Debug info emission fix when scratch offset is encoded

in LSC message

Teach VISA debug info emission to emit scratch offset when it's directly
embedded in LSC message.

Author: pratikashar

visa/DebugInfo.cpp

@@ -1991,10 +1991,12 @@ void SaveRestoreManager::addInst(G4_INST *inst) {
         srInfo[srInfo.size() - 2].saveRestoreMap;
   }
 
+  auto *kernel = visaKernel->getKernel();
+
   if (inst->opcode() == G4_add && inst->getSrc(1) && inst->getSrc(1)->isImm() &&
       inst->getSrc(0) && inst->getSrc(0)->isSrcRegRegion() &&
       GetTopDclFromRegRegion(inst->getSrc(0)) ==
-          visaKernel->getKernel()->fg.builder->getBEFP()) {
+          kernel->fg.builder->getBEFP()) {
     memOffset = (int32_t)inst->getSrc(1)->asImm()->getImm();
     regWithMemOffset =
         inst->getDst()->getLinearizedStart() / builder.numEltPerGRF<Type_UB>();
@@ -2011,6 +2013,29 @@ void SaveRestoreManager::addInst(G4_INST *inst) {
     absOffset = true;
   }
 
+  if (inst->isSend()) {
+    auto *sendInst = inst->asSendInst();
+    if (GlobalRA::LSCUsesImmOff(*kernel->fg.builder) &&
+        sendInst->getMsgDesc()->isLSC() && sendInst->getMsgDescRaw() &&
+        sendInst->getMsgDescRaw()->getLscAddrType() == LSC_ADDR_TYPE_SS &&
+        (inst->getSrc(0)->getLinearizedStart() /
+         builder.numEltPerGRF<Type_UB>()) ==
+            kernel->stackCall.getSpillHeaderGRF()) {
+      // Spill offset is inlined in LSC message as:
+      // GlobalRA::SPILL_FILL_IMMOFF_MAX + Inlined offset in LSC
+      auto maxSpillFillImmOffMax = GlobalRA::SPILL_FILL_IMMOFF_MAX;
+
+      auto msgDesc = inst->asSendInst()->getMsgDesc();
+      auto off = msgDesc->getOffset();
+      if (off) {
+        memOffset = (maxSpillFillImmOffMax + off.value().immOff);
+        absOffset = false;
+        regWithMemOffset = inst->getSrc(0)->getLinearizedStart() /
+                           builder.numEltPerGRF<Type_UB>();
+      }
+    }
+  }
+
   srInfo.back().update(inst, memOffset, regWithMemOffset, absOffset);
 }
 

visa/GraphColor.cpp

@@ -10954,17 +10954,7 @@ int GlobalRA::coloringRegAlloc() {
   // b. Spill size exceeds what can be represented using hword msg on PVC+
   // c. Xe2+ requires LSC stack (can force on DG2+ via -lscNonStackSpill)
   if (builder.supportsLSC()) {
-
-    const auto scratchAddrType = VISA_LSC_IMMOFF_ADDR_TYPE_SS;
-    const uint32_t immOffOpts =
-        builder.getuint32Option(vISA_lscEnableImmOffsFor);
-    canUseLscImmediateOffsetSpillFill =
-        // HW supports it
-        builder.getPlatform() >= Xe2 &&
-        // the spill/fill is enabled in options
-        (immOffOpts & (1 << VISA_LSC_IMMOFF_SPILL_FILL)) != 0 &&
-        // address type is also enabled in options
-        (immOffOpts & (1 << scratchAddrType)) != 0;
+    canUseLscImmediateOffsetSpillFill = LSCUsesImmOff(builder);
   }
 
   stackCallSaveRestore(hasStackCall);

visa/GraphColor.h

@@ -1348,6 +1348,23 @@ class GlobalRA {
 
   const bool use4GRFAlign = false;
 
+  // [-2^16...2^16) in bytes
+  //   (frame pointer is biased 2^16 so that -2^16 references scratch[0x0])
+  static constexpr unsigned SPILL_FILL_IMMOFF_MAX = 0x10000; // 64k
+
+  static bool LSCUsesImmOff(IR_Builder &builder) {
+    const auto scratchAddrType = VISA_LSC_IMMOFF_ADDR_TYPE_SS;
+    const uint32_t immOffOpts =
+        builder.getuint32Option(vISA_lscEnableImmOffsFor);
+    return
+        // HW supports it
+        builder.getPlatform() >= Xe2 &&
+        // the spill/fill is enabled in options
+        (immOffOpts & (1 << VISA_LSC_IMMOFF_SPILL_FILL)) != 0 &&
+        // address type is also enabled in options
+        (immOffOpts & (1 << scratchAddrType)) != 0;
+  }
+
 private:
   template <class REGION_TYPE>
   static unsigned getRegionDisp(REGION_TYPE *region, const IR_Builder &irb);
@@ -1392,10 +1409,6 @@ class GlobalRA {
   // store instructions that shouldnt be rematerialized.
   std::unordered_set<G4_INST *> dontRemat;
 
-  // [-2^16...2^16) in bytes
-  //   (frame pointer is biased 2^16 so that -2^16 references scratch[0x0])
-  static constexpr unsigned SPILL_FILL_IMMOFF_MAX = 0x10000; // 64k
-
   // map each BB to its local RA GRF usage summary, populated in local RA.
   std::map<G4_BB *, PhyRegSummary *> bbLocalRAMap;
   llvm::SpecificBumpPtrAllocator<PhyRegSummary> PRSAlloc;