[CorrelatedValuePropagation] Fold calls to UCMP/SCMP when we know that ranges of operands do not overlap

llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp

@@ -548,6 +548,27 @@ static bool processAbsIntrinsic(IntrinsicInst *II, LazyValueInfo *LVI) {
   return false;
 }
 
+static bool processCmpIntrinsic(IntrinsicInst *II, LazyValueInfo *LVI) {
+  bool IsSigned = II->getIntrinsicID() == Intrinsic::scmp;
+  ConstantRange LHS_CR = LVI->getConstantRangeAtUse(II->getOperandUse(0),
+                                                    /*UndefAllowed*/ false);
+  ConstantRange RHS_CR = LVI->getConstantRangeAtUse(II->getOperandUse(1),
+                                                    /*UndefAllowed*/ false);
+
+  if (LHS_CR.icmp(IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT, RHS_CR)) {
+    II->replaceAllUsesWith(ConstantInt::get(II->getType(), 1));
+    II->eraseFromParent();
+    return true;
+  }
+  if (LHS_CR.icmp(IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, RHS_CR)) {
+    II->replaceAllUsesWith(ConstantInt::getSigned(II->getType(), -1));
+    II->eraseFromParent();
+    return true;
+  }
+
+  return false;
+}
+
 // See if this min/max intrinsic always picks it's one specific operand.
 // If not, check whether we can canonicalize signed minmax into unsigned version
 static bool processMinMaxIntrinsic(MinMaxIntrinsic *MM, LazyValueInfo *LVI) {
@@ -639,6 +660,11 @@ static bool processCallSite(CallBase &CB, LazyValueInfo *LVI) {
     return processAbsIntrinsic(&cast<IntrinsicInst>(CB), LVI);
   }
 
+  if (CB.getIntrinsicID() == Intrinsic::scmp ||
+      CB.getIntrinsicID() == Intrinsic::ucmp) {
+    return processCmpIntrinsic(&cast<IntrinsicInst>(CB), LVI);
+  }
+
   if (auto *MM = dyn_cast<MinMaxIntrinsic>(&CB)) {
     return processMinMaxIntrinsic(MM, LVI);
   }

llvm/test/Transforms/CorrelatedValuePropagation/uscmp.ll

@@ -0,0 +1,118 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=correlated-propagation -S | FileCheck %s
+
+; If nothing is known we can't change anything
+define i8 @ucmp_0(i32 %x, i32 %y) {
+  %1 = call i8 @llvm.ucmp(i32 %x, i32 %y)
+  ret i8 %1
+}
+
+define i8 @scmp_0(i32 %x, i32 %y) {
+  %1 = call i8 @llvm.scmp(i32 %x, i32 %y)
+  ret i8 %1
+}
+
+; If we know that range of LHS < range of RHS then return -1
+define i8 @ucmp_1(i32 %x, i32 %y) {
+  ; X is within [4, 8)
+  %cond1 = icmp uge i32 %x, 4
+  call void @llvm.assume(i1 %cond1)
+  %cond2 = icmp ult i32 %x, 8
+  call void @llvm.assume(i1 %cond2)
+  ; Y is within [8, +INF)
+  %cond3 = icmp uge i32 %y, 8
+  call void @llvm.assume(i1 %cond3)
+
+  %1 = call i8 @llvm.ucmp(i32 %x, i32 %y)
+  ret i8 %1
+}
+
+define i8 @scmp_1(i32 %x, i32 %y) {
+  ; X is within [-5, 3)
+  %cond1 = icmp sge i32 %x, -5
+  call void @llvm.assume(i1 %cond1)
+  %cond2 = icmp slt i32 %x, 3
+  call void @llvm.assume(i1 %cond2)
+  ; Y is within [3, +INF)
+  %cond3 = icmp sge i32 %y, 3
+  call void @llvm.assume(i1 %cond3)
+
+  %1 = call i8 @llvm.scmp(i32 %x, i32 %y)
+  ret i8 %1
+}
+
+; If we know that range of LHS > range of RHS then return 1
+define i8 @ucmp_2(i32 %x, i32 %y) {
+  ; X is within [4, +INF)
+  %cond1 = icmp uge i32 %x, 4
+  call void @llvm.assume(i1 %cond1)
+  ; Y is within [0, 4)
+  %cond2 = icmp ult i32 %y, 4
+  call void @llvm.assume(i1 %cond2)
+
+  %1 = call i8 @llvm.ucmp(i32 %x, i32 %y)
+  ret i8 %1
+}
+
+define i8 @scmp_2(i32 %x, i32 %y) {
+  ; X is within [4, +INF)
+  %cond1 = icmp sge i32 %x, 4
+  call void @llvm.assume(i1 %cond1)
+  ; Y is within [-INF, 4)
+  %cond2 = icmp slt i32 %y, 4
+  call void @llvm.assume(i1 %cond2)
+
+  %1 = call i8 @llvm.scmp(i32 %x, i32 %y)
+  ret i8 %1
+}
+
+; Negative test: ranges overlap
+define i8 @ucmp_3(i32 %x, i32 %y) {
+  ; X is within [4, +INF)
+  %cond1 = icmp uge i32 %x, 4
+  call void @llvm.assume(i1 %cond1)
+  ; Y is within [0, 6)
+  %cond2 = icmp ult i32 %y, 6
+  call void @llvm.assume(i1 %cond2)
+
+  %1 = call i8 @llvm.ucmp(i32 %x, i32 %y)
+  ret i8 %1
+}
+
+define i8 @scmp_3(i32 %x, i32 %y) {
+  ; X is within [2, +INF)
+  %cond1 = icmp sge i32 %x, 2
+  call void @llvm.assume(i1 %cond1)
+  ; Y is within [-INF, 4)
+  %cond2 = icmp slt i32 %y, 4
+  call void @llvm.assume(i1 %cond2)
+
+  %1 = call i8 @llvm.scmp(i32 %x, i32 %y)
+  ret i8 %1
+}
+
+; Negative test: mismatched signedness of range-establishing comparisons and
+;                of the intrinsic
+define i8 @ucmp_4(i32 %x, i32 %y) {
+  ; X is within [4, +INF)
+  %cond1 = icmp sge i32 %x, 4
+  call void @llvm.assume(i1 %cond1)
+  ; Y is within [0, 4)
+  %cond2 = icmp slt i32 %y, 4
+  call void @llvm.assume(i1 %cond2)
+
+  %1 = call i8 @llvm.ucmp(i32 %x, i32 %y)
+  ret i8 %1
+}
+
+define i8 @scmp_4(i32 %x, i32 %y) {
+  ; X is within [4, +INF)
+  %cond1 = icmp uge i32 %x, 4
+  call void @llvm.assume(i1 %cond1)
+  ; Y is within [0, 4)
+  %cond2 = icmp ult i32 %y, 4
+  call void @llvm.assume(i1 %cond2)
+
+  %1 = call i8 @llvm.scmp(i32 %x, i32 %y)
+  ret i8 %1
+}