[llvm] 22a280d - LICM: teach hoistMinMax about samesign (#122730)

[via llvm-commits]

Author: Ramkumar Ramachandra
Date: 2025-01-13T20:03:58Z
New Revision: 22a280d3924abf2cd2587dcff247bba884303c57

URL: https://github.com/llvm/llvm-project/commit/22a280d3924abf2cd2587dcff247bba884303c57
DIFF: https://github.com/llvm/llvm-project/commit/22a280d3924abf2cd2587dcff247bba884303c57.diff

LOG: LICM: teach hoistMinMax about samesign (#122730)

Follow up on 4a0d53a (PatternMatch: migrate to CmpPredicate) to get rid
of one of the FIXMEs it introduced by replacing a predicate comparison
with CmpPredicate::getMatching.

Added: 
    

Modified: 
    llvm/lib/Transforms/Scalar/LICM.cpp
    llvm/test/Transforms/LICM/min_max.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Scalar/LICM.cpp b/llvm/lib/Transforms/Scalar/LICM.cpp
index a5d5eecb1ebf82..0bab01904406ea 100644
--- a/llvm/lib/Transforms/Scalar/LICM.cpp
+++ b/llvm/lib/Transforms/Scalar/LICM.cpp
@@ -2453,16 +2453,17 @@ static bool hoistMinMax(Instruction &I, Loop &L, ICFLoopSafetyInfo &SafetyInfo,
   if (!MatchICmpAgainstInvariant(Cond1, P1, LHS1, RHS1) ||
       !MatchICmpAgainstInvariant(Cond2, P2, LHS2, RHS2))
     return false;
-  // FIXME: Use CmpPredicate::getMatching here.
-  if (P1 != static_cast<CmpInst::Predicate>(P2) || LHS1 != LHS2)
+  auto MatchingPred = CmpPredicate::getMatching(P1, P2);
+  if (!MatchingPred || LHS1 != LHS2)
     return false;
 
   // Everything is fine, we can do the transform.
-  bool UseMin = ICmpInst::isLT(P1) || ICmpInst::isLE(P1);
+  bool UseMin = ICmpInst::isLT(*MatchingPred) || ICmpInst::isLE(*MatchingPred);
   assert(
-      (UseMin || ICmpInst::isGT(P1) || ICmpInst::isGE(P1)) &&
+      (UseMin || ICmpInst::isGT(*MatchingPred) ||
+       ICmpInst::isGE(*MatchingPred)) &&
       "Relational predicate is either less (or equal) or greater (or equal)!");
-  Intrinsic::ID id = ICmpInst::isSigned(P1)
+  Intrinsic::ID id = ICmpInst::isSigned(*MatchingPred)
                          ? (UseMin ? Intrinsic::smin : Intrinsic::smax)
                          : (UseMin ? Intrinsic::umin : Intrinsic::umax);
   auto *Preheader = L.getLoopPreheader();
@@ -2475,11 +2476,12 @@ static bool hoistMinMax(Instruction &I, Loop &L, ICFLoopSafetyInfo &SafetyInfo,
   if (isa<SelectInst>(I))
     RHS2 = Builder.CreateFreeze(RHS2, RHS2->getName() + ".fr");
   Value *NewRHS = Builder.CreateBinaryIntrinsic(
-      id, RHS1, RHS2, nullptr, StringRef("invariant.") +
-                                   (ICmpInst::isSigned(P1) ? "s" : "u") +
-                                   (UseMin ? "min" : "max"));
+      id, RHS1, RHS2, nullptr,
+      StringRef("invariant.") +
+          (ICmpInst::isSigned(*MatchingPred) ? "s" : "u") +
+          (UseMin ? "min" : "max"));
   Builder.SetInsertPoint(&I);
-  ICmpInst::Predicate P = P1;
+  ICmpInst::Predicate P = *MatchingPred;
   if (Inverse)
     P = ICmpInst::getInversePredicate(P);
   Value *NewCond = Builder.CreateICmp(P, LHS1, NewRHS);

diff  --git a/llvm/test/Transforms/LICM/min_max.ll b/llvm/test/Transforms/LICM/min_max.ll
index c2bf0a7f20cc12..04f309b785ebcd 100644
--- a/llvm/test/Transforms/LICM/min_max.ll
+++ b/llvm/test/Transforms/LICM/min_max.ll
@@ -242,6 +242,35 @@ exit:
   ret i32 %iv
 }
 
+define i32 @test_sgt_samesign(i32 %start, i32 %inv_1, i32 %inv_2) {
+; CHECK-LABEL: @test_sgt_samesign(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[INVARIANT_SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[INV_1:%.*]], i32 [[INV_2:%.*]])
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[START:%.*]], [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
+; CHECK-NEXT:    [[LOOP_COND:%.*]] = icmp sgt i32 [[IV]], [[INVARIANT_SMAX]]
+; CHECK-NEXT:    [[IV_NEXT]] = add i32 [[IV]], 1
+; CHECK-NEXT:    br i1 [[LOOP_COND]], label [[LOOP]], label [[EXIT:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[IV_LCSSA:%.*]] = phi i32 [ [[IV]], [[LOOP]] ]
+; CHECK-NEXT:    ret i32 [[IV_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i32 [%start, %entry], [%iv.next, %loop]
+  %cmp_1 = icmp samesign ugt i32 %iv, %inv_1
+  %cmp_2 = icmp sgt i32 %iv, %inv_2
+  %loop_cond = and i1 %cmp_1, %cmp_2
+  %iv.next = add i32 %iv, 1
+  br i1 %loop_cond, label %loop, label %exit
+
+exit:
+  ret i32 %iv
+}
+
 ; turn to %iv >=s smax(inv_1, inv_2) and hoist it out of loop.
 define i32 @test_sge(i32 %start, i32 %inv_1, i32 %inv_2) {
 ; CHECK-LABEL: @test_sge(
@@ -272,6 +301,35 @@ exit:
   ret i32 %iv
 }
 
+define i32 @test_sge_samesign(i32 %start, i32 %inv_1, i32 %inv_2) {
+; CHECK-LABEL: @test_sge_samesign(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[INVARIANT_SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[INV_1:%.*]], i32 [[INV_2:%.*]])
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[START:%.*]], [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
+; CHECK-NEXT:    [[LOOP_COND:%.*]] = icmp sge i32 [[IV]], [[INVARIANT_SMAX]]
+; CHECK-NEXT:    [[IV_NEXT]] = add i32 [[IV]], 1
+; CHECK-NEXT:    br i1 [[LOOP_COND]], label [[LOOP]], label [[EXIT:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[IV_LCSSA:%.*]] = phi i32 [ [[IV]], [[LOOP]] ]
+; CHECK-NEXT:    ret i32 [[IV_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i32 [%start, %entry], [%iv.next, %loop]
+  %cmp_1 = icmp sge i32 %iv, %inv_1
+  %cmp_2 = icmp samesign uge i32 %iv, %inv_2
+  %loop_cond = and i1 %cmp_1, %cmp_2
+  %iv.next = add i32 %iv, 1
+  br i1 %loop_cond, label %loop, label %exit
+
+exit:
+  ret i32 %iv
+}
+
 ; Turn OR to AND and handle accordingly.
 define i32 @test_ult_inv(i32 %start, i32 %inv_1, i32 %inv_2) {
 ; CHECK-LABEL: @test_ult_inv(