refactor(SBuffer, StoreDifftest): move diff store

event calculate part from scala to cpp

BREAKING CHANGE: struct DifftestStoreEvent is
changed and broke backward compatibility!

Author: skyhgzsh

src/main/scala/Bundles.scala

@@ -245,16 +245,25 @@ class UncacheMMStoreEvent extends DifftestBaseBundle with HasValid {
 }
 
 class StoreEvent extends DifftestBaseBundle with HasValid {
-  val addr = UInt(64.W)
-  val data = UInt(64.W)
-  val highData = UInt(64.W)
-  val mask = UInt(16.W)
   val wLine = Bool()
-  val vecNeedSplit = Bool()
-  val eew = UInt(8.W)
   val offset = UInt(16.W)
   val pc = UInt(64.W)
   val robidx = UInt(10.W)
+
+  val isNonCacheable = Bool()
+  val isVStore = Bool()
+  val isUStride = Bool()
+  val isMasked = Bool()
+  val isWhole = Bool()
+  val veew = UInt(8.W)
+  val nf = UInt(8.W)
+  val rawDataLow = UInt(64.W)
+  val rawDataHigh = UInt(64.W)
+  val rawMask = UInt(16.W)
+  val rawAddr = UInt(64.W)
+  val wline = Bool()
+  val vecValid = Bool()
+  val isHighPart = Bool()
 }
 
 class LoadEvent extends DifftestBaseBundle with HasValid {

src/test/csrc/difftest/checkers/store.cpp

@@ -27,6 +27,13 @@ void StoreRecorder::clear_valid(DifftestStoreEvent &probe) {
   probe.valid = 0;
 }
 
+// Expand 8-bit mask to bit-and with 64-bit wide data
+static uint64_t MaskExpand(uint8_t mask) {
+  uint64_t expander = 0x01010101'01010101ULL;
+  uint64_t selector = 0x80402010'08040201ULL;
+  return (((((mask * expander) & selector) * 0xFFULL) >> 7) & expander) * 0xFFULL;
+}
+
 int StoreRecorder::check(const DifftestStoreEvent &probe) {
   if (!probe.valid)
     return STATE_OK;
@@ -35,18 +42,91 @@ int StoreRecorder::check(const DifftestStoreEvent &probe) {
   int WORDBYTES = 8;
   int COMMITBYTES = 16;
 
-  auto addr = probe.addr;
-  auto lowData = probe.data;
-  auto highData = probe.highData;
-  auto mask = probe.mask;
+  uint64_t calcAddr;
+  uint64_t calcDataLow;
+  uint64_t calcDataHigh;
+  uint16_t calcMask;
+  uint8_t calcEEW;
+  bool calcIsVSLine;
+
+  auto isNonCacheable = probe.isNonCacheable;
+  auto isVStore = probe.isVStore;
+  auto isUStride = probe.isUStride;
+  auto isMasked = probe.isMasked;
+  auto isWhole = probe.isWhole;
+  auto veew = probe.veew;
+  auto nf = probe.nf;
+  auto rawDataLow = probe.rawDataLow;
+  auto rawDataHigh = probe.rawDataHigh;
+  auto rawMask = probe.rawMask;
+  auto rawAddr = probe.rawAddr;
+  auto wline = probe.wline;
+  auto vecValid = probe.vecValid;
+  auto isHighPart = probe.isHighPart;
+
+  if (!isNonCacheable) {
+    bool isVse = isVStore && isUStride;
+    bool isVsm = isVStore && isMasked;
+    bool isVsr = isVStore && isWhole;
+
+    uint8_t eew = veew;
+    uint32_t EEB = 1 << eew;
+    uint32_t EEWBits = EEB << 3;
+    uint8_t nf2 = isVsr ? nf : 0;
+
+    bool isSegment = nf != 0 && !isVsm;
+    bool isVSLine = (isVse || isVsm || isVsr) && !isSegment;
+    bool isWline = wline;
+    calcIsVSLine = isVSLine;
+
+    uint8_t flow = isVSLine ? (16 >> eew) : 0;
+
+    if (isVSLine) {
+      calcAddr = rawAddr;
+      calcDataLow = rawDataLow;
+      calcDataHigh = rawDataHigh;
+      calcMask = rawMask;
+    } else if (isWline) {
+      calcAddr = rawAddr;
+      calcDataLow = rawDataLow;
+      calcDataHigh = rawDataHigh;
+      calcMask = rawMask;
+      Assert(rawDataLow == 0 && rawDataHigh == 0, "wline only supports whole zero write now");
+    } else if (isHighPart) {
+      calcAddr = (rawAddr & (~0xFULL)) | 0b1000;
+      calcMask = (rawMask >> 8) & 0xFF;
+      uint64_t tmpData = rawDataHigh;
+      calcDataLow = tmpData & MaskExpand(calcMask);
+      calcDataHigh = 0;
+    } else {
+      calcAddr = rawAddr & (~0xFULL);
+      calcMask = rawMask & 0xFF;
+      uint64_t tmpData = rawDataLow;
+      calcDataLow = tmpData & MaskExpand(calcMask);
+      calcDataHigh = 0;
+    }
+    calcEEW = EEB;
+  } else {
+    calcAddr = rawAddr & (~7ULL);
+    calcMask = rawMask;
+    calcDataLow = rawDataLow & MaskExpand(calcMask);
+    calcDataHigh = 0;
+    calcEEW = 0;
+    calcIsVSLine = false;
+  }
+
+  auto addr = calcAddr;
+  auto lowData = calcDataLow;
+  auto highData = calcDataHigh;
+  auto mask = calcMask;
   auto offset = probe.offset;
-  auto eew = probe.eew;
+  auto eew = calcEEW;
   auto pc = probe.pc;
   auto robIdx = probe.robidx;
 
   uint64_t rawVecAddr = addr + offset;
 
-  if (probe.vecNeedSplit) {
+  if (calcIsVSLine) {
     uint16_t flow = COMMITBYTES / eew;
     uint64_t flowMask = (eew == 1) ? 0x1ULL : (eew == 2) ? 0x3ULL : (eew == 4) ? 0xfULL : (eew == 8) ? 0xffULL : 0x0ULL;
     uint64_t flowMaskBit = (eew == 1)   ? 0xffULL
@@ -103,7 +183,7 @@ int StoreRecorder::check(const DifftestStoreEvent &probe) {
 
         nextOffset = 8 - rawOffset;
         auto nextDataOffset = nextOffset * 8;
-        auto nextData = probe.highData << nextDataOffset;
+        auto nextData = calcDataHigh << nextDataOffset;
 
         refStoreCommitData = (nextData + presentData) & flowMaskBit;
       } else {
@@ -145,9 +225,9 @@ int StoreRecorder::check(const DifftestStoreEvent &probe) {
     }
   } else {
     DiffState::StoreCommit storeCommit = {probe.valid,
-                                          probe.addr,
+                                          calcAddr,
                                           lowData,
-                                          static_cast<uint8_t>(probe.mask & 0xFF),
+                                          static_cast<uint8_t>(calcMask & 0xFF),
                                           pc,
                                           robIdx
 #ifdef CONFIG_DIFFTEST_SQUASH