Fix perf issues discovered in "For software performance, can you always trust inlining" blog (#61408)

Author: EgorBo

src/coreclr/jit/fgbasic.cpp

@@ -894,12 +894,10 @@ void Compiler::fgFindJumpTargets(const BYTE* codeAddr, IL_OFFSET codeSize, Fixed
     const bool  isForceInline          = (info.compFlags & CORINFO_FLG_FORCEINLINE) != 0;
     const bool  makeInlineObservations = (compInlineResult != nullptr);
     const bool  isInlining             = compIsForInlining();
-    const bool  isPreJit               = opts.jitFlags->IsSet(JitFlags::JIT_FLAG_PREJIT);
-    const bool  isTier1                = opts.jitFlags->IsSet(JitFlags::JIT_FLAG_TIER1);
     unsigned    retBlocks              = 0;
     int         prefixFlags            = 0;
     bool        preciseScan            = makeInlineObservations && compInlineResult->GetPolicy()->RequiresPreciseScan();
-    const bool  resolveTokens          = preciseScan && (isPreJit || isTier1);
+    const bool  resolveTokens          = preciseScan;
 
     // Track offsets where IL instructions begin in DEBUG builds. Used to
     // validate debug info generated by the JIT.

src/coreclr/jit/gentree.cpp

@@ -2620,29 +2620,58 @@ unsigned Compiler::gtSetMultiOpOrder(GenTreeMultiOp* multiOp)
     unsigned level  = 0;
     unsigned lvl2   = 0;
 
-#if defined(FEATURE_HW_INTRINSICS) && defined(TARGET_XARCH)
-    if (multiOp->OperIs(GT_HWINTRINSIC) && (multiOp->GetOperandCount() == 1) &&
-        multiOp->AsHWIntrinsic()->OperIsMemoryLoadOrStore())
+#if defined(FEATURE_HW_INTRINSICS)
+    if (multiOp->OperIs(GT_HWINTRINSIC))
     {
-        costEx = IND_COST_EX;
-        costSz = 2;
+        GenTreeHWIntrinsic* hwTree = multiOp->AsHWIntrinsic();
+#if defined(TARGET_XARCH)
+        if ((hwTree->GetOperandCount() == 1) && hwTree->OperIsMemoryLoadOrStore())
+        {
+            costEx = IND_COST_EX;
+            costSz = 2;
 
-        GenTree* addr = multiOp->Op(1)->gtEffectiveVal();
-        level         = gtSetEvalOrder(addr);
+            GenTree* addr = hwTree->Op(1)->gtEffectiveVal();
+            level         = gtSetEvalOrder(addr);
 
-        // See if we can form a complex addressing mode.
-        if (addr->OperIs(GT_ADD) && gtMarkAddrMode(addr, &costEx, &costSz, multiOp->TypeGet()))
-        {
-            // Nothing to do, costs have been set.
+            // See if we can form a complex addressing mode.
+            if (addr->OperIs(GT_ADD) && gtMarkAddrMode(addr, &costEx, &costSz, hwTree->TypeGet()))
+            {
+                // Nothing to do, costs have been set.
+            }
+            else
+            {
+                costEx += addr->GetCostEx();
+                costSz += addr->GetCostSz();
+            }
+
+            hwTree->SetCosts(costEx, costSz);
+            return level;
         }
-        else
+#endif
+        switch (hwTree->GetHWIntrinsicId())
         {
-            costEx += addr->GetCostEx();
-            costSz += addr->GetCostSz();
+#if defined(TARGET_XARCH)
+            case NI_Vector128_Create:
+            case NI_Vector256_Create:
+#elif defined(TARGET_ARM64)
+            case NI_Vector64_Create:
+            case NI_Vector128_Create:
+#endif
+            {
+                if ((hwTree->GetOperandCount() == 1) && hwTree->Op(1)->OperIsConst())
+                {
+                    // Vector.Create(cns) is cheap but not that cheap to be (1,1)
+                    costEx = IND_COST_EX;
+                    costSz = 2;
+                    level  = gtSetEvalOrder(hwTree->Op(1));
+                    hwTree->SetCosts(costEx, costSz);
+                    return level;
+                }
+                break;
+            }
+            default:
+                break;
         }
-
-        multiOp->SetCosts(costEx, costSz);
-        return level;
     }
 #endif // defined(FEATURE_SIMD) || defined(FEATURE_HW_INTRINSICS)