[InstCombine] Improve range computation in foldICmpAddConstant (#155096)

Address comment
#110511 (comment).

Author: dtcxzyw

llvm/include/llvm/IR/ConstantRange.h

@@ -381,6 +381,7 @@ class [[nodiscard]] ConstantRange {
   /// strictly smaller than the current type.  The returned range will
   /// correspond to the possible range of values if the source range had been
   /// truncated to the specified type with wrap type \p NoWrapKind.
+  /// Note that the result of trunc nuw is exact.
   LLVM_ABI ConstantRange truncate(uint32_t BitWidth,
                                   unsigned NoWrapKind = 0) const;
 

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

@@ -3320,20 +3320,18 @@ Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
     Type *NewCmpTy = V->getType();
     unsigned NewCmpBW = NewCmpTy->getScalarSizeInBits();
     if (shouldChangeType(Ty, NewCmpTy)) {
-      if (CR.getActiveBits() <= NewCmpBW) {
-        ConstantRange SrcCR = CR.truncate(NewCmpBW);
-        CmpInst::Predicate EquivPred;
-        APInt EquivInt;
-        APInt EquivOffset;
-
-        SrcCR.getEquivalentICmp(EquivPred, EquivInt, EquivOffset);
-        return new ICmpInst(
-            EquivPred,
-            EquivOffset.isZero()
-                ? V
-                : Builder.CreateAdd(V, ConstantInt::get(NewCmpTy, EquivOffset)),
-            ConstantInt::get(NewCmpTy, EquivInt));
-      }
+      ConstantRange SrcCR = CR.truncate(NewCmpBW, TruncInst::NoUnsignedWrap);
+      CmpInst::Predicate EquivPred;
+      APInt EquivInt;
+      APInt EquivOffset;
+
+      SrcCR.getEquivalentICmp(EquivPred, EquivInt, EquivOffset);
+      return new ICmpInst(
+          EquivPred,
+          EquivOffset.isZero()
+              ? V
+              : Builder.CreateAdd(V, ConstantInt::get(NewCmpTy, EquivOffset)),
+          ConstantInt::get(NewCmpTy, EquivInt));
     }
   }
 

llvm/test/Transforms/InstCombine/icmp-add.ll

@@ -3209,13 +3209,11 @@ entry:
   ret i1 %cmp
 }
 
-; TODO: should be canonicalized to (x - 4) u> 2
 define i1 @zext_range_check_ugt(i8 %x) {
 ; CHECK-LABEL: @zext_range_check_ugt(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[CONV:%.*]] = zext i8 [[X:%.*]] to i32
-; CHECK-NEXT:    [[TMP0:%.*]] = add nsw i32 [[CONV]], -7
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[TMP0]], -3
+; CHECK-NEXT:    [[TMP0:%.*]] = add i8 [[X:%.*]], -7
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[TMP0]], -3
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
 entry:
@@ -3225,13 +3223,11 @@ entry:
   ret i1 %cmp
 }
 
-; TODO: should be canonicalized to (x - 4) u> 2
 define i1 @zext_range_check_ult_alter(i8 %x) {
 ; CHECK-LABEL: @zext_range_check_ult_alter(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[CONV:%.*]] = zext i8 [[X:%.*]] to i32
-; CHECK-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], -7
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[ADD]], -3
+; CHECK-NEXT:    [[TMP0:%.*]] = add i8 [[X:%.*]], -7
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i8 [[TMP0]], -3
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
 entry:
@@ -3241,6 +3237,19 @@ entry:
   ret i1 %cmp
 }
 
+define i1 @zext_range_check_ult_alter2(i8 %x) {
+; CHECK-LABEL: @zext_range_check_ult_alter2(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i8 [[X:%.*]], 3
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+entry:
+  %conv = zext i8 %x to i32
+  %add = add i32 %conv, -4
+  %cmp = icmp ult i32 %add, 253
+  ret i1 %cmp
+}
+
 define i1 @zext_range_check_mergable(i8 %x) {
 ; CHECK-LABEL: @zext_range_check_mergable(
 ; CHECK-NEXT:    [[COND:%.*]] = icmp slt i8 [[X:%.*]], 7
@@ -3286,21 +3295,6 @@ entry:
   ret i1 %cmp
 }
 
-define i1 @zext_range_check_ult_range_check_failure(i8 %x) {
-; CHECK-LABEL: @zext_range_check_ult_range_check_failure(
-; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[CONV:%.*]] = zext i8 [[X:%.*]] to i32
-; CHECK-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], -4
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[ADD]], 253
-; CHECK-NEXT:    ret i1 [[CMP]]
-;
-entry:
-  %conv = zext i8 %x to i32
-  %add = add i32 %conv, -4
-  %cmp = icmp ult i32 %add, 253
-  ret i1 %cmp
-}
-
 ; PR 152851
 
 define i1 @val_is_aligend_const_pow2(i32 %num) {