[InstCombinePHI] Enhance PHI CSE to remove redundant phis

llvm/lib/Analysis/InstructionSimplify.cpp

@@ -4866,6 +4866,89 @@ static Value *simplifySelectWithFCmp(Value *Cond, Value *T, Value *F,
   return nullptr;
 }
 
+// Look for the following pattern and simplify %1 to %identicalPhi.
+// Here %phi, %1 and %phi.next perform the same functionality as
+// %identicalPhi and hence the select instruction %1 can be folded
+// into %identicalPhi.
+//
+// BB1:
+//   %identicalPhi = phi [ X, %BB0 ], [ %identicalPhi.next, %BB1 ]
+//   %phi = phi [ X, %BB0 ], [ %phi.next, %BB1 ]
+//   ...
+//   %identicalPhi.next = select %cmp, %val, %identicalPhi
+//                      (or select %cmp, %identicalPhi, %val)
+//   %1 = select %cmp2, %identicalPhi, %phi
+//   %phi.next = select %cmp, %val, %1
+//             (or select %cmp, %1, %val)
+//
+// Prove that %phi and %identicalPhi are the same by induction:
+//
+// Base case: Both %phi and %identicalPhi are equal on entry to the loop.
+// Inductive case:
+// Suppose %phi and %identicalPhi are equal at iteration i.
+// We look at their values at iteration i+1 which are %phi.next and
+// %identicalPhi.next. They would have become different only when %cmp is
+// false and the corresponding values %1 and %identicalPhi differ
+// (similar reason for the other "or" case in the bracket).
+//
+// The only condition when %1 and %identicalPh could differ is when %cmp2
+// is false and %1 is %phi, which contradicts our inductive hypothesis
+// that %phi and %identicalPhi are equal. Thus %phi and %identicalPhi are
+// always equal at iteration i+1.
+bool isSimplifierIdenticalPHI(PHINode &PN, PHINode &IdenticalPN) {
+  if (PN.getParent() != IdenticalPN.getParent())
+    return false;
+  if (PN.getNumIncomingValues() != 2)
+    return false;
+
+  // Check that only the backedge incoming value is different.
+  unsigned DiffVals = 0;
+  BasicBlock *DiffValBB = nullptr;
+  for (unsigned i = 0; i < 2; i++) {
+    BasicBlock *PredBB = PN.getIncomingBlock(i);
+    if (PN.getIncomingValueForBlock(PredBB) !=
+        IdenticalPN.getIncomingValueForBlock(PredBB)) {
+      DiffVals++;
+      DiffValBB = PredBB;
+    }
+  }
+  if (DiffVals != 1)
+    return false;
+  // Now check that the backedge incoming values are two select
+  // instructions with the same condition. Either their true
+  // values are the same, or their false values are the same.
+  auto *SI = dyn_cast<SelectInst>(PN.getIncomingValueForBlock(DiffValBB));
+  auto *IdenticalSI =
+      dyn_cast<SelectInst>(IdenticalPN.getIncomingValueForBlock(DiffValBB));
+  if (!SI || !IdenticalSI)
+    return false;
+  if (SI->getCondition() != IdenticalSI->getCondition() ||
+      (SI->getTrueValue() != IdenticalSI->getTrueValue() &&
+       SI->getFalseValue() != IdenticalSI->getFalseValue()))
+    return false;
+
+  SelectInst *SIOtherVal = nullptr;
+  Value *IdenticalSIOtherVal = nullptr;
+  if (SI->getTrueValue() == IdenticalSI->getTrueValue()) {
+    SIOtherVal = dyn_cast<SelectInst>(SI->getFalseValue());
+    IdenticalSIOtherVal = IdenticalSI->getFalseValue();
+  } else {
+    SIOtherVal = dyn_cast<SelectInst>(SI->getTrueValue());
+    IdenticalSIOtherVal = IdenticalSI->getTrueValue();
+  }
+
+  // Now check that the other values in select, i.e., %1 and %identicalPhi,
+  // are essentially the same value within the same BB.
+  if (!SIOtherVal || IdenticalSIOtherVal != &IdenticalPN)
+    return false;
+  if (!(SIOtherVal->getTrueValue() == &IdenticalPN &&
+        SIOtherVal->getFalseValue() == &PN) &&
+      !(SIOtherVal->getTrueValue() == &PN &&
+        SIOtherVal->getFalseValue() == &IdenticalPN))
+    return false;
+  return true;
+}
+
 /// Given operands for a SelectInst, see if we can fold the result.
 /// If not, this returns null.
 static Value *simplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
@@ -5041,7 +5124,14 @@ static Value *simplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
   std::optional<bool> Imp = isImpliedByDomCondition(Cond, Q.CxtI, Q.DL);
   if (Imp)
     return *Imp ? TrueVal : FalseVal;
-
+  // Look for same PHIs in the true and false values.
+  if (auto *TruePHI = dyn_cast<PHINode>(TrueVal))
+    if (auto *FalsePHI = dyn_cast<PHINode>(FalseVal)) {
+      if (isSimplifierIdenticalPHI(*TruePHI, *FalsePHI))
+        return FalseVal;
+      if (isSimplifierIdenticalPHI(*FalsePHI, *TruePHI))
+        return TrueVal;
+    }
   return nullptr;
 }
 

llvm/test/Transforms/InstCombine/select_with_identical_phi.ll

@@ -0,0 +1,243 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -S -passes=instcombine | FileCheck %s
+@A = extern_weak global float, align 4
+
+; %same.as.v1 is a select with two phis %v1 and %phi.to.remove as the true
+; and false values, while %v1 and %phi.to.remove are actually the same.
+; Fold the selection instruction %same.as.v1 to %v1.
+define void @select_with_identical_phi(ptr %m, ptr %n, i32 %count) {
+; CHECK-LABEL: @select_with_identical_phi(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
+; CHECK:       for.body:
+; CHECK-NEXT:    [[V0:%.*]] = phi float [ 0x4415AF1D80000000, [[ENTRY:%.*]] ], [ [[V0_1:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[V1:%.*]] = phi float [ 0xC415AF1D80000000, [[ENTRY]] ], [ [[V1_1:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[INC_I:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[Q:%.*]] = phi ptr [ [[M:%.*]], [[ENTRY]] ], [ [[Q_NEXT:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[C:%.*]] = phi ptr [ [[N:%.*]], [[ENTRY]] ], [ [[C_NEXT:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[Q_LOAD:%.*]] = load float, ptr [[Q]], align 4
+; CHECK-NEXT:    [[C_LOAD:%.*]] = load float, ptr [[C]], align 4
+; CHECK-NEXT:    [[SUB:%.*]] = fsub float [[Q_LOAD]], [[C_LOAD]]
+; CHECK-NEXT:    [[CMP1:%.*]] = fcmp olt float [[SUB]], [[V0]]
+; CHECK-NEXT:    [[V0_1]] = select i1 [[CMP1]], float [[SUB]], float [[V0]]
+; CHECK-NEXT:    [[CMP2:%.*]] = fcmp ogt float [[SUB]], [[V1]]
+; CHECK-NEXT:    [[V1_1]] = select i1 [[CMP2]], float [[SUB]], float [[V1]]
+; CHECK-NEXT:    [[INC_I]] = add nuw nsw i32 [[I]], 1
+; CHECK-NEXT:    [[Q_NEXT]] = getelementptr inbounds nuw i8, ptr [[Q]], i64 4
+; CHECK-NEXT:    [[C_NEXT]] = getelementptr inbounds nuw i8, ptr [[C]], i64 4
+; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC_I]], [[COUNT:%.*]]
+; CHECK-NEXT:    br i1 [[EXITCOND]], label [[EXIT:%.*]], label [[FOR_BODY]]
+; CHECK:       exit:
+; CHECK-NEXT:    store float [[V1_1]], ptr @A, align 4
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %for.body
+
+for.body:                                    ; preds = %entry, %for.body
+  %v0 = phi float [ 0x4415AF1D80000000, %entry ], [ %v0.1, %for.body ]
+  %v1 = phi float [ 0xC415AF1D80000000, %entry ], [ %v1.1, %for.body ]
+  %phi.to.remove = phi float [ 0xC415AF1D80000000, %entry ], [ %phi.to.remove.next, %for.body ]
+  %i = phi i32 [ 0, %entry ], [ %inc.i, %for.body ]
+  %q = phi ptr [ %m, %entry ], [ %q.next, %for.body ]
+  %c = phi ptr [ %n, %entry ], [ %c.next, %for.body ]
+  %q.load = load float, ptr %q
+  %c.load = load float, ptr %c
+  %sub = fsub float %q.load, %c.load
+  %cmp1 = fcmp olt float %sub, %v0
+  %v0.1 = select i1 %cmp1, float %sub, float %v0
+  %same.as.v1 = select i1 %cmp1, float %v1, float %phi.to.remove
+  %cmp2 = fcmp ogt float  %sub, %same.as.v1
+  %v1.1 = select i1 %cmp2, float %sub, float %v1
+  %phi.to.remove.next = select i1 %cmp2, float %sub, float %same.as.v1
+  %inc.i = add nuw nsw i32 %i, 1
+  %q.next = getelementptr inbounds i8, ptr %q, i64 4
+  %c.next = getelementptr inbounds i8, ptr %c, i64 4
+  %exitcond = icmp eq i32 %inc.i, %count
+  br i1 %exitcond, label %exit, label %for.body
+
+exit:
+  %vl.1.lcssa = phi float [ %v1.1, %for.body ]
+  store float %vl.1.lcssa, ptr @A
+  ret void
+}
+
+; The difference from select_with_identical_phi() is that the true and false values in
+; %phi.to.remove.next and %v1.1 are swapped.
+; Check that %same.as.v1 can be folded.
+define void @select_with_identical_phi_2(ptr %m, ptr %n, i32 %count) {
+; CHECK-LABEL: @select_with_identical_phi_2(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
+; CHECK:       for.body:
+; CHECK-NEXT:    [[V0:%.*]] = phi float [ 0x4415AF1D80000000, [[ENTRY:%.*]] ], [ [[V0_1:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[V1:%.*]] = phi float [ 0xC415AF1D80000000, [[ENTRY]] ], [ [[V1_1:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[INC_I:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[Q:%.*]] = phi ptr [ [[M:%.*]], [[ENTRY]] ], [ [[Q_NEXT:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[C:%.*]] = phi ptr [ [[N:%.*]], [[ENTRY]] ], [ [[C_NEXT:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[Q_LOAD:%.*]] = load float, ptr [[Q]], align 4
+; CHECK-NEXT:    [[C_LOAD:%.*]] = load float, ptr [[C]], align 4
+; CHECK-NEXT:    [[SUB:%.*]] = fsub float [[Q_LOAD]], [[C_LOAD]]
+; CHECK-NEXT:    [[CMP1:%.*]] = fcmp olt float [[SUB]], [[V0]]
+; CHECK-NEXT:    [[V0_1]] = select i1 [[CMP1]], float [[SUB]], float [[V0]]
+; CHECK-NEXT:    [[CMP2:%.*]] = fcmp ogt float [[SUB]], [[V1]]
+; CHECK-NEXT:    [[V1_1]] = select i1 [[CMP2]], float [[V1]], float [[SUB]]
+; CHECK-NEXT:    [[INC_I]] = add nuw nsw i32 [[I]], 1
+; CHECK-NEXT:    [[Q_NEXT]] = getelementptr inbounds nuw i8, ptr [[Q]], i64 4
+; CHECK-NEXT:    [[C_NEXT]] = getelementptr inbounds nuw i8, ptr [[C]], i64 4
+; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC_I]], [[COUNT:%.*]]
+; CHECK-NEXT:    br i1 [[EXITCOND]], label [[EXIT:%.*]], label [[FOR_BODY]]
+; CHECK:       exit:
+; CHECK-NEXT:    store float [[V1_1]], ptr @A, align 4
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %for.body
+
+for.body:                                    ; preds = %entry, %for.body
+  %v0 = phi float [ 0x4415AF1D80000000, %entry ], [ %v0.1, %for.body ]
+  %v1 = phi float [ 0xC415AF1D80000000, %entry ], [ %v1.1, %for.body ]
+  %phi.to.remove = phi float [ 0xC415AF1D80000000, %entry ], [ %phi.to.remove.next, %for.body ]
+  %i = phi i32 [ 0, %entry ], [ %inc.i, %for.body ]
+  %q = phi ptr [ %m, %entry ], [ %q.next, %for.body ]
+  %c = phi ptr [ %n, %entry ], [ %c.next, %for.body ]
+  %q.load = load float, ptr %q
+  %c.load = load float, ptr %c
+  %sub = fsub float %q.load, %c.load
+  %cmp1 = fcmp olt float %sub, %v0
+  %v0.1 = select i1 %cmp1, float %sub, float %v0
+  %same.as.v1 = select i1 %cmp1, float %v1, float %phi.to.remove
+  %cmp2 = fcmp ogt float  %sub, %same.as.v1
+  %v1.1 = select i1 %cmp2, float %v1, float %sub
+  %phi.to.remove.next = select i1 %cmp2, float %same.as.v1, float %sub
+  %inc.i = add nuw nsw i32 %i, 1
+  %q.next = getelementptr inbounds i8, ptr %q, i64 4
+  %c.next = getelementptr inbounds i8, ptr %c, i64 4
+  %exitcond = icmp eq i32 %inc.i, %count
+  br i1 %exitcond, label %exit, label %for.body
+
+exit:
+  %vl.1.lcssa = phi float [ %v1.1, %for.body ]
+  store float %vl.1.lcssa, ptr @A
+  ret void
+}
+
+; The difference from select_with_identical_phi() is that the true and false values in
+; same.as.v1 are swapped.
+; Check that %same.as.v1 can be folded.
+define void @select_with_identical_phi_3(ptr %m, ptr %n, i32 %count) {
+; CHECK-LABEL: @select_with_identical_phi_3(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
+; CHECK:       for.body:
+; CHECK-NEXT:    [[V0:%.*]] = phi float [ 0x4415AF1D80000000, [[ENTRY:%.*]] ], [ [[V0_1:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[V1:%.*]] = phi float [ 0xC415AF1D80000000, [[ENTRY]] ], [ [[V1_1:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[INC_I:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[Q:%.*]] = phi ptr [ [[M:%.*]], [[ENTRY]] ], [ [[Q_NEXT:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[C:%.*]] = phi ptr [ [[N:%.*]], [[ENTRY]] ], [ [[C_NEXT:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[Q_LOAD:%.*]] = load float, ptr [[Q]], align 4
+; CHECK-NEXT:    [[C_LOAD:%.*]] = load float, ptr [[C]], align 4
+; CHECK-NEXT:    [[SUB:%.*]] = fsub float [[Q_LOAD]], [[C_LOAD]]
+; CHECK-NEXT:    [[CMP1:%.*]] = fcmp olt float [[SUB]], [[V0]]
+; CHECK-NEXT:    [[V0_1]] = select i1 [[CMP1]], float [[SUB]], float [[V0]]
+; CHECK-NEXT:    [[CMP2:%.*]] = fcmp ogt float [[SUB]], [[V1]]
+; CHECK-NEXT:    [[V1_1]] = select i1 [[CMP2]], float [[SUB]], float [[V1]]
+; CHECK-NEXT:    [[INC_I]] = add nuw nsw i32 [[I]], 1
+; CHECK-NEXT:    [[Q_NEXT]] = getelementptr inbounds nuw i8, ptr [[Q]], i64 4
+; CHECK-NEXT:    [[C_NEXT]] = getelementptr inbounds nuw i8, ptr [[C]], i64 4
+; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC_I]], [[COUNT:%.*]]
+; CHECK-NEXT:    br i1 [[EXITCOND]], label [[EXIT:%.*]], label [[FOR_BODY]]
+; CHECK:       exit:
+; CHECK-NEXT:    store float [[V1_1]], ptr @A, align 4
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %for.body
+
+for.body:                                    ; preds = %entry, %for.body
+  %v0 = phi float [ 0x4415AF1D80000000, %entry ], [ %v0.1, %for.body ]
+  %v1 = phi float [ 0xC415AF1D80000000, %entry ], [ %v1.1, %for.body ]
+  %phi.to.remove = phi float [ 0xC415AF1D80000000, %entry ], [ %phi.to.remove.next, %for.body ]
+  %i = phi i32 [ 0, %entry ], [ %inc.i, %for.body ]
+  %q = phi ptr [ %m, %entry ], [ %q.next, %for.body ]
+  %c = phi ptr [ %n, %entry ], [ %c.next, %for.body ]
+  %q.load = load float, ptr %q
+  %c.load = load float, ptr %c
+  %sub = fsub float %q.load, %c.load
+  %cmp1 = fcmp olt float %sub, %v0
+  %v0.1 = select i1 %cmp1, float %sub, float %v0
+  %same.as.v1 = select i1 %cmp1, float %phi.to.remove, float %v1
+  %cmp2 = fcmp ogt float  %sub, %same.as.v1
+  %v1.1 = select i1 %cmp2, float %sub, float %v1
+  %phi.to.remove.next = select i1 %cmp2, float %sub, float %same.as.v1
+  %inc.i = add nuw nsw i32 %i, 1
+  %q.next = getelementptr inbounds i8, ptr %q, i64 4
+  %c.next = getelementptr inbounds i8, ptr %c, i64 4
+  %exitcond = icmp eq i32 %inc.i, %count
+  br i1 %exitcond, label %exit, label %for.body
+
+exit:
+  %vl.1.lcssa = phi float [ %v1.1, %for.body ]
+  store float %vl.1.lcssa, ptr @A
+  ret void
+}
+
+; The difference from select_with_identical_phi() is that the true and false values in
+; %same.as.v1, %phi.to.remove.next and %v1.1 are swapped.
+; Check that %same.as.v1 can be folded.
+define void @select_with_identical_phi_4(ptr %m, ptr %n, i32 %count) {
+; CHECK-LABEL: @select_with_identical_phi_4(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
+; CHECK:       for.body:
+; CHECK-NEXT:    [[V0:%.*]] = phi float [ 0x4415AF1D80000000, [[ENTRY:%.*]] ], [ [[V0_1:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[V1:%.*]] = phi float [ 0xC415AF1D80000000, [[ENTRY]] ], [ [[V1_1:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[INC_I:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[Q:%.*]] = phi ptr [ [[M:%.*]], [[ENTRY]] ], [ [[Q_NEXT:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[C:%.*]] = phi ptr [ [[N:%.*]], [[ENTRY]] ], [ [[C_NEXT:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[Q_LOAD:%.*]] = load float, ptr [[Q]], align 4
+; CHECK-NEXT:    [[C_LOAD:%.*]] = load float, ptr [[C]], align 4
+; CHECK-NEXT:    [[SUB:%.*]] = fsub float [[Q_LOAD]], [[C_LOAD]]
+; CHECK-NEXT:    [[CMP1:%.*]] = fcmp olt float [[SUB]], [[V0]]
+; CHECK-NEXT:    [[V0_1]] = select i1 [[CMP1]], float [[SUB]], float [[V0]]
+; CHECK-NEXT:    [[CMP2:%.*]] = fcmp ogt float [[SUB]], [[V1]]
+; CHECK-NEXT:    [[V1_1]] = select i1 [[CMP2]], float [[V1]], float [[SUB]]
+; CHECK-NEXT:    [[INC_I]] = add nuw nsw i32 [[I]], 1
+; CHECK-NEXT:    [[Q_NEXT]] = getelementptr inbounds nuw i8, ptr [[Q]], i64 4
+; CHECK-NEXT:    [[C_NEXT]] = getelementptr inbounds nuw i8, ptr [[C]], i64 4
+; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC_I]], [[COUNT:%.*]]
+; CHECK-NEXT:    br i1 [[EXITCOND]], label [[EXIT:%.*]], label [[FOR_BODY]]
+; CHECK:       exit:
+; CHECK-NEXT:    store float [[V1_1]], ptr @A, align 4
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %for.body
+
+for.body:                                    ; preds = %entry, %for.body
+  %v0 = phi float [ 0x4415AF1D80000000, %entry ], [ %v0.1, %for.body ]
+  %v1 = phi float [ 0xC415AF1D80000000, %entry ], [ %v1.1, %for.body ]
+  %phi.to.remove = phi float [ 0xC415AF1D80000000, %entry ], [ %phi.to.remove.next, %for.body ]
+  %i = phi i32 [ 0, %entry ], [ %inc.i, %for.body ]
+  %q = phi ptr [ %m, %entry ], [ %q.next, %for.body ]
+  %c = phi ptr [ %n, %entry ], [ %c.next, %for.body ]
+  %q.load = load float, ptr %q
+  %c.load = load float, ptr %c
+  %sub = fsub float %q.load, %c.load
+  %cmp1 = fcmp olt float %sub, %v0
+  %v0.1 = select i1 %cmp1, float %sub, float %v0
+  %same.as.v1 = select i1 %cmp1, float %phi.to.remove, float %v1
+  %cmp2 = fcmp ogt float  %sub, %same.as.v1
+  %v1.1 = select i1 %cmp2, float %v1, float %sub
+  %phi.to.remove.next = select i1 %cmp2, float %same.as.v1, float %sub
+  %inc.i = add nuw nsw i32 %i, 1
+  %q.next = getelementptr inbounds i8, ptr %q, i64 4
+  %c.next = getelementptr inbounds i8, ptr %c, i64 4
+  %exitcond = icmp eq i32 %inc.i, %count
+  br i1 %exitcond, label %exit, label %for.body
+
+exit:
+  %vl.1.lcssa = phi float [ %v1.1, %for.body ]
+  store float %vl.1.lcssa, ptr @A
+  ret void
+}