Added PerfScore support for Arm64

Based upon arm_cortex_a55_software_optimization_guide_v2.pdf

src/coreclr/src/jit/emit.cpp

@@ -1107,7 +1107,7 @@ float emitter::insEvaluateExecutionCost(instrDesc* id)
 
     if (memAccessKind == PERFSCORE_MEMORY_WRITE)
     {
-        // We assume that we won't read back from memory for the next WR_GENERAL (3) cycles
+        // We assume that we won't read back from memory for the next WR_GENERAL cycles
         // Thus we normally won't pay latency costs for writes.
         latency = max(0.0f, latency - PERFSCORE_LATENCY_WR_GENERAL);
     }
@@ -1121,6 +1121,34 @@ float emitter::insEvaluateExecutionCost(instrDesc* id)
     return max(throughput, latency);
 }
 
+//------------------------------------------------------------------------------------
+// perfScoreUnhandledInstruction:
+//    Helper method used to report an unhandled instruction
+//
+// Arguments:
+//    id  - The current instruction descriptor to be evaluated
+//    pResult - pointer to struct holding the instruction characteristics
+//              if we return these are updated with default values
+//
+// Notes:
+//     When validating that the PerfScore handles every instruction.
+//     the #if 0 block is changed into a #ifdef DEBUG
+//     We will print the instruction and instruction group
+//     and instead of returning we will assert
+//
+//     Otherwise we will return default latencies of 1 cycle.
+//
+void emitter::perfScoreUnhandledInstruction(instrDesc* id, insExecutionCharacteristics* pResult)
+{
+// Change this to #ifdef DEBUG to assert on any unhandled instructions
+#if 0
+    printf("PerfScore: unhandled instruction: %s, format %s", codeGen->genInsName(id->idIns()), emitIfName(id->idInsFmt()));
+    assert(!"PerfScore: unhandled instruction");
+#endif
+    pResult->insThroughput = PERFSCORE_THROUGHPUT_1C;
+    pResult->insLatency    = PERFSCORE_LATENCY_1C;
+}
+
 #endif // defined(DEBUG) || defined(LATE_DISASM)
 
 //----------------------------------------------------------------------------------------

src/coreclr/src/jit/emit.h

@@ -1233,19 +1233,17 @@ class emitter
 #define PERFSCORE_THROUGHPUT_4C 4.0f   // slower - 4 cycles
 #define PERFSCORE_THROUGHPUT_5C 5.0f   // slower - 5 cycles
 #define PERFSCORE_THROUGHPUT_6C 6.0f   // slower - 6 cycles
+#define PERFSCORE_THROUGHPUT_9C 9.0f   // slower - 9 cycles
 #define PERFSCORE_THROUGHPUT_10C 10.0f // slower - 10 cycles
 #define PERFSCORE_THROUGHPUT_13C 13.0f // slower - 13 cycles
+#define PERFSCORE_THROUGHPUT_19C 19.0f // slower - 19 cycles
 #define PERFSCORE_THROUGHPUT_25C 25.0f // slower - 25 cycles
 #define PERFSCORE_THROUGHPUT_33C 33.0f // slower - 33 cycles
 #define PERFSCORE_THROUGHPUT_52C 52.0f // slower - 52 cycles
 #define PERFSCORE_THROUGHPUT_57C 57.0f // slower - 57 cycles
 
-#define PERFSCORE_THROUGHPUT_DEFAULT PERFSCORE_THROUGHPUT_1C
-
 #define PERFSCORE_LATENCY_ILLEGAL -1024.0f
 
-#define PERFSCORE_LATENCY_DEFAULT 1.0f
-
 #define PERFSCORE_LATENCY_ZERO 0.0f
 #define PERFSCORE_LATENCY_1C 1.0f
 #define PERFSCORE_LATENCY_2C 2.0f
@@ -1260,7 +1258,11 @@ class emitter
 #define PERFSCORE_LATENCY_11C 11.0f
 #define PERFSCORE_LATENCY_12C 12.0f
 #define PERFSCORE_LATENCY_13C 13.0f
+#define PERFSCORE_LATENCY_15C 15.0f
 #define PERFSCORE_LATENCY_16C 16.0f
+#define PERFSCORE_LATENCY_18C 18.0f
+#define PERFSCORE_LATENCY_20C 20.0f
+#define PERFSCORE_LATENCY_22C 22.0f
 #define PERFSCORE_LATENCY_23C 23.0f
 #define PERFSCORE_LATENCY_26C 26.0f
 #define PERFSCORE_LATENCY_62C 62.0f
@@ -1271,22 +1273,46 @@ class emitter
 #define PERFSCORE_LATENCY_BRANCH_COND 2.0f     // includes cost of a possible misprediction
 #define PERFSCORE_LATENCY_BRANCH_INDIRECT 2.0f // includes cost of a possible misprediction
 
+#if defined(_TARGET_XARCH_)
+
+// a read,write or modify from stack location, possible def to use latency from L0 cache
+#define PERFSCORE_LATENCY_RD_STACK PERFSCORE_LATENCY_2C
+#define PERFSCORE_LATENCY_WR_STACK PERFSCORE_LATENCY_2C
+#define PERFSCORE_LATENCY_RD_WR_STACK PERFSCORE_LATENCY_5C
+
+// a read, write or modify from constant location, possible def to use latency from L0 cache
+#define PERFSCORE_LATENCY_RD_CONST_ADDR PERFSCORE_LATENCY_2C
+#define PERFSCORE_LATENCY_WR_CONST_ADDR PERFSCORE_LATENCY_2C
+#define PERFSCORE_LATENCY_RD_WR_CONST_ADDR PERFSCORE_LATENCY_5C
+
+// a read, write or modify from memory location, possible def to use latency from L0 or L1 cache
+// plus an extra cost  (of 1.0) for a increased chance  of a cache miss
+#define PERFSCORE_LATENCY_RD_GENERAL PERFSCORE_LATENCY_3C
+#define PERFSCORE_LATENCY_WR_GENERAL PERFSCORE_LATENCY_3C
+#define PERFSCORE_LATENCY_RD_WR_GENERAL PERFSCORE_LATENCY_6C
+
+#elif defined(_TARGET_ARM64_) || defined(_TARGET_ARM_)
+
 // a read,write or modify from stack location, possible def to use latency from L0 cache
-#define PERFSCORE_LATENCY_RD_STACK 2.0f
-#define PERFSCORE_LATENCY_WR_STACK 2.0f
-#define PERFSCORE_LATENCY_RD_WR_STACK 5.0f
+#define PERFSCORE_LATENCY_RD_STACK PERFSCORE_LATENCY_3C
+#define PERFSCORE_LATENCY_WR_STACK PERFSCORE_LATENCY_1C
+#define PERFSCORE_LATENCY_RD_WR_STACK PERFSCORE_LATENCY_3C
 
 // a read, write or modify from constant location, possible def to use latency from L0 cache
-#define PERFSCORE_LATENCY_RD_CONST_ADDR 2.0f
-#define PERFSCORE_LATENCY_WR_CONST_ADDR 2.0f
-#define PERFSCORE_LATENCY_RD_WR_CONST_ADDR 5.0f
+#define PERFSCORE_LATENCY_RD_CONST_ADDR PERFSCORE_LATENCY_3C
+#define PERFSCORE_LATENCY_WR_CONST_ADDR PERFSCORE_LATENCY_1C
+#define PERFSCORE_LATENCY_RD_WR_CONST_ADDR PERFSCORE_LATENCY_3C
 
 // a read, write or modify from memory location, possible def to use latency from L0 or L1 cache
 // plus an extra cost  (of 1.0) for a increased chance  of a cache miss
-#define PERFSCORE_LATENCY_RD_GENERAL 3.0f
-#define PERFSCORE_LATENCY_WR_GENERAL 3.0f
-#define PERFSCORE_LATENCY_RD_WR_GENERAL 6.0f
+#define PERFSCORE_LATENCY_RD_GENERAL PERFSCORE_LATENCY_4C
+#define PERFSCORE_LATENCY_WR_GENERAL PERFSCORE_LATENCY_1C
+#define PERFSCORE_LATENCY_RD_WR_GENERAL PERFSCORE_LATENCY_4C
 
+#endif // _TARGET_XXX
+
+// Make this an enum:
+//
 #define PERFSCORE_MEMORY_NONE 0
 #define PERFSCORE_MEMORY_READ 1
 #define PERFSCORE_MEMORY_WRITE 2
@@ -1295,8 +1321,7 @@ class emitter
 #define PERFSCORE_CODESIZE_COST_HOT 0.10f
 #define PERFSCORE_CODESIZE_COST_COLD 0.01f
 
-#define PERFSCORE_CALLEE_SPILL_COST                                                                                    \
-    0.75f // heuristicly derived - actual cost is one push and one pop, in the prolog/epilog
+#define PERFSCORE_CALLEE_SPILL_COST 0.75f
 
     struct insExecutionCharacteristics
     {
@@ -1305,9 +1330,12 @@ class emitter
         unsigned insMemoryAccessKind;
     };
 
-    insExecutionCharacteristics getInsExecutionCharacteristics(instrDesc* id);
     float insEvaluateExecutionCost(instrDesc* id);
 
+    insExecutionCharacteristics getInsExecutionCharacteristics(instrDesc* id);
+
+    void emitter::perfScoreUnhandledInstruction(instrDesc* id, insExecutionCharacteristics* result);
+
 #endif // defined(DEBUG) || defined(LATE_DISASM)
 
     BasicBlock::weight_t getCurrentBlockWeight();

src/coreclr/src/jit/emitarm.cpp

@@ -7991,8 +7991,8 @@ emitter::insExecutionCharacteristics emitter::getInsExecutionCharacteristics(ins
 
     // ToDo: Calculate actual throughput and latency values
     //
-    result.insThroughput = PERFSCORE_THROUGHPUT_DEFAULT;
-    result.insLatency    = PERFSCORE_LATENCY_DEFAULT;
+    result.insThroughput = PERFSCORE_THROUGHPUT_1C;
+    result.insLatency    = PERFSCORE_LATENCY_1C;
 
     return result;
 }