[CVP] Refactor processMinMaxIntrinsic to check non-strict predicate…

… in both directions (#82596) This patch uses `getConstantRangeAtUse` in `processMinMaxIntrinsic` to address the comment #82478 (comment). After this patch we can reuse the range result in #82478.
llvm · Feb 22, 2024 · 3ef63a7 · 3ef63a7
1 parent 4235e44
commit 3ef63a7
Show file tree

Hide file tree

Showing 2 changed files with 76 additions and 13 deletions.
diff --git a/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp b/llvm/lib/Transforms/Scalar/CorrelatedValuePropagation.cpp
@@ -530,15 +530,23 @@ static bool processAbsIntrinsic(IntrinsicInst *II, LazyValueInfo *LVI) {
 // See if this min/max intrinsic always picks it's one specific operand.
 static bool processMinMaxIntrinsic(MinMaxIntrinsic *MM, LazyValueInfo *LVI) {
   CmpInst::Predicate Pred = CmpInst::getNonStrictPredicate(MM->getPredicate());
-  LazyValueInfo::Tristate Result = LVI->getPredicateAt(
-      Pred, MM->getLHS(), MM->getRHS(), MM, /*UseBlockValue=*/true);
-  if (Result == LazyValueInfo::Unknown)
-    return false;
-
-  ++NumMinMax;
-  MM->replaceAllUsesWith(MM->getOperand(!Result));
-  MM->eraseFromParent();
-  return true;
+  ConstantRange LHS_CR = LVI->getConstantRangeAtUse(MM->getOperandUse(0),
+                                                    /*UndefAllowed*/ false);
+  ConstantRange RHS_CR = LVI->getConstantRangeAtUse(MM->getOperandUse(1),
+                                                    /*UndefAllowed*/ false);
+  if (LHS_CR.icmp(Pred, RHS_CR)) {
+    ++NumMinMax;
+    MM->replaceAllUsesWith(MM->getLHS());
+    MM->eraseFromParent();
+    return true;
+  }
+  if (RHS_CR.icmp(Pred, LHS_CR)) {
+    ++NumMinMax;
+    MM->replaceAllUsesWith(MM->getRHS());
+    MM->eraseFromParent();
+    return true;
+  }
+  return false;
 }
 
 // Rewrite this with.overflow intrinsic as non-overflowing.

diff --git a/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll b/llvm/test/Transforms/CorrelatedValuePropagation/min-max.ll
@@ -71,7 +71,6 @@ define i8 @test6(i8 %x) {
 ; CHECK-LABEL: @test6(
 ; CHECK-NEXT:    [[LIM:%.*]] = icmp uge i8 [[X:%.*]], 42
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[LIM]])
-; CHECK-NEXT:    [[R:%.*]] = call i8 @llvm.umin.i8(i8 [[X]], i8 42)
 ; CHECK-NEXT:    ret i8 42
 ;
   %lim = icmp uge i8 %x, 42
@@ -119,7 +118,6 @@ define i8 @test10(i8 %x) {
 ; CHECK-LABEL: @test10(
 ; CHECK-NEXT:    [[LIM:%.*]] = icmp ule i8 [[X:%.*]], 42
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[LIM]])
-; CHECK-NEXT:    [[R:%.*]] = call i8 @llvm.umax.i8(i8 [[X]], i8 42)
 ; CHECK-NEXT:    ret i8 42
 ;
   %lim = icmp ule i8 %x, 42
@@ -167,7 +165,6 @@ define i8 @test14(i8 %x) {
 ; CHECK-LABEL: @test14(
 ; CHECK-NEXT:    [[LIM:%.*]] = icmp sge i8 [[X:%.*]], 42
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[LIM]])
-; CHECK-NEXT:    [[R:%.*]] = call i8 @llvm.smin.i8(i8 [[X]], i8 42)
 ; CHECK-NEXT:    ret i8 42
 ;
   %lim = icmp sge i8 %x, 42
@@ -215,7 +212,6 @@ define i8 @test18(i8 %x) {
 ; CHECK-LABEL: @test18(
 ; CHECK-NEXT:    [[LIM:%.*]] = icmp sle i8 [[X:%.*]], 42
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[LIM]])
-; CHECK-NEXT:    [[R:%.*]] = call i8 @llvm.smax.i8(i8 [[X]], i8 42)
 ; CHECK-NEXT:    ret i8 42
 ;
   %lim = icmp sle i8 %x, 42
@@ -235,3 +231,62 @@ define i8 @test19(i8 %x) {
   %r = call i8 @llvm.smax(i8 %x, i8 42)
   ret i8 %r
 }
+
+declare void @body(i32)
+
+define void @test_bidirectional() {
+; CHECK-LABEL: @test_bidirectional(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
+; CHECK:       for.body:
+; CHECK-NEXT:    [[INDVAR:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    call void @body(i32 65535)
+; CHECK-NEXT:    [[INC]] = add nsw i32 [[INDVAR]], 1
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i32 [[INDVAR]], 65535
+; CHECK-NEXT:    br i1 [[CMP]], label [[FOR_BODY]], label [[EXIT:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %for.body
+
+for.body:
+  %indvar = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %smax = call i32 @llvm.smax.i32(i32 %indvar, i32 65535)
+  call void @body(i32 %smax)
+  %inc = add nsw i32 %indvar, 1
+  %cmp = icmp slt i32 %indvar, 65535
+  br i1 %cmp, label %for.body, label %exit
+
+exit:
+  ret void
+}
+
+define i64 @test_at_use(i1 %cond, i64 %x) {
+; CHECK-LABEL: @test_at_use(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 [[COND:%.*]], label [[BB1:%.*]], label [[IF_END:%.*]]
+; CHECK:       bb1:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i64 [[X:%.*]], 0
+; CHECK-NEXT:    br i1 [[CMP]], label [[IF_THEN:%.*]], label [[IF_END]]
+; CHECK:       if.then:
+; CHECK-NEXT:    ret i64 0
+; CHECK:       if.end:
+; CHECK-NEXT:    [[PHI:%.*]] = phi i64 [ [[X]], [[BB1]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECK-NEXT:    ret i64 [[PHI]]
+;
+entry:
+  br i1 %cond, label %bb1, label %if.end
+
+bb1:
+  %val = call i64 @llvm.smax.i64(i64 %x, i64 -1)
+  %cmp = icmp slt i64 %x, 0
+  br i1 %cmp, label %if.then, label %if.end
+
+if.then:
+  ret i64 0
+
+if.end:
+  %phi = phi i64 [%val, %bb1], [0, %entry]
+  ret i64 %phi
+}