[llvm] [SCEV] Remove expensive inverted reasoning from isImpliedCond. (PR #227302)

[Florian Hahn via llvm-commits]

https://github.com/fhahn created https://github.com/llvm/llvm-project/pull/227302

The constructing the inverted expressions and the additional reasoning is quite expensive (0.35% for CTMark O3), for marginal gain (3 small regressions on https://github.com/dtcxzyw/llvm-opt-benchmark-nightly/pull/1494)

I think that should allow us to spend compile-time on SCEV on areas with higher impact.

2 of those could be recovered by constant-based reasoning about the inverted condition.

Compile-time improvements:
stage1-O3: -0.35%
stage1-ReleaseThinLTO: -0.32%
stage1-ReleaseLTO-g: -0.28%
stage1-aarch64-O3: -0.31%
stage2-O3: -0.32%
stage2-clang: -0.06%

https://llvm-compile-time-tracker.com/compare.php?from=197c16add5b16f4bd68c4f4bae115723e74330e2&to=30dc497653f0daafd602d29a1ead334599055619&stat=instructions:u

>From c9ca0323010ad23f178ff45ba01bee2342664312 Mon Sep 17 00:00:00 2001
From: Florian Hahn <flo at fhahn.com>
Date: Fri, 25 Sep 2026 20:21:54 +0100
Subject: [PATCH] [SCEV] Remove expensive inverted reasoning from
 isImpliedCond.

The constructing the inverted expressions and the additional reasoning
is quite expensive (0.35% for CTMark O3), for marginal gain
(3 small regressions on https://github.com/dtcxzyw/llvm-opt-benchmark-nightly/pull/1494)

I think that should allow us to spend compile-time on SCEV on areas with
higher impact.

2 of those could be recovered by constant-based reasoning about the
inverted condition.

Compile-time improvements:
stage1-O3: -0.35%
stage1-ReleaseThinLTO: -0.32%
stage1-ReleaseLTO-g: -0.28%
stage1-aarch64-O3: -0.31%
stage2-O3: -0.32%
stage2-clang: -0.06%

https://llvm-compile-time-tracker.com/compare.php?from=197c16add5b16f4bd68c4f4bae115723e74330e2&to=30dc497653f0daafd602d29a1ead334599055619&stat=instructions:u
---
 llvm/lib/Analysis/ScalarEvolution.cpp | 21 ++-------------------
 1 file changed, 2 insertions(+), 19 deletions(-)

diff --git a/llvm/lib/Analysis/ScalarEvolution.cpp b/llvm/lib/Analysis/ScalarEvolution.cpp
index cbadeaf7b347d..c03e2b5b206f7 100644
--- a/llvm/lib/Analysis/ScalarEvolution.cpp
+++ b/llvm/lib/Analysis/ScalarEvolution.cpp
@@ -12256,31 +12256,14 @@ bool ScalarEvolution::isImpliedCondBalancedTypes(
     // using one of the following ways:
     // 1.  LHS Pred      RHS  <-   FoundRHS Pred      FoundLHS
     // 2.  RHS SwapPred  LHS  <-   FoundLHS SwapPred  FoundRHS
-    // 3.  LHS Pred      RHS  <-  ~FoundLHS Pred     ~FoundRHS
-    // 4. ~LHS SwapPred ~RHS  <-   FoundLHS SwapPred  FoundRHS
-    // Forms 1. and 2. require swapping the operands of one condition. Don't
-    // do this if it would break canonical constant/addrec ordering.
+    // Both require swapping the operands of one condition. Don't do this if it
+    // would break canonical constant/addrec ordering.
     if (!isa<SCEVConstant>(RHS) && !isa<SCEVAddRecExpr>(LHS))
       return isImpliedCondOperands(ICmpInst::getSwappedCmpPredicate(*P), RHS,
                                    LHS, FoundLHS, FoundRHS, CtxI);
     if (!isa<SCEVConstant>(FoundRHS) && !isa<SCEVAddRecExpr>(FoundLHS))
       return isImpliedCondOperands(*P, LHS, RHS, FoundRHS, FoundLHS, CtxI);
 
-    // There's no clear preference between forms 3. and 4., try both.  Avoid
-    // forming getNotSCEV of pointer values as the resulting subtract is
-    // not legal.
-    if (!LHS->getType()->isPointerTy() && !RHS->getType()->isPointerTy() &&
-        isImpliedCondOperands(ICmpInst::getSwappedCmpPredicate(*P),
-                              getNotSCEV(LHS), getNotSCEV(RHS), FoundLHS,
-                              FoundRHS, CtxI))
-      return true;
-
-    if (!FoundLHS->getType()->isPointerTy() &&
-        !FoundRHS->getType()->isPointerTy() &&
-        isImpliedCondOperands(*P, LHS, RHS, getNotSCEV(FoundLHS),
-                              getNotSCEV(FoundRHS), CtxI))
-      return true;
-
     return false;
   }