From c1de4ba68eccad2d450ab2fb5ba39fdf498a373e Mon Sep 17 00:00:00 2001
From: Egor Bogatov <egorbo@gmail.com>
Date: Sun, 7 Jan 2024 16:45:46 +0100
Subject: [PATCH] PGO: Profiled SequenceEqual (#96571)

---
 src/coreclr/jit/compiler.h        | 11 ++++++++-
 src/coreclr/jit/fgprofile.cpp     | 12 ++++++----
 src/coreclr/jit/importercalls.cpp | 39 +++++++++++++++++++++++++++++--
 3 files changed, 55 insertions(+), 7 deletions(-)

diff --git a/src/coreclr/jit/compiler.h b/src/coreclr/jit/compiler.h
index 05bd2e53dfef8..7cb0dfb564f88 100644
--- a/src/coreclr/jit/compiler.h
+++ b/src/coreclr/jit/compiler.h
@@ -9515,7 +9515,9 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
     {
         Memset,
         Memcpy,
-        Memmove
+        Memmove,
+        ProfiledMemmove,
+        ProfiledMemcmp
     };
 
     //------------------------------------------------------------------------
@@ -9589,6 +9591,13 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
             threshold = maxRegSize * 4;
         }
 
+        // For profiled memcmp/memmove we don't want to unroll too much as it's just a guess,
+        // and it works better for small sizes.
+        if ((type == UnrollKind::ProfiledMemcmp) || (type == UnrollKind::ProfiledMemmove))
+        {
+            threshold = maxRegSize * 2;
+        }
+
         return threshold;
     }
 
diff --git a/src/coreclr/jit/fgprofile.cpp b/src/coreclr/jit/fgprofile.cpp
index f3fe032a4c363..a4d9c21722471 100644
--- a/src/coreclr/jit/fgprofile.cpp
+++ b/src/coreclr/jit/fgprofile.cpp
@@ -1950,9 +1950,13 @@ class ValueHistogramProbeVisitor final : public GenTreeVisitor<ValueHistogramPro
     Compiler::fgWalkResult PreOrderVisit(GenTree** use, GenTree* user)
     {
         GenTree* const node = *use;
-        if (node->IsCall() && node->AsCall()->IsSpecialIntrinsic(m_compiler, NI_System_Buffer_Memmove))
+        if (node->IsCall() && node->AsCall()->IsSpecialIntrinsic())
         {
-            m_functor(m_compiler, node);
+            const NamedIntrinsic ni = m_compiler->lookupNamedIntrinsic(node->AsCall()->gtCallMethHnd);
+            if ((ni == NI_System_Buffer_Memmove) || (ni == NI_System_SpanHelpers_SequenceEqual))
+            {
+                m_functor(m_compiler, node);
+            }
         }
         return Compiler::WALK_CONTINUE;
     }
@@ -2276,8 +2280,8 @@ class ValueHistogramProbeInserter
             return;
         }
 
-        // Only Buffer.Memmove call is currently expected
-        assert(node->IsCall() && (node->AsCall()->IsSpecialIntrinsic(compiler, NI_System_Buffer_Memmove)));
+        assert(node->AsCall()->IsSpecialIntrinsic(compiler, NI_System_Buffer_Memmove) ||
+               node->AsCall()->IsSpecialIntrinsic(compiler, NI_System_SpanHelpers_SequenceEqual));
 
         const ICorJitInfo::PgoInstrumentationSchema& countEntry = m_schema[*m_currentSchemaIndex];
         if (countEntry.ILOffset !=
diff --git a/src/coreclr/jit/importercalls.cpp b/src/coreclr/jit/importercalls.cpp
index 4692bd53fc60c..cf698110ef129 100644
--- a/src/coreclr/jit/importercalls.cpp
+++ b/src/coreclr/jit/importercalls.cpp
@@ -1440,6 +1440,32 @@ var_types Compiler::impImportCall(OPCODE                  opcode,
                 {
                     impSpillSideEffects(true, CHECK_SPILL_ALL DEBUGARG("non-inline candidate call"));
                 }
+
+                if (JitConfig.JitProfileValues() && call->IsCall() &&
+                    call->AsCall()->IsSpecialIntrinsic(this, NI_System_SpanHelpers_SequenceEqual))
+                {
+                    if (opts.IsOptimizedWithProfile())
+                    {
+                        call = impDuplicateWithProfiledArg(call->AsCall(), rawILOffset);
+                        if (call->OperIs(GT_QMARK))
+                        {
+                            // QMARK has to be a root node
+                            unsigned tmp = lvaGrabTemp(true DEBUGARG("Grabbing temp for Qmark"));
+                            impStoreTemp(tmp, call, CHECK_SPILL_ALL);
+                            call = gtNewLclvNode(tmp, call->TypeGet());
+                        }
+                    }
+                    else if (opts.IsInstrumented())
+                    {
+                        // We might want to instrument it for optimized versions too, but we don't currently.
+                        HandleHistogramProfileCandidateInfo* pInfo =
+                            new (this, CMK_Inlining) HandleHistogramProfileCandidateInfo;
+                        pInfo->ilOffset                                       = rawILOffset;
+                        pInfo->probeIndex                                     = 0;
+                        call->AsCall()->gtHandleHistogramProfileCandidateInfo = pInfo;
+                        compCurBB->SetFlags(BBF_HAS_VALUE_PROFILE);
+                    }
+                }
             }
 
             //-------------------------------------------------------------------------
@@ -1536,7 +1562,15 @@ GenTree* Compiler::impDuplicateWithProfiledArg(GenTreeCall* call, IL_OFFSET ilOf
             argNum = 2;
 
             minValue = 1; // TODO: enable for 0 as well.
-            maxValue = (ssize_t)getUnrollThreshold(Memmove);
+            maxValue = (ssize_t)getUnrollThreshold(ProfiledMemmove);
+        }
+        else if (call->IsSpecialIntrinsic(this, NI_System_SpanHelpers_SequenceEqual))
+        {
+            // dst(0), src(1), len(2)
+            argNum = 2;
+
+            minValue = 1; // TODO: enable for 0 as well.
+            maxValue = (ssize_t)getUnrollThreshold(ProfiledMemcmp);
         }
         else
         {
@@ -2729,7 +2763,8 @@ GenTree* Compiler::impIntrinsic(GenTree*                newobjThis,
                 break;
 
             case NI_System_Buffer_Memmove:
-                // We're going to instrument this one
+            case NI_System_SpanHelpers_SequenceEqual:
+                // We're going to instrument these
                 betterToExpand = opts.IsInstrumented();
                 break;