[InstCombine] Perform sext transformation for values with >1 users

Author: SavchenkoValeriy

llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp

@@ -416,6 +416,7 @@ class TypeEvaluationHelper {
                                           InstCombinerImpl &IC,
                                           Instruction *CxtI);
   [[nodiscard]] bool canEvaluateSExtdImpl(Value *V, Type *Ty);
+  [[nodiscard]] bool canEvaluateSExtdPred(Value *V, Type *Ty);
 
   /// A bookkeeping map to memorize an already made decision for a traversed
   /// value.
@@ -1332,6 +1333,8 @@ bool TypeEvaluationHelper::canEvaluateZExtdImpl(Value *V, Type *Ty,
   BitsToClear = 0;
   if (canAlwaysEvaluateInType(V, Ty))
     return true;
+  // We stick to the one-user limit for the ZExt transform due to the fact
+  // that this predicate returns two values: predicate result and BitsToClear.
   if (canNotEvaluateInType(V, Ty))
     return false;
 
@@ -1681,16 +1684,18 @@ Instruction *InstCombinerImpl::transformSExtICmp(ICmpInst *Cmp,
 ///
 bool TypeEvaluationHelper::canEvaluateSExtd(Value *V, Type *Ty) {
   TypeEvaluationHelper TYH;
-  return TYH.canEvaluateSExtdImpl(V, Ty);
+  return TYH.canEvaluateSExtdImpl(V, Ty) && TYH.allPendingVisited();
 }
 
 bool TypeEvaluationHelper::canEvaluateSExtdImpl(Value *V, Type *Ty) {
+  return canEvaluate(V, Ty, [this](Value *V, Type *Ty) {
+    return canEvaluateSExtdPred(V, Ty);
+  });
+}
+
+bool TypeEvaluationHelper::canEvaluateSExtdPred(Value *V, Type *Ty) {
   assert(V->getType()->getScalarSizeInBits() < Ty->getScalarSizeInBits() &&
          "Can't sign extend type to a smaller type");
-  if (canAlwaysEvaluateInType(V, Ty))
-    return true;
-  if (canNotEvaluateInType(V, Ty))
-    return false;
 
   auto *I = cast<Instruction>(V);
   switch (I->getOpcode()) {
@@ -1708,20 +1713,21 @@ bool TypeEvaluationHelper::canEvaluateSExtdImpl(Value *V, Type *Ty) {
     return canEvaluateSExtdImpl(I->getOperand(0), Ty) &&
            canEvaluateSExtdImpl(I->getOperand(1), Ty);
 
-  //case Instruction::Shl:   TODO
-  //case Instruction::LShr:  TODO
+    // case Instruction::Shl:   TODO
+    // case Instruction::LShr:  TODO
 
   case Instruction::Select:
     return canEvaluateSExtdImpl(I->getOperand(1), Ty) &&
            canEvaluateSExtdImpl(I->getOperand(2), Ty);
 
   case Instruction::PHI: {
     // We can change a phi if we can change all operands.  Note that we never
-    // get into trouble with cyclic PHIs here because we only consider
-    // instructions with a single use.
+    // get into trouble with cyclic PHIs here because canEvaluate handles use
+    // chain loops.
     PHINode *PN = cast<PHINode>(I);
     for (Value *IncValue : PN->incoming_values())
-      if (!canEvaluateSExtdImpl(IncValue, Ty)) return false;
+      if (!canEvaluateSExtdImpl(IncValue, Ty))
+        return false;
     return true;
   }
   default:

llvm/test/Transforms/InstCombine/cast-mul-select.ll

@@ -139,25 +139,21 @@ define i32 @eval_zext_multi_use_in_one_inst(i32 %x) {
 
 define i32 @eval_sext_multi_use_in_one_inst(i32 %x) {
 ; CHECK-LABEL: @eval_sext_multi_use_in_one_inst(
-; CHECK-NEXT:    [[T:%.*]] = trunc i32 [[X:%.*]] to i16
-; CHECK-NEXT:    [[A:%.*]] = and i16 [[T]], 14
-; CHECK-NEXT:    [[M:%.*]] = mul nuw nsw i16 [[A]], [[A]]
-; CHECK-NEXT:    [[O:%.*]] = or disjoint i16 [[M]], -32768
-; CHECK-NEXT:    [[R:%.*]] = sext i16 [[O]] to i32
+; CHECK-NEXT:    [[A:%.*]] = and i32 [[X:%.*]], 14
+; CHECK-NEXT:    [[M:%.*]] = mul nuw nsw i32 [[A]], [[A]]
+; CHECK-NEXT:    [[R:%.*]] = or disjoint i32 [[M]], -32768
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
 ; DBGINFO-LABEL: @eval_sext_multi_use_in_one_inst(
-; DBGINFO-NEXT:    [[T:%.*]] = trunc i32 [[X:%.*]] to i16, !dbg [[DBG81:![0-9]+]]
-; DBGINFO-NEXT:      #dbg_value(i16 [[T]], [[META76:![0-9]+]], !DIExpression(), [[DBG81]])
-; DBGINFO-NEXT:    [[A:%.*]] = and i16 [[T]], 14, !dbg [[DBG82:![0-9]+]]
-; DBGINFO-NEXT:      #dbg_value(i16 [[A]], [[META77:![0-9]+]], !DIExpression(), [[DBG82]])
-; DBGINFO-NEXT:    [[M:%.*]] = mul nuw nsw i16 [[A]], [[A]], !dbg [[DBG83:![0-9]+]]
-; DBGINFO-NEXT:      #dbg_value(i16 [[M]], [[META78:![0-9]+]], !DIExpression(), [[DBG83]])
-; DBGINFO-NEXT:    [[O:%.*]] = or disjoint i16 [[M]], -32768, !dbg [[DBG84:![0-9]+]]
-; DBGINFO-NEXT:      #dbg_value(i16 [[O]], [[META79:![0-9]+]], !DIExpression(), [[DBG84]])
-; DBGINFO-NEXT:    [[R:%.*]] = sext i16 [[O]] to i32, !dbg [[DBG85:![0-9]+]]
-; DBGINFO-NEXT:      #dbg_value(i32 [[R]], [[META80:![0-9]+]], !DIExpression(), [[DBG85]])
-; DBGINFO-NEXT:    ret i32 [[R]], !dbg [[DBG86:![0-9]+]]
+; DBGINFO-NEXT:      #dbg_value(i32 [[X:%.*]], [[META76:![0-9]+]], !DIExpression(DW_OP_LLVM_convert, 32, DW_ATE_unsigned, DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_stack_value), [[META81:![0-9]+]])
+; DBGINFO-NEXT:    [[A:%.*]] = and i32 [[X]], 14, !dbg [[DBG82:![0-9]+]]
+; DBGINFO-NEXT:      #dbg_value(i32 [[X]], [[META77:![0-9]+]], !DIExpression(DW_OP_LLVM_convert, 32, DW_ATE_unsigned, DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_constu, 14, DW_OP_and, DW_OP_stack_value), [[DBG82]])
+; DBGINFO-NEXT:    [[M:%.*]] = mul nuw nsw i32 [[A]], [[A]], !dbg [[DBG83:![0-9]+]]
+; DBGINFO-NEXT:      #dbg_value(!DIArgList(i32 [[X]], i32 [[X]]), [[META78:![0-9]+]], !DIExpression(DW_OP_LLVM_arg, 0, DW_OP_LLVM_convert, 32, DW_ATE_unsigned, DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_constu, 14, DW_OP_and, DW_OP_LLVM_arg, 1, DW_OP_LLVM_convert, 32, DW_ATE_unsigned, DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_constu, 14, DW_OP_and, DW_OP_mul, DW_OP_stack_value), [[DBG83]])
+; DBGINFO-NEXT:    [[O:%.*]] = or disjoint i32 [[M]], -32768, !dbg [[DBG84:![0-9]+]]
+; DBGINFO-NEXT:      #dbg_value(!DIArgList(i32 [[X]], i32 [[X]]), [[META79:![0-9]+]], !DIExpression(DW_OP_LLVM_arg, 0, DW_OP_LLVM_convert, 32, DW_ATE_unsigned, DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_constu, 14, DW_OP_and, DW_OP_LLVM_arg, 1, DW_OP_LLVM_convert, 32, DW_ATE_unsigned, DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_constu, 14, DW_OP_and, DW_OP_mul, DW_OP_constu, 18446744073709518848, DW_OP_or, DW_OP_stack_value), [[DBG84]])
+; DBGINFO-NEXT:      #dbg_value(i32 [[O]], [[META80:![0-9]+]], !DIExpression(), [[META85:![0-9]+]])
+; DBGINFO-NEXT:    ret i32 [[O]], !dbg [[DBG86:![0-9]+]]
 ;
   %t = trunc i32 %x to i16
   %a = and i16 %t, 14

llvm/test/Transforms/InstCombine/cast.ll

@@ -2312,13 +2312,12 @@ exit:
 
 define i32 @test99(i32 %x, i32 %y) {
 ; ALL-LABEL: @test99(
-; ALL-NEXT:    [[TX:%.*]] = trunc i32 [[X:%.*]] to i16
-; ALL-NEXT:    [[TY:%.*]] = trunc i32 [[Y:%.*]] to i16
-; ALL-NEXT:    [[A:%.*]] = add i16 [[TX]], [[TY]]
-; ALL-NEXT:    [[B:%.*]] = add i16 [[A]], 5
-; ALL-NEXT:    [[C:%.*]] = mul i16 [[A]], 3
-; ALL-NEXT:    [[D:%.*]] = or i16 [[B]], [[C]]
-; ALL-NEXT:    [[S:%.*]] = sext i16 [[D]] to i32
+; ALL-NEXT:    [[A:%.*]] = add i32 [[X:%.*]], [[Y:%.*]]
+; ALL-NEXT:    [[B:%.*]] = add i32 [[A]], 5
+; ALL-NEXT:    [[C:%.*]] = mul i32 [[A]], 3
+; ALL-NEXT:    [[D:%.*]] = or i32 [[B]], [[C]]
+; ALL-NEXT:    [[SEXT:%.*]] = shl i32 [[D]], 16
+; ALL-NEXT:    [[S:%.*]] = ashr exact i32 [[SEXT]], 16
 ; ALL-NEXT:    ret i32 [[S]]
 ;
   %tx = trunc i32 %x to i16