JIT: Boost inversion for oversize loops with bounds checks (#118078)

Addresses some of the regressions in #116486. If a loop has bounds checks in it, it might benefit from cloning, so perhaps we ought to tolerate going a bit over the inversion size limit to enable downstream optimizations. I ought to do something for GDV checks; perhaps as a follow-up, I'll move the checks in loop cloning to something I can reuse here. The 1.25x figure isn't all that scientific -- I found it to be the smallest factor necessary to make a dent in the regressions from less cloning I looked at, and despite the size increases, it's a net PerfScore win across collections.

Author: amanasifkhalid

src/coreclr/jit/compiler.hpp

@@ -5573,14 +5573,7 @@ bool Compiler::optLoopComplexityExceeds(FlowGraphNaturalLoop* loop, unsigned lim
                                                                     : BasicBlockVisit::Continue;
     });
 
-    if (result == BasicBlockVisit::Abort)
-    {
-        JITDUMP("Loop " FMT_LP ": exceeds complexity limit %u\n", loop->GetIndex(), limit);
-        return true;
-    }
-
-    JITDUMP("Loop " FMT_LP ": complexity %u does not exceed limit %u\n", loop->GetIndex(), loopComplexity, limit);
-    return false;
+    return (result == BasicBlockVisit::Abort);
 }
 
 /*****************************************************************************/

src/coreclr/jit/loopcloning.cpp

@@ -3108,6 +3108,7 @@ PhaseStatus Compiler::optCloneLoops()
             // then the method returns false.
             else if ((sizeLimit >= 0) && optLoopComplexityExceeds(loop, (unsigned)sizeLimit, countNode))
             {
+                JITDUMP(FMT_LP " exceeds cloning size limit %d\n", loop->GetIndex(), sizeLimit);
                 context.CancelLoopOptInfo(loop->GetIndex());
             }
         }

src/coreclr/jit/optimizer.cpp

@@ -1903,14 +1903,45 @@ bool Compiler::optTryInvertWhileLoop(FlowGraphNaturalLoop* loop)
 
     // Check if loop is small enough to consider for inversion.
     // Large loops are less likely to benefit from inversion.
-    const int sizeLimit = JitConfig.JitLoopInversionSizeLimit();
-    auto      countNode = [](GenTree* tree) -> unsigned {
-        return 1;
-    };
+    const int invertSizeLimit = JitConfig.JitLoopInversionSizeLimit();
+    if (invertSizeLimit >= 0)
+    {
+        const int cloneSizeLimit          = JitConfig.JitCloneLoopsSizeLimit();
+        bool      mightBenefitFromCloning = false;
+        unsigned  loopSize                = 0;
+
+        // Loops with bounds checks can benefit from cloning, which depends on inversion running.
+        // Thus, we will try to proceed with inversion for slightly oversize loops if they show potential for cloning.
+        auto countNode = [&mightBenefitFromCloning, &loopSize](GenTree* tree) -> unsigned {
+            mightBenefitFromCloning |= tree->OperIs(GT_BOUNDS_CHECK);
+            loopSize++;
+            return 1;
+        };
 
-    if ((sizeLimit >= 0) && optLoopComplexityExceeds(loop, (unsigned)sizeLimit, countNode))
-    {
-        return false;
+        optLoopComplexityExceeds(loop, (unsigned)max(invertSizeLimit, cloneSizeLimit), countNode);
+        if (loopSize > (unsigned)invertSizeLimit)
+        {
+            // Don't try to invert oversize loops if they don't show cloning potential,
+            // or if they're too big to be cloned anyway.
+            JITDUMP(FMT_LP " exceeds inversion size limit of %d\n", loop->GetIndex(), invertSizeLimit);
+            const bool tooBigToClone = (cloneSizeLimit >= 0) && (loopSize > (unsigned)cloneSizeLimit);
+            if (!mightBenefitFromCloning || tooBigToClone)
+            {
+                JITDUMP("No inversion for " FMT_LP ": %s\n", loop->GetIndex(),
+                        tooBigToClone ? "too big to clone" : "unlikely to benefit from cloning");
+                return false;
+            }
+
+            // If the loop shows cloning potential, tolerate some excess size.
+            const unsigned liberalInvertSizeLimit = (unsigned)(invertSizeLimit * 1.25);
+            if (loopSize > liberalInvertSizeLimit)
+            {
+                JITDUMP(FMT_LP " might benefit from cloning, but is too large to invert.\n", loop->GetIndex());
+                return false;
+            }
+
+            JITDUMP(FMT_LP " might benefit from cloning. Continuing.\n", loop->GetIndex());
+        }
     }
 
     unsigned estDupCostSz = 0;