[VPlan] Introduce VPGEPInstruction by artagnon · Pull Request #193510 · llvm/llvm-project

artagnon · 2026-04-22T14:22:13Z

Introduce a first-class VPGEPInstruction, similar to the existing "scalar cast", with the additional benefit of being able to query the source element type of the GEP directly without going to the underlying value. Planned follow-ups include unifying WidenGEP and Replicate GEPs with this.

llvmbot · 2026-04-22T14:22:49Z

@llvm/pr-subscribers-vectorizers

Author: Ramkumar Ramachandra (artagnon)

Changes

Currently, WidenGEP, Replicate, and VPInstruction recipes can all hold a Instruction::GetElementPtr. Introduce a first-class "scalar GEP" VPInstructionWithType, similar to the existing "scalar cast", with the additional benefit of being able to query the source element type of the GEP directly without going to the underlying value. Planned follow-ups include unifying WidenGEP and Replicate GEPs with this VPInstruction GEP.

Full diff: https://github.com/llvm/llvm-project/pull/193510.diff

8 Files Affected:

(modified) llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h (+8)
(modified) llvm/lib/Transforms/Vectorize/VPlan.h (+23-1)
(modified) llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp (+6)
(modified) llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp (+22-1)
(modified) llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp (+22-18)
(modified) llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll (+7-7)
(modified) llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll (+3-3)
(modified) llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll (+2-2)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h b/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h
index f06b959700687..fd859fcefe7df 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h
@@ -364,6 +364,14 @@ class VPBuilder {
         new VPInstructionWithType(Opcode, Op, ResultTy, Flags, Metadata, DL));
   }
 
+  VPInstruction *createScalarGEP(Type *SourceElementTy, ArrayRef<VPValue *> Ops,
+                                 Type *ResultTy, DebugLoc DL,
+                                 const VPIRFlags &Flags,
+                                 const VPIRMetadata &Metadata = {}) {
+    return tryInsertInstruction(new VPInstructionWithType(
+        SourceElementTy, Ops, ResultTy, Flags, Metadata, DL));
+  }
+
   VPValue *createScalarZExtOrTrunc(VPValue *Op, Type *ResultTy, Type *SrcTy,
                                    DebugLoc DL) {
     if (ResultTy == SrcTy)
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 934cca006e91c..005cae1c94552 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -556,6 +556,9 @@ class LLVM_ABI_FOR_TEST VPRecipeBase
   /// Return true if the recipe is a scalar cast.
   bool isScalarCast() const;
 
+  /// Return true if the recipe is a scalar getelementptr.
+  bool isScalarGEP() const;
+
   /// Set the recipe's debug location to \p NewDL.
   void setDebugLoc(DebugLoc NewDL) { DL = NewDL; }
 
@@ -1517,6 +1520,10 @@ class VPInstructionWithType : public VPInstruction {
   /// Scalar result type produced by the recipe.
   Type *ResultTy;
 
+  /// The source element type, which is present when the recipe has a
+  /// getelementptr opcode.
+  Type *SourceElementTy = nullptr;
+
 public:
   VPInstructionWithType(unsigned Opcode, ArrayRef<VPValue *> Operands,
                         Type *ResultTy, const VPIRFlags &Flags = {},
@@ -1526,10 +1533,20 @@ class VPInstructionWithType : public VPInstruction {
       : VPInstruction(Opcode, Operands, Flags, Metadata, DL, Name),
         ResultTy(ResultTy) {}
 
+  /// Constructor for GEPs.
+  VPInstructionWithType(Type *SourceElementTy, ArrayRef<VPValue *> Operands,
+                        Type *ResultTy, const VPIRFlags &Flags = {},
+                        const VPIRMetadata &Metadata = {},
+                        DebugLoc DL = DebugLoc::getUnknown(),
+                        const Twine &Name = "")
+      : VPInstruction(Instruction::GetElementPtr, Operands, Flags, Metadata, DL,
+                      Name),
+        ResultTy(ResultTy), SourceElementTy(SourceElementTy) {}
+
   static inline bool classof(const VPRecipeBase *R) {
     // VPInstructionWithType are VPInstructions with specific opcodes requiring
     // type information.
-    if (R->isScalarCast())
+    if (R->isScalarCast() || R->isScalarGEP())
       return true;
     auto *VPI = dyn_cast<VPInstruction>(R);
     if (!VPI)
@@ -1567,6 +1584,11 @@ class VPInstructionWithType : public VPInstruction {
   }
 
   Type *getResultType() const { return ResultTy; }
+  Type *getSourceElementType() const {
+    assert(isScalarGEP() &&
+           "Source element type requested for non-getelementptr");
+    return SourceElementTy;
+  }
 
 protected:
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
index 9b052047dcc8b..6cbe36b44af58 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
@@ -267,6 +267,12 @@ void PlainCFGBuilder::createVPInstructionsForVPBB(VPBasicBlock *VPBB,
                                             CI->getType(), CI->getDebugLoc(),
                                             VPIRFlags(*CI), MD);
         NewR->setUnderlyingValue(CI);
+      } else if (auto *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
+        NewR = VPIRBuilder.createScalarGEP(
+            GEP->getSourceElementType(), VPOperands,
+            GEP->getPointerOperandType(), GEP->getDebugLoc(), VPIRFlags(*GEP),
+            MD);
+        NewR->setUnderlyingValue(GEP);
       } else if (auto *LI = dyn_cast<LoadInst>(Inst)) {
         NewR = VPIRBuilder.createScalarLoad(LI->getType(), VPOperands[0],
                                             LI->getDebugLoc(), MD);
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 25fe37124b017..573fa419b4bf1 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -317,6 +317,11 @@ bool VPRecipeBase::isScalarCast() const {
   return VPI && Instruction::isCast(VPI->getOpcode());
 }
 
+bool VPRecipeBase::isScalarGEP() const {
+  auto *VPI = dyn_cast<VPInstruction>(this);
+  return VPI && VPI->getOpcode() == Instruction::GetElementPtr;
+}
+
 void VPIRFlags::intersectFlags(const VPIRFlags &Other) {
   assert(OpType == Other.OpType && "OpType must match");
   switch (OpType) {
@@ -1291,7 +1296,7 @@ bool VPInstruction::isSingleScalar() const {
   case VPInstruction::VScale:
     return true;
   default:
-    return isScalarCast();
+    return isScalarCast() || isScalarGEP();
   }
 }
 
@@ -1583,6 +1588,16 @@ void VPInstructionWithType::execute(VPTransformState &State) {
     State.set(this, Cast, VPLane(0));
     return;
   }
+  if (isScalarGEP()) {
+    Value *Ptr = State.get(getOperand(0), true);
+    auto IdxList =
+        to_vector(map_range(drop_begin(operands()),
+                            [&](VPValue *Op) { return State.get(Op, true); }));
+    Value *GEP = State.Builder.CreateGEP(ResultTy, Ptr, IdxList, "",
+                                         getGEPNoWrapFlags());
+    State.set(this, GEP, true);
+    return;
+  }
   switch (getOpcode()) {
   case VPInstruction::StepVector: {
     Value *StepVector =
@@ -1623,6 +1638,12 @@ void VPInstructionWithType::printRecipe(raw_ostream &O, const Twine &Indent,
     O << "load ";
     printOperands(O, SlotTracker);
     break;
+  case Instruction::GetElementPtr:
+    O << "getelementptr";
+    printFlags(O);
+    O << *ResultTy << " ";
+    printOperands(O, SlotTracker);
+    break;
   default:
     assert(Instruction::isCast(getOpcode()) && "unhandled opcode");
     O << Instruction::getOpcodeName(getOpcode()) << " ";
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 2234b2c5d56a3..39187c613f620 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1227,6 +1227,27 @@ getOpcodeOrIntrinsicID(const VPSingleDefRecipe *R) {
       .Default([](auto *) { return std::nullopt; });
 }
 
+/// If recipe \p R will lower to a GEP with a non-i8 source element type,
+/// return that source element type.
+static Type *getGEPSourceElementType(const VPSingleDefRecipe *R) {
+  // All VPInstructions that lower to GEPs must have the i8 source element
+  // type (as they are PtrAdds), so we omit it.
+  return TypeSwitch<const VPSingleDefRecipe *, Type *>(R)
+      .Case([](const VPReplicateRecipe *I) -> Type * {
+        if (auto *GEP = dyn_cast<GetElementPtrInst>(I->getUnderlyingValue()))
+          return GEP->getSourceElementType();
+        return nullptr;
+      })
+      .Case<VPVectorPointerRecipe, VPWidenGEPRecipe>(
+          [](auto *I) { return I->getSourceElementType(); })
+      .Case<VPInstructionWithType>([](auto *I) {
+        return I->getOpcode() == Instruction::GetElementPtr
+                   ? I->getSourceElementType()
+                   : nullptr;
+      })
+      .Default([](auto *) { return nullptr; });
+}
+
 /// Try to fold \p R using InstSimplifyFolder. Will succeed and return a
 /// non-nullptr VPValue for a handled opcode or intrinsic ID if corresponding \p
 /// Operands are foldable live-ins.
@@ -1279,8 +1300,7 @@ static VPIRValue *tryToFoldLiveIns(VPSingleDefRecipe &R,
                             Ops[1]);
     case Instruction::GetElementPtr: {
       auto &RFlags = cast<VPRecipeWithIRFlags>(R);
-      auto *GEP = cast<GetElementPtrInst>(RFlags.getUnderlyingInstr());
-      return Folder.FoldGEP(GEP->getSourceElementType(), Ops[0],
+      return Folder.FoldGEP(getGEPSourceElementType(&RFlags), Ops[0],
                             drop_begin(Ops), RFlags.getGEPNoWrapFlags());
     }
     case VPInstruction::PtrAdd:
@@ -2577,22 +2597,6 @@ struct VPCSEDenseMapInfo : public DenseMapInfo<VPSingleDefRecipe *> {
     return Def == getEmptyKey() || Def == getTombstoneKey();
   }
 
-  /// If recipe \p R will lower to a GEP with a non-i8 source element type,
-  /// return that source element type.
-  static Type *getGEPSourceElementType(const VPSingleDefRecipe *R) {
-    // All VPInstructions that lower to GEPs must have the i8 source element
-    // type (as they are PtrAdds), so we omit it.
-    return TypeSwitch<const VPSingleDefRecipe *, Type *>(R)
-        .Case([](const VPReplicateRecipe *I) -> Type * {
-          if (auto *GEP = dyn_cast<GetElementPtrInst>(I->getUnderlyingValue()))
-            return GEP->getSourceElementType();
-          return nullptr;
-        })
-        .Case<VPVectorPointerRecipe, VPWidenGEPRecipe>(
-            [](auto *I) { return I->getSourceElementType(); })
-        .Default([](auto *) { return nullptr; });
-  }
-
   /// Returns true if recipe \p Def can be safely handed for CSE.
   static bool canHandle(const VPSingleDefRecipe *Def) {
     // We can extend the list of handled recipes in the future,
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll b/llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll
index 2857fcd246cef..da5789714b852 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll
@@ -8,7 +8,7 @@ define void @diamond_phi(ptr %a) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0:%[0-9]+]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT ir<%c0> = icmp sle ir<%iv>, ir<0>
 ; CHECK-NEXT:    Successor(s): bb2
 ; CHECK-EMPTY:
@@ -73,7 +73,7 @@ define void @mask_reuse(ptr %a) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0:%[0-9]+]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT ir<%c0> = icmp sle ir<%iv>, ir<0>
 ; CHECK-NEXT:      EMIT ir<%add0> = add ir<%iv>, ir<0>
 ; CHECK-NEXT:    Successor(s): bb1
@@ -159,7 +159,7 @@ define void @optimized_mask(ptr %a) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0:%[0-9]+]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT ir<%c0> = icmp sle ir<%iv>, ir<0>
 ; CHECK-NEXT:    Successor(s): bb6
 ; CHECK-EMPTY:
@@ -285,7 +285,7 @@ define void @switch(ptr %a) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0:%[0-9]+]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT ir<%c0> = icmp sle ir<%iv>, ir<0>
 ; CHECK-NEXT:      EMIT ir<%add0> = add ir<%iv>, ir<0>
 ; CHECK-NEXT:    Successor(s): bb2
@@ -420,7 +420,7 @@ define void @diamond_phi2(ptr %a, i1 %c1, i1 %c2) {
 ; CHECK-NEXT:      EMIT vp<[[VP6:%[0-9]+]]> = logical-and ir<%c0>, ir<%c1>
 ; CHECK-NEXT:      EMIT vp<[[VP7:%[0-9]+]]> = or vp<[[VP5]]>, vp<[[VP6]]>
 ; CHECK-NEXT:      BLEND ir<%phi> = ir<%add2>/vp<[[VP5]]> ir<%add1>/vp<[[VP6]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT store ir<%phi>, ir<%gep>, vp<[[VP7]]>
 ; CHECK-NEXT:    Successor(s): bb5
 ; CHECK-EMPTY:
@@ -521,7 +521,7 @@ define void @blend_masks(ptr noalias %p, i1 %c0, i1 %c1, i1 %c2, i1 %c3, i1 %c4)
 ; CHECK-NEXT:      EMIT vp<[[VP15:%[0-9]+]]> = logical-and vp<[[VP9]]>, ir<%c4>
 ; CHECK-NEXT:      EMIT vp<[[VP16:%[0-9]+]]> = or vp<[[VP15]]>, vp<[[VP14]]>
 ; CHECK-NEXT:      BLEND ir<%phi> = ir<1>/vp<[[VP15]]> ir<0>/vp<[[VP14]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT store ir<%phi>, ir<%gep>, vp<[[VP16]]>
 ; CHECK-NEXT:    Successor(s): bb8
 ; CHECK-EMPTY:
@@ -607,7 +607,7 @@ define void @blend_masks_triangle_phi(ptr noalias %p, i1 %c0, i1 %c1) {
 ; CHECK-NEXT:    bb3:
 ; CHECK-NEXT:      EMIT vp<[[VP8:%[0-9]+]]> = logical-and ir<%c0>, ir<%c1>
 ; CHECK-NEXT:      BLEND ir<%phi> = ir<1>/vp<[[VP7]]> ir<0>/vp<[[VP8]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT store ir<%phi>, ir<%gep>
 ; CHECK-NEXT:      EMIT ir<%iv.next> = add ir<%iv>, ir<1>
 ; CHECK-NEXT:      EMIT ir<%ec> = icmp eq ir<%iv.next>, ir<128>
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll b/llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll
index d02fd486ec609..741bfd4d90822 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll
@@ -27,7 +27,7 @@ define i32 @live_out(ptr noalias %p, i32 %n) {
 ; CHECK-NEXT:    Successor(s): vector.body.split, vector.latch
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body.split:
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT-SCALAR ir<%x> = load ir<%gep>
 ; CHECK-NEXT:      EMIT ir<%y> = add ir<%x>, ir<1>
 ; CHECK-NEXT:      EMIT store ir<%y>, ir<%gep>
@@ -118,7 +118,7 @@ define i32 @conditional_live_out(ptr noalias %p, i32 %n, i1 %c) {
 ; CHECK-NEXT:    Successor(s): if, latch
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    if:
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT-SCALAR ir<%x> = load ir<%gep>
 ; CHECK-NEXT:      EMIT ir<%y> = add ir<%x>, ir<1>
 ; CHECK-NEXT:      EMIT store ir<%y>, ir<%gep>
@@ -284,7 +284,7 @@ define i32 @reduction(ptr noalias %p, i32 %n) {
 ; CHECK-NEXT:    Successor(s): vector.body.split, vector.latch
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body.split:
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT-SCALAR ir<%x> = load ir<%gep>
 ; CHECK-NEXT:      EMIT ir<%rdx.next> = add ir<%rdx>, ir<%x>
 ; CHECK-NEXT:      EMIT ir<%iv.next> = add ir<%iv>, ir<1>
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll b/llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll
index 6eff6d1df3523..a27d0d9ecd227 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll
@@ -15,14 +15,14 @@ define void @foo(i64 %n) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:  outer.header:
 ; CHECK-NEXT:    EMIT-SCALAR ir<%outer.iv> = phi [ ir<%outer.iv.next>, outer.latch ], [ ir<0>, ir-bb<entry> ]
-; CHECK-NEXT:    EMIT ir<%gep.1> = getelementptr inbounds ir<@arr2>, ir<0>, ir<%outer.iv>
+; CHECK-NEXT:    EMIT-SCALAR ir<%gep.1> = getelementptr inbounds ptr ir<@arr2>, ir<0>, ir<%outer.iv>
 ; CHECK-NEXT:    EMIT store ir<%outer.iv>, ir<%gep.1>
 ; CHECK-NEXT:    EMIT ir<%add> = add nsw ir<%outer.iv>, ir<%n>
 ; CHECK-NEXT:  Successor(s): inner
 ; CHECK-EMPTY:
 ; CHECK-NEXT:  inner:
 ; CHECK-NEXT:    EMIT-SCALAR ir<%inner.iv> = phi [ ir<%inner.iv.next>, inner ], [ ir<0>, outer.header ]
-; CHECK-NEXT:    EMIT ir<%gep.2> = getelementptr inbounds ir<@arr>, ir<0>, ir<%inner.iv>, ir<%outer.iv>
+; CHECK-NEXT:    EMIT-SCALAR ir<%gep.2> = getelementptr inbounds ptr ir<@arr>, ir<0>, ir<%inner.iv>, ir<%outer.iv>
 ; CHECK-NEXT:    EMIT store ir<%add>, ir<%gep.2>
 ; CHECK-NEXT:    EMIT ir<%inner.iv.next> = add nuw nsw ir<%inner.iv>, ir<1>
 ; CHECK-NEXT:    EMIT ir<%inner.ec> = icmp eq ir<%inner.iv.next>, ir<8>

llvmbot · 2026-04-22T14:22:49Z

@llvm/pr-subscribers-llvm-transforms

Author: Ramkumar Ramachandra (artagnon)

Changes

Currently, WidenGEP, Replicate, and VPInstruction recipes can all hold a Instruction::GetElementPtr. Introduce a first-class "scalar GEP" VPInstructionWithType, similar to the existing "scalar cast", with the additional benefit of being able to query the source element type of the GEP directly without going to the underlying value. Planned follow-ups include unifying WidenGEP and Replicate GEPs with this VPInstruction GEP.

Full diff: https://github.com/llvm/llvm-project/pull/193510.diff

8 Files Affected:

(modified) llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h (+8)
(modified) llvm/lib/Transforms/Vectorize/VPlan.h (+23-1)
(modified) llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp (+6)
(modified) llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp (+22-1)
(modified) llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp (+22-18)
(modified) llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll (+7-7)
(modified) llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll (+3-3)
(modified) llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll (+2-2)

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h b/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h
index f06b959700687..fd859fcefe7df 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h
@@ -364,6 +364,14 @@ class VPBuilder {
         new VPInstructionWithType(Opcode, Op, ResultTy, Flags, Metadata, DL));
   }
 
+  VPInstruction *createScalarGEP(Type *SourceElementTy, ArrayRef<VPValue *> Ops,
+                                 Type *ResultTy, DebugLoc DL,
+                                 const VPIRFlags &Flags,
+                                 const VPIRMetadata &Metadata = {}) {
+    return tryInsertInstruction(new VPInstructionWithType(
+        SourceElementTy, Ops, ResultTy, Flags, Metadata, DL));
+  }
+
   VPValue *createScalarZExtOrTrunc(VPValue *Op, Type *ResultTy, Type *SrcTy,
                                    DebugLoc DL) {
     if (ResultTy == SrcTy)
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 934cca006e91c..005cae1c94552 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -556,6 +556,9 @@ class LLVM_ABI_FOR_TEST VPRecipeBase
   /// Return true if the recipe is a scalar cast.
   bool isScalarCast() const;
 
+  /// Return true if the recipe is a scalar getelementptr.
+  bool isScalarGEP() const;
+
   /// Set the recipe's debug location to \p NewDL.
   void setDebugLoc(DebugLoc NewDL) { DL = NewDL; }
 
@@ -1517,6 +1520,10 @@ class VPInstructionWithType : public VPInstruction {
   /// Scalar result type produced by the recipe.
   Type *ResultTy;
 
+  /// The source element type, which is present when the recipe has a
+  /// getelementptr opcode.
+  Type *SourceElementTy = nullptr;
+
 public:
   VPInstructionWithType(unsigned Opcode, ArrayRef<VPValue *> Operands,
                         Type *ResultTy, const VPIRFlags &Flags = {},
@@ -1526,10 +1533,20 @@ class VPInstructionWithType : public VPInstruction {
       : VPInstruction(Opcode, Operands, Flags, Metadata, DL, Name),
         ResultTy(ResultTy) {}
 
+  /// Constructor for GEPs.
+  VPInstructionWithType(Type *SourceElementTy, ArrayRef<VPValue *> Operands,
+                        Type *ResultTy, const VPIRFlags &Flags = {},
+                        const VPIRMetadata &Metadata = {},
+                        DebugLoc DL = DebugLoc::getUnknown(),
+                        const Twine &Name = "")
+      : VPInstruction(Instruction::GetElementPtr, Operands, Flags, Metadata, DL,
+                      Name),
+        ResultTy(ResultTy), SourceElementTy(SourceElementTy) {}
+
   static inline bool classof(const VPRecipeBase *R) {
     // VPInstructionWithType are VPInstructions with specific opcodes requiring
     // type information.
-    if (R->isScalarCast())
+    if (R->isScalarCast() || R->isScalarGEP())
       return true;
     auto *VPI = dyn_cast<VPInstruction>(R);
     if (!VPI)
@@ -1567,6 +1584,11 @@ class VPInstructionWithType : public VPInstruction {
   }
 
   Type *getResultType() const { return ResultTy; }
+  Type *getSourceElementType() const {
+    assert(isScalarGEP() &&
+           "Source element type requested for non-getelementptr");
+    return SourceElementTy;
+  }
 
 protected:
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
index 9b052047dcc8b..6cbe36b44af58 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
@@ -267,6 +267,12 @@ void PlainCFGBuilder::createVPInstructionsForVPBB(VPBasicBlock *VPBB,
                                             CI->getType(), CI->getDebugLoc(),
                                             VPIRFlags(*CI), MD);
         NewR->setUnderlyingValue(CI);
+      } else if (auto *GEP = dyn_cast<GetElementPtrInst>(Inst)) {
+        NewR = VPIRBuilder.createScalarGEP(
+            GEP->getSourceElementType(), VPOperands,
+            GEP->getPointerOperandType(), GEP->getDebugLoc(), VPIRFlags(*GEP),
+            MD);
+        NewR->setUnderlyingValue(GEP);
       } else if (auto *LI = dyn_cast<LoadInst>(Inst)) {
         NewR = VPIRBuilder.createScalarLoad(LI->getType(), VPOperands[0],
                                             LI->getDebugLoc(), MD);
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 25fe37124b017..573fa419b4bf1 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -317,6 +317,11 @@ bool VPRecipeBase::isScalarCast() const {
   return VPI && Instruction::isCast(VPI->getOpcode());
 }
 
+bool VPRecipeBase::isScalarGEP() const {
+  auto *VPI = dyn_cast<VPInstruction>(this);
+  return VPI && VPI->getOpcode() == Instruction::GetElementPtr;
+}
+
 void VPIRFlags::intersectFlags(const VPIRFlags &Other) {
   assert(OpType == Other.OpType && "OpType must match");
   switch (OpType) {
@@ -1291,7 +1296,7 @@ bool VPInstruction::isSingleScalar() const {
   case VPInstruction::VScale:
     return true;
   default:
-    return isScalarCast();
+    return isScalarCast() || isScalarGEP();
   }
 }
 
@@ -1583,6 +1588,16 @@ void VPInstructionWithType::execute(VPTransformState &State) {
     State.set(this, Cast, VPLane(0));
     return;
   }
+  if (isScalarGEP()) {
+    Value *Ptr = State.get(getOperand(0), true);
+    auto IdxList =
+        to_vector(map_range(drop_begin(operands()),
+                            [&](VPValue *Op) { return State.get(Op, true); }));
+    Value *GEP = State.Builder.CreateGEP(ResultTy, Ptr, IdxList, "",
+                                         getGEPNoWrapFlags());
+    State.set(this, GEP, true);
+    return;
+  }
   switch (getOpcode()) {
   case VPInstruction::StepVector: {
     Value *StepVector =
@@ -1623,6 +1638,12 @@ void VPInstructionWithType::printRecipe(raw_ostream &O, const Twine &Indent,
     O << "load ";
     printOperands(O, SlotTracker);
     break;
+  case Instruction::GetElementPtr:
+    O << "getelementptr";
+    printFlags(O);
+    O << *ResultTy << " ";
+    printOperands(O, SlotTracker);
+    break;
   default:
     assert(Instruction::isCast(getOpcode()) && "unhandled opcode");
     O << Instruction::getOpcodeName(getOpcode()) << " ";
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 2234b2c5d56a3..39187c613f620 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1227,6 +1227,27 @@ getOpcodeOrIntrinsicID(const VPSingleDefRecipe *R) {
       .Default([](auto *) { return std::nullopt; });
 }
 
+/// If recipe \p R will lower to a GEP with a non-i8 source element type,
+/// return that source element type.
+static Type *getGEPSourceElementType(const VPSingleDefRecipe *R) {
+  // All VPInstructions that lower to GEPs must have the i8 source element
+  // type (as they are PtrAdds), so we omit it.
+  return TypeSwitch<const VPSingleDefRecipe *, Type *>(R)
+      .Case([](const VPReplicateRecipe *I) -> Type * {
+        if (auto *GEP = dyn_cast<GetElementPtrInst>(I->getUnderlyingValue()))
+          return GEP->getSourceElementType();
+        return nullptr;
+      })
+      .Case<VPVectorPointerRecipe, VPWidenGEPRecipe>(
+          [](auto *I) { return I->getSourceElementType(); })
+      .Case<VPInstructionWithType>([](auto *I) {
+        return I->getOpcode() == Instruction::GetElementPtr
+                   ? I->getSourceElementType()
+                   : nullptr;
+      })
+      .Default([](auto *) { return nullptr; });
+}
+
 /// Try to fold \p R using InstSimplifyFolder. Will succeed and return a
 /// non-nullptr VPValue for a handled opcode or intrinsic ID if corresponding \p
 /// Operands are foldable live-ins.
@@ -1279,8 +1300,7 @@ static VPIRValue *tryToFoldLiveIns(VPSingleDefRecipe &R,
                             Ops[1]);
     case Instruction::GetElementPtr: {
       auto &RFlags = cast<VPRecipeWithIRFlags>(R);
-      auto *GEP = cast<GetElementPtrInst>(RFlags.getUnderlyingInstr());
-      return Folder.FoldGEP(GEP->getSourceElementType(), Ops[0],
+      return Folder.FoldGEP(getGEPSourceElementType(&RFlags), Ops[0],
                             drop_begin(Ops), RFlags.getGEPNoWrapFlags());
     }
     case VPInstruction::PtrAdd:
@@ -2577,22 +2597,6 @@ struct VPCSEDenseMapInfo : public DenseMapInfo<VPSingleDefRecipe *> {
     return Def == getEmptyKey() || Def == getTombstoneKey();
   }
 
-  /// If recipe \p R will lower to a GEP with a non-i8 source element type,
-  /// return that source element type.
-  static Type *getGEPSourceElementType(const VPSingleDefRecipe *R) {
-    // All VPInstructions that lower to GEPs must have the i8 source element
-    // type (as they are PtrAdds), so we omit it.
-    return TypeSwitch<const VPSingleDefRecipe *, Type *>(R)
-        .Case([](const VPReplicateRecipe *I) -> Type * {
-          if (auto *GEP = dyn_cast<GetElementPtrInst>(I->getUnderlyingValue()))
-            return GEP->getSourceElementType();
-          return nullptr;
-        })
-        .Case<VPVectorPointerRecipe, VPWidenGEPRecipe>(
-            [](auto *I) { return I->getSourceElementType(); })
-        .Default([](auto *) { return nullptr; });
-  }
-
   /// Returns true if recipe \p Def can be safely handed for CSE.
   static bool canHandle(const VPSingleDefRecipe *Def) {
     // We can extend the list of handled recipes in the future,
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll b/llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll
index 2857fcd246cef..da5789714b852 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/predicator.ll
@@ -8,7 +8,7 @@ define void @diamond_phi(ptr %a) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0:%[0-9]+]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT ir<%c0> = icmp sle ir<%iv>, ir<0>
 ; CHECK-NEXT:    Successor(s): bb2
 ; CHECK-EMPTY:
@@ -73,7 +73,7 @@ define void @mask_reuse(ptr %a) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0:%[0-9]+]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT ir<%c0> = icmp sle ir<%iv>, ir<0>
 ; CHECK-NEXT:      EMIT ir<%add0> = add ir<%iv>, ir<0>
 ; CHECK-NEXT:    Successor(s): bb1
@@ -159,7 +159,7 @@ define void @optimized_mask(ptr %a) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0:%[0-9]+]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT ir<%c0> = icmp sle ir<%iv>, ir<0>
 ; CHECK-NEXT:    Successor(s): bb6
 ; CHECK-EMPTY:
@@ -285,7 +285,7 @@ define void @switch(ptr %a) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body:
 ; CHECK-NEXT:      ir<%iv> = WIDEN-INDUCTION nuw nsw ir<0>, ir<1>, vp<[[VP0:%[0-9]+]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT ir<%c0> = icmp sle ir<%iv>, ir<0>
 ; CHECK-NEXT:      EMIT ir<%add0> = add ir<%iv>, ir<0>
 ; CHECK-NEXT:    Successor(s): bb2
@@ -420,7 +420,7 @@ define void @diamond_phi2(ptr %a, i1 %c1, i1 %c2) {
 ; CHECK-NEXT:      EMIT vp<[[VP6:%[0-9]+]]> = logical-and ir<%c0>, ir<%c1>
 ; CHECK-NEXT:      EMIT vp<[[VP7:%[0-9]+]]> = or vp<[[VP5]]>, vp<[[VP6]]>
 ; CHECK-NEXT:      BLEND ir<%phi> = ir<%add2>/vp<[[VP5]]> ir<%add1>/vp<[[VP6]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%a>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%a>, ir<%iv>
 ; CHECK-NEXT:      EMIT store ir<%phi>, ir<%gep>, vp<[[VP7]]>
 ; CHECK-NEXT:    Successor(s): bb5
 ; CHECK-EMPTY:
@@ -521,7 +521,7 @@ define void @blend_masks(ptr noalias %p, i1 %c0, i1 %c1, i1 %c2, i1 %c3, i1 %c4)
 ; CHECK-NEXT:      EMIT vp<[[VP15:%[0-9]+]]> = logical-and vp<[[VP9]]>, ir<%c4>
 ; CHECK-NEXT:      EMIT vp<[[VP16:%[0-9]+]]> = or vp<[[VP15]]>, vp<[[VP14]]>
 ; CHECK-NEXT:      BLEND ir<%phi> = ir<1>/vp<[[VP15]]> ir<0>/vp<[[VP14]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT store ir<%phi>, ir<%gep>, vp<[[VP16]]>
 ; CHECK-NEXT:    Successor(s): bb8
 ; CHECK-EMPTY:
@@ -607,7 +607,7 @@ define void @blend_masks_triangle_phi(ptr noalias %p, i1 %c0, i1 %c1) {
 ; CHECK-NEXT:    bb3:
 ; CHECK-NEXT:      EMIT vp<[[VP8:%[0-9]+]]> = logical-and ir<%c0>, ir<%c1>
 ; CHECK-NEXT:      BLEND ir<%phi> = ir<1>/vp<[[VP7]]> ir<0>/vp<[[VP8]]>
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT store ir<%phi>, ir<%gep>
 ; CHECK-NEXT:      EMIT ir<%iv.next> = add ir<%iv>, ir<1>
 ; CHECK-NEXT:      EMIT ir<%ec> = icmp eq ir<%iv.next>, ir<128>
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll b/llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll
index d02fd486ec609..741bfd4d90822 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/tail-folding.ll
@@ -27,7 +27,7 @@ define i32 @live_out(ptr noalias %p, i32 %n) {
 ; CHECK-NEXT:    Successor(s): vector.body.split, vector.latch
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body.split:
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT-SCALAR ir<%x> = load ir<%gep>
 ; CHECK-NEXT:      EMIT ir<%y> = add ir<%x>, ir<1>
 ; CHECK-NEXT:      EMIT store ir<%y>, ir<%gep>
@@ -118,7 +118,7 @@ define i32 @conditional_live_out(ptr noalias %p, i32 %n, i1 %c) {
 ; CHECK-NEXT:    Successor(s): if, latch
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    if:
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT-SCALAR ir<%x> = load ir<%gep>
 ; CHECK-NEXT:      EMIT ir<%y> = add ir<%x>, ir<1>
 ; CHECK-NEXT:      EMIT store ir<%y>, ir<%gep>
@@ -284,7 +284,7 @@ define i32 @reduction(ptr noalias %p, i32 %n) {
 ; CHECK-NEXT:    Successor(s): vector.body.split, vector.latch
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    vector.body.split:
-; CHECK-NEXT:      EMIT ir<%gep> = getelementptr ir<%p>, ir<%iv>
+; CHECK-NEXT:      EMIT-SCALAR ir<%gep> = getelementptr ptr ir<%p>, ir<%iv>
 ; CHECK-NEXT:      EMIT-SCALAR ir<%x> = load ir<%gep>
 ; CHECK-NEXT:      EMIT ir<%rdx.next> = add ir<%rdx>, ir<%x>
 ; CHECK-NEXT:      EMIT ir<%iv.next> = add ir<%iv>, ir<1>
diff --git a/llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll b/llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll
index 6eff6d1df3523..a27d0d9ecd227 100644
--- a/llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll
+++ b/llvm/test/Transforms/LoopVectorize/VPlan/vplan-printing-outer-loop.ll
@@ -15,14 +15,14 @@ define void @foo(i64 %n) {
 ; CHECK-EMPTY:
 ; CHECK-NEXT:  outer.header:
 ; CHECK-NEXT:    EMIT-SCALAR ir<%outer.iv> = phi [ ir<%outer.iv.next>, outer.latch ], [ ir<0>, ir-bb<entry> ]
-; CHECK-NEXT:    EMIT ir<%gep.1> = getelementptr inbounds ir<@arr2>, ir<0>, ir<%outer.iv>
+; CHECK-NEXT:    EMIT-SCALAR ir<%gep.1> = getelementptr inbounds ptr ir<@arr2>, ir<0>, ir<%outer.iv>
 ; CHECK-NEXT:    EMIT store ir<%outer.iv>, ir<%gep.1>
 ; CHECK-NEXT:    EMIT ir<%add> = add nsw ir<%outer.iv>, ir<%n>
 ; CHECK-NEXT:  Successor(s): inner
 ; CHECK-EMPTY:
 ; CHECK-NEXT:  inner:
 ; CHECK-NEXT:    EMIT-SCALAR ir<%inner.iv> = phi [ ir<%inner.iv.next>, inner ], [ ir<0>, outer.header ]
-; CHECK-NEXT:    EMIT ir<%gep.2> = getelementptr inbounds ir<@arr>, ir<0>, ir<%inner.iv>, ir<%outer.iv>
+; CHECK-NEXT:    EMIT-SCALAR ir<%gep.2> = getelementptr inbounds ptr ir<@arr>, ir<0>, ir<%inner.iv>, ir<%outer.iv>
 ; CHECK-NEXT:    EMIT store ir<%add>, ir<%gep.2>
 ; CHECK-NEXT:    EMIT ir<%inner.iv.next> = add nuw nsw ir<%inner.iv>, ir<1>
 ; CHECK-NEXT:    EMIT ir<%inner.ec> = icmp eq ir<%inner.iv.next>, ir<8>

github-actions · 2026-04-22T14:50:13Z

🐧 Linux x64 Test Results

196306 tests passed
5339 tests skipped

✅ The build succeeded and all tests passed.

github-actions · 2026-04-22T14:50:13Z

🪟 Windows x64 Test Results

135450 tests passed
3395 tests skipped

✅ The build succeeded and all tests passed.

artagnon · 2026-04-30T08:53:48Z

Gentle ping.

artagnon · 2026-05-14T00:23:59Z

Gentle ping.

artagnon · 2026-05-28T14:16:59Z

Gentle ping.

Currently, WidenGEP, Replicate, and VPInstruction recipes can all hold a Instruction::GetElementPtr. Introduce a first-class "scalar GEP" VPInstructionWithType, similar to the existing "scalar cast", with the additional benefit of being able to query the source element type of the GEP directly without going to the underlying value. Planned follow-ups include unifying WidenGEP and Replicate GEPs with this VPInstruction GEP.

artagnon requested review from fhahn and lukel97 April 22, 2026 14:22

llvmbot added vectorizers llvm:transforms labels Apr 22, 2026

artagnon mentioned this pull request Apr 22, 2026

[VPlan] Avoid assuming underlying values for GEPs (NFC) #193460

Closed

artagnon force-pushed the vplan-scalargep branch 3 times, most recently from fcbb92c to b79ff4a Compare April 28, 2026 07:09

artagnon force-pushed the vplan-scalargep branch from b79ff4a to a072b3d Compare April 30, 2026 09:15

This was referenced May 3, 2026

[VPlan] Embed widening semantics of recipes (NFCI) #195385

Closed

[VPlan] Simplify WidenGEP::execute #193543

Merged

artagnon force-pushed the vplan-scalargep branch 5 times, most recently from 205ebeb to bbb6169 Compare May 4, 2026 14:29

artagnon mentioned this pull request May 4, 2026

[LV] Skip replicating already-scalar casts #195646

Closed

fhahn reviewed May 4, 2026

View reviewed changes

Comment thread llvm/lib/Transforms/Vectorize/VPlan.h Outdated

artagnon changed the title ~~[VPlan] Introduce scalar GEP VPInstruction~~ [VPlan] Introduce VPGEPInstruction May 5, 2026

artagnon force-pushed the vplan-scalargep branch 3 times, most recently from f3d1b67 to a585365 Compare May 6, 2026 08:53

artagnon force-pushed the vplan-scalargep branch from a585365 to 130d7e1 Compare May 14, 2026 00:23

artagnon force-pushed the vplan-scalargep branch from 130d7e1 to e297bb4 Compare May 25, 2026 12:31

artagnon force-pushed the vplan-scalargep branch 2 times, most recently from 48f5b9f to 9e0e2a2 Compare May 28, 2026 14:03

artagnon force-pushed the vplan-scalargep branch from 9e0e2a2 to 4039e34 Compare June 9, 2026 08:20

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[VPlan] Introduce VPGEPInstruction#193510

[VPlan] Introduce VPGEPInstruction#193510
artagnon wants to merge 1 commit into
llvm:mainfrom
artagnon:vplan-scalargep

artagnon commented Apr 22, 2026 •

edited

Loading

Uh oh!

llvmbot commented Apr 22, 2026

Uh oh!

llvmbot commented Apr 22, 2026

Uh oh!

github-actions Bot commented Apr 22, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Apr 22, 2026 •

edited

Loading

Uh oh!

artagnon commented Apr 30, 2026

Uh oh!

Uh oh!

artagnon commented May 14, 2026

Uh oh!

artagnon commented May 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

artagnon commented Apr 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Apr 22, 2026

Uh oh!

llvmbot commented Apr 22, 2026

Uh oh!

github-actions Bot commented Apr 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🐧 Linux x64 Test Results

Uh oh!

github-actions Bot commented Apr 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🪟 Windows x64 Test Results

Uh oh!

artagnon commented Apr 30, 2026

Uh oh!

Uh oh!

artagnon commented May 14, 2026

Uh oh!

artagnon commented May 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

artagnon commented Apr 22, 2026 •

edited

Loading

github-actions Bot commented Apr 22, 2026 •

edited

Loading

github-actions Bot commented Apr 22, 2026 •

edited

Loading