}
+!36 = distinct !{!36, !30}
+!37 = !{!38, !39}
+!38 = distinct !{!38, !28}
+!39 = distinct !{!39, !28}