Merge branch 'main' into modify-tutorial-typo

Author: baramatsubonzo

llvm/include/llvm/ExecutionEngine/Orc/WaitingOnGraph.h

@@ -500,7 +500,7 @@ template <typename ContainerIdT, typename ElementIdT> class WaitingOnGraph {
         if (I == SN->Deps.end())
           continue;
         for (auto &[DefElem, DefSN] : DefElems)
-          if (I->second.erase(DefElem))
+          if (I->second.erase(DefElem) && DefSN != SN.get())
             SNDeps.insert(DefSN);
         if (I->second.empty())
           SN->Deps.erase(I);
@@ -511,11 +511,13 @@ template <typename ContainerIdT, typename ElementIdT> class WaitingOnGraph {
   // Compute transitive closure of deps for each node.
   static void propagateSuperNodeDeps(SuperNodeDepsMap &SuperNodeDeps) {
     for (auto &[SN, Deps] : SuperNodeDeps) {
-      DenseSet<SuperNode *> Reachable({SN});
+      DenseSet<SuperNode *> Reachable;
       SmallVector<SuperNode *> Worklist(Deps.begin(), Deps.end());
 
       while (!Worklist.empty()) {
         auto *DepSN = Worklist.pop_back_val();
+        if (DepSN == SN)
+          continue;
         if (!Reachable.insert(DepSN).second)
           continue;
         auto I = SuperNodeDeps.find(DepSN);
@@ -537,9 +539,11 @@ template <typename ContainerIdT, typename ElementIdT> class WaitingOnGraph {
       if (I == SuperNodeDeps.end())
         continue;
 
-      for (auto *DepSN : I->second)
+      for (auto *DepSN : I->second) {
+        assert(DepSN != SN.get() && "Unexpected self-dependence for SN");
         for (auto &[Container, Elems] : DepSN->Deps)
           SN->Deps[Container].insert(Elems.begin(), Elems.end());
+      }
     }
   }
 

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp

@@ -140,18 +140,23 @@ Instruction *InstCombinerImpl::scalarizePHI(ExtractElementInst &EI,
     Value *Elt = EI.getIndexOperand();
     // If the operand is the PHI induction variable:
     if (PHIInVal == PHIUser) {
-      // Scalarize the binary operation. Its first operand is the
-      // scalar PHI, and the second operand is extracted from the other
+      // Scalarize the binary operation. One operand is the
+      // scalar PHI, and the other is extracted from the other
       // vector operand.
       BinaryOperator *B0 = cast<BinaryOperator>(PHIUser);
       unsigned opId = (B0->getOperand(0) == PN) ? 1 : 0;
       Value *Op = InsertNewInstWith(
           ExtractElementInst::Create(B0->getOperand(opId), Elt,
                                      B0->getOperand(opId)->getName() + ".Elt"),
           B0->getIterator());
-      Value *newPHIUser = InsertNewInstWith(
-          BinaryOperator::CreateWithCopiedFlags(B0->getOpcode(),
-                                                scalarPHI, Op, B0), B0->getIterator());
+      // Preserve operand order for binary operation to preserve semantics of
+      // non-commutative operations.
+      Value *FirstOp = (B0->getOperand(0) == PN) ? scalarPHI : Op;
+      Value *SecondOp = (B0->getOperand(0) == PN) ? Op : scalarPHI;
+      Value *newPHIUser =
+          InsertNewInstWith(BinaryOperator::CreateWithCopiedFlags(
+                                B0->getOpcode(), FirstOp, SecondOp, B0),
+                            B0->getIterator());
       scalarPHI->addIncoming(newPHIUser, inBB);
     } else {
       // Scalarize PHI input:

llvm/test/Transforms/InstCombine/scalarization.ll

@@ -108,6 +108,50 @@ for.end:
   ret void
 }
 
+define void @scalarize_phi_sub(ptr %n, ptr %inout) {
+;
+; CHECK-LABEL: @scalarize_phi_sub(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[T0:%.*]] = load volatile float, ptr [[INOUT:%.*]], align 4
+; CHECK-NEXT:    br label [[FOR_COND:%.*]]
+; CHECK:       for.cond:
+; CHECK-NEXT:    [[TMP0:%.*]] = phi float [ [[T0]], [[ENTRY:%.*]] ], [ [[TMP1:%.*]], [[FOR_BODY:%.*]] ]
+; CHECK-NEXT:    [[I_0:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[T1:%.*]] = load i32, ptr [[N:%.*]], align 4
+; CHECK-NEXT:    [[CMP_NOT:%.*]] = icmp eq i32 [[I_0]], [[T1]]
+; CHECK-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; CHECK:       for.body:
+; CHECK-NEXT:    store volatile float [[TMP0]], ptr [[INOUT]], align 4
+; CHECK-NEXT:    [[TMP1]] = fsub float 0.000000e+00, [[TMP0]]
+; CHECK-NEXT:    [[INC]] = add nuw nsw i32 [[I_0]], 1
+; CHECK-NEXT:    br label [[FOR_COND]]
+; CHECK:       for.end:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %t0 = load volatile float, ptr %inout, align 4
+  %insert = insertelement <4 x float> poison, float %t0, i32 0
+  %splat = shufflevector <4 x float> %insert, <4 x float> poison, <4 x i32> zeroinitializer
+  br label %for.cond
+
+for.cond:
+  %x.0 = phi <4 x float> [ %splat, %entry ], [ %sub, %for.body ]
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %t1 = load i32, ptr %n, align 4
+  %cmp = icmp ne i32 %i.0, %t1
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:
+  %t2 = extractelement <4 x float> %x.0, i32 1
+  store volatile float %t2, ptr %inout, align 4
+  %sub = fsub <4 x float> zeroinitializer, %x.0
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:
+  ret void
+}
+
 define float @extract_element_binop_splat_constant_index(<4 x float> %x) {
 ;
 ; CHECK-LABEL: @extract_element_binop_splat_constant_index(

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

@@ -2610,8 +2610,13 @@ cc_library(
     hdrs = glob(["include/mlir/Dialect/X86Vector/TransformOps/*.h"]),
     includes = ["include"],
     deps = [
+        ":IR",
+        ":LLVMCommonConversion",
+        ":LLVMDialect",
         ":TransformDialect",
+        ":TransformDialectInterfaces",
         ":VectorDialect",
+        ":X86VectorDialect",
         ":X86VectorTransformOpsIncGen",
         ":X86VectorTransforms",
     ],
@@ -2628,6 +2633,9 @@ cc_library(
         ":LLVMCommonConversion",
         ":LLVMDialect",
         ":LinalgDialect",
+        ":LinalgInterfaces",
+        ":Pass",
+        ":TransformUtils",
         ":VectorDialect",
         ":VectorUtils",
         ":X86VectorDialect",