dedicate load store emitters runtime info for segfault

src/plugins/intel_cpu/src/emitters/x64/jit_load_store_emitters.cpp

@@ -97,6 +97,24 @@ size_t jit_load_emitter::aux_gprs_count() const {
 }
 
 void jit_load_emitter::emit_impl(const std::vector<size_t> &in_idxs, const std::vector<size_t> &out_idxs) const {
+    // save runtime debug info
+    h->push(h->r15);
+    Xbyak::Label label_set_current;
+    h->mov(h->r15, reinterpret_cast<size_t>(&start_address));
+    h->cmp(h->qword[h->r15], 0);
+    h->jne(label_set_current);
+    h->mov(h->qword[h->r15], Xbyak::Reg64(in_idxs[0]));
+    h->L(label_set_current);
+    {
+        h->mov(h->r15, reinterpret_cast<size_t>(&current_address));
+        h->mov(h->qword[h->r15], Xbyak::Reg64(in_idxs[0]));
+
+        // iteration++
+        h->mov(h->r15, reinterpret_cast<size_t>(&iteration));
+        h->add(h->qword[h->r15], 0x01);
+    }
+    h->pop(h->r15);
+
     const int offset = in_idxs.size() == 2 ? in_idxs[1] : 0;
     if (host_isa_ == cpu::x64::sse41) {
         emit_isa<cpu::x64::sse41>(Reg64(in_idxs[0]), static_cast<int>(out_idxs[0]), offset);
@@ -601,6 +619,7 @@ void jit_load_emitter::register_table_entries() {
 
 void jit_load_emitter::print_debug_info() const {
     std::cerr << "ERROR is from jit_load_emitter." << "\n";
+    std::cerr << "where start_address:" << start_address << " current_address:" << current_address << " iteration:" << iteration << "\n";
     std::cerr << "Emitter name:" << name_ << "\n";
     std::cerr << "load_num_:" << load_num_ << "\n";
     std::cerr << "src_prc_:" << src_prc_ << "\n";
@@ -667,6 +686,24 @@ void jit_store_emitter::emit_data() const {
 }
 
 void jit_store_emitter::emit_impl(const std::vector<size_t> &in_idxs, const std::vector<size_t> &out_idxs) const {
+    // save runtime debug info
+    h->push(h->r15);
+    Xbyak::Label label_set_current;
+    h->mov(h->r15, reinterpret_cast<size_t>(&start_address));
+    h->cmp(h->qword[h->r15], 0);
+    h->jne(label_set_current);
+    h->mov(h->qword[h->r15], Xbyak::Reg64(out_idxs[0]));
+    h->L(label_set_current);
+    {
+        h->mov(h->r15, reinterpret_cast<size_t>(&current_address));
+        h->mov(h->qword[h->r15], Xbyak::Reg64(out_idxs[0]));
+
+        // iteration++
+        h->mov(h->r15, reinterpret_cast<size_t>(&iteration));
+        h->add(h->qword[h->r15], 0x01);
+    }
+    h->pop(h->r15);
+
     const int offset = in_idxs.size() == 2 ? in_idxs[1] : 0;
     if (host_isa_ == cpu::x64::sse41) {
         emit_isa<cpu::x64::sse41>(static_cast<int>(in_idxs[0]), Reg64(out_idxs[0]), offset);
@@ -1273,6 +1310,7 @@ void jit_store_emitter::register_table_entries() {
 
 void jit_store_emitter::print_debug_info() const {
     std::cerr << "Segfault happens in jit_store_emitter." << "\n";
+    std::cerr << "where start_address:" << start_address << " current_address:" << current_address << " iteration:" << iteration << "\n";
     std::cerr << "Emitter name:" << name_ << "\n";
     std::cerr << "store_num_:" << store_num_ << "\n";
     std::cerr << "src_prc_:" << src_prc_ << "\n";

src/plugins/intel_cpu/src/emitters/x64/jit_load_store_emitters.hpp

@@ -101,6 +101,11 @@ class jit_load_emitter : public jit_emitter {
     InferenceEngine::Precision dst_prc_;
     bool is_fill_;
     std::string fill_value_;
+
+    size_t start_address = 0;
+    size_t current_address = 0;
+    size_t iteration = 0;
+    std::shared_ptr<snippets::op::Store> m_load_node;
 };
 
 class jit_store_emitter : public jit_emitter {
@@ -174,6 +179,11 @@ class jit_store_emitter : public jit_emitter {
     mutable bool data_reg_updated = false;
     mutable int data_idx = 0;
     mutable int aux_src_idx = 0;
+
+    size_t start_address = 0;
+    size_t current_address = 0;
+    size_t iteration = 0;
+    std::shared_ptr<snippets::op::Store> m_store_node;
 };
 
 }   // namespace intel_cpu

src/plugins/intel_cpu/src/emitters/x64/jit_snippets_emitters.hpp

@@ -248,9 +248,9 @@ class ScalarEmitter : public jit_emitter {
 /// Memory emitters:
 ///
 /// *Note*: post increment is embedded into Load/Store operation which means that
-/// it's illigal to load/store to the same address multiple times
+/// it's illegal to load/store to the same address multiple times
 /// Typical application can be if Load and BroadcastLoad are performed from the same pointer.
-/// If Load goes before BroadcastLoad topologicaly the resilt will be incorrect
+/// If Load goes before BroadcastLoad topologicaly the result will be incorrect
 /// For scalar loads we can use different loops. Tiling indeed can be arbitrary and post increment should be somehow coded into ISA.
 /// Blocked parameter to tell if input is actually blocked. Broadcast means broadcast by W in other cases no need to substitute load.
 class MemoryEmitter : public jit_emitter  {

src/plugins/intel_cpu/src/nodes/subgraph.cpp

@@ -466,16 +466,15 @@ void Snippet::SnippetJitExecutor::update_ptrs(jit_snippets_call_args& call_args,
 void Snippet::SnippetJitExecutor::schedule_6d(const std::vector<MemoryPtr>& inMemPtrs, const std::vector<MemoryPtr>& outMemPtrs) {
     const auto& dom = parallel_exec_domain;
     // < N, C, H, W > < 1, 1, N, C*H*W>
-    // parallel_for5d(dom[0], dom[1], dom[2], dom[3], dom[4],
-    parallel_for5d(1, 1, 1, 1, 1,
+    parallel_for5d(dom[0], dom[1], dom[2], dom[3], dom[4],
         [&](int64_t d0, int64_t d1, int64_t d2, int64_t d3, int64_t d4) {
             int64_t indexes[] = {d0, d1, d2, d3, d4};
             jit_snippets_call_args call_args;
             update_ptrs(call_args, inMemPtrs, outMemPtrs);
-#ifndef _WIN32
+#ifdef __linux__
             __sighandler_t signal_handler = [](int signal) {
-                std::cerr << "Segfault was caught by the signal handler.\n";
                 ov::intel_cpu::g_debug_err_handler->print_debug_info();
+                OPENVINO_THROW("Segfault was caught by the signal handler");
             };
             struct sigaction new_handler{};
             new_handler.sa_handler = signal_handler;
@@ -501,10 +500,10 @@ void Snippet::SnippetJitExecutor::schedule_nt(const std::vector<MemoryPtr>& inMe
                 indexes[j] = tmp % work_size[j];
                 tmp /= work_size[j];
             }
-#ifndef _WIN32
+#ifdef __linux__
             __sighandler_t signal_handler = [](int signal) {
-                std::cerr << "Segfault was caught by the signal handler.\n";
                 ov::intel_cpu::g_debug_err_handler->print_debug_info();
+                OPENVINO_THROW("Segfault was caught by the signal handler");
             };
             struct sigaction new_handler{};
             new_handler.sa_handler = signal_handler;