Mark loads/stores for private variables

Mark loads/stores which are created for fill/spill of user private variables. Adding this mark, because during compilation the orginal addrspace is changed (for ex. from PRIVATE to GLOBAL) and is not visible on end stages of compilation. This will help to identify - which load/store is related for the private variables of user. In llvm-IR load/stores it has mark "user_addrspace_priv". In asm dumps it has mark "address space: private;"
intel · Sep 5, 2023 · 427f3d3 · 427f3d3
1 parent 5a8e871
commit 427f3d3
Show file tree

Hide file tree

Showing 3 changed files with 157 additions and 1 deletion.
diff --git a/IGC/Compiler/CISACodeGen/LowerGEPForPrivMem.cpp b/IGC/Compiler/CISACodeGen/LowerGEPForPrivMem.cpp
@@ -73,6 +73,7 @@ namespace IGC {
         StatusPrivArr2Reg CheckIfAllocaPromotable(llvm::AllocaInst* pAlloca);
         bool IsNativeType(Type* type);
 
+        void MarkNotPromtedAllocas(llvm::AllocaInst& I, IGC::StatusPrivArr2Reg status);
     public:
         static char ID;
 
@@ -491,6 +492,90 @@ bool IGC::CanUseSOALayout(AllocaInst* I, Type*& base, bool& allUsesAreVector)
     return useSOA;
 }
 
+void LowerGEPForPrivMem::MarkNotPromtedAllocas(llvm::AllocaInst& I, IGC::StatusPrivArr2Reg status)
+{
+    const char* reason = nullptr;
+    // The reason why the user private variable
+    // wasn't promoted to grfs
+    switch (status)
+    {
+    case StatusPrivArr2Reg::CannotUseSOALayout:
+        reason = "CannotUseSOALayout";
+        break;
+    case StatusPrivArr2Reg::IsDynamicAlloca:
+        reason = "IsDynamicAlloca";
+        break;
+    case StatusPrivArr2Reg::IsNotNativeType:
+        reason = "IsNotNativeType";
+        break;
+    case StatusPrivArr2Reg::OutOfAllocSizeLimit:
+        reason = "OutOfAllocSizeLimit";
+        break;
+    case StatusPrivArr2Reg::OutOfMaxGRFPressure:
+        reason = "OutOfMaxGRFPressure";
+        break;
+    default:
+        reason = "NotDefine";
+        break;
+    }
+    MDNode* node = MDNode::get(
+        I.getContext(),
+        MDString::get(I.getContext(), reason));
+
+    UserAddrSpaceMD& userASMD = m_ctx->m_UserAddrSpaceMD;
+    std::function<void(Instruction*, MDNode*)> markAS_PRIV;
+    markAS_PRIV = [&markAS_PRIV, &userASMD](Instruction* instr, MDNode* node) -> void
+    {
+        // Avoid instruction which has already md set
+        if (!userASMD.Has(instr, LSC_DOC_ADDR_SPACE::PRIVATE))
+        {
+            // Adding this mark because, during compilation the orginal
+            // addrspace is changed (for ex. from PRIVATE to GLOBAL) and
+            // is not visible on end stages of compilation. This will help
+            // to identify - which load/store is related for the private
+            // variables of user.
+
+            bool isLoadStore =
+                instr->getOpcode() == Instruction::Store ||
+                instr->getOpcode() == Instruction::Load;
+
+            if (isLoadStore)
+            {
+                // Add mark for any load/store which will read/write the data from
+                // user private variable. This information will be passed
+                // to the assembly level.
+                userASMD.Set(instr, LSC_DOC_ADDR_SPACE::PRIVATE, node);
+            }
+            else
+            {
+                // Special case to avoid stack overflow
+                userASMD.Set(instr, LSC_DOC_ADDR_SPACE::PRIVATE);
+
+                bool allowedInst =
+                    instr->getOpcode() == Instruction::Alloca ||
+                    instr->getOpcode() == Instruction::PHI ||
+                    instr->getOpcode() == Instruction::GetElementPtr ||
+                    instr->getOpcode() == Instruction::PtrToInt ||
+                    instr->getOpcode() == Instruction::IntToPtr ||
+                    instr->isBinaryOp();
+
+
+                if (allowedInst)
+                {
+                    for (auto user_i = instr->user_begin();
+                        user_i != instr->user_end();
+                        ++user_i)
+                    {
+                        markAS_PRIV(llvm::dyn_cast<Instruction>(*user_i), node);
+                    }
+                }
+            }
+        }
+    };
+
+    markAS_PRIV(&I, node);
+}
+
 void LowerGEPForPrivMem::visitAllocaInst(AllocaInst& I)
 {
     // Alloca should always be private memory
@@ -504,6 +589,7 @@ void LowerGEPForPrivMem::visitAllocaInst(AllocaInst& I)
     }
     if (status != StatusPrivArr2Reg::OK)
     {
+        MarkNotPromtedAllocas(I, status);
         // alloca size extends remain per-lane-reg space
         return;
     }

diff --git a/IGC/Compiler/DebugInfo/ScalarVISAModule.cpp b/IGC/Compiler/DebugInfo/ScalarVISAModule.cpp
@@ -376,7 +376,6 @@ ScalarVisaModule::GetVariableLocation(const llvm::Instruction* pInst) const
     {
         if (!isa<GlobalVariable>(pVal) && !isa<ConstantExpr>(pVal))
         {
-            IGC_ASSERT_MESSAGE(!isDbgDclInst, "address cannot be immediate!");
             return VISAVariableLocation(pConstVal, this);
         }
     }

diff --git a/IGC/ocloc_tests/features/metadata_travel_check/user_private_var.cl b/IGC/ocloc_tests/features/metadata_travel_check/user_private_var.cl
@@ -0,0 +1,71 @@
+/*========================== begin_copyright_notice ============================
+
+Copyright (C) 2023 Intel Corporation
+
+SPDX-License-Identifier: MIT
+
+============================= end_copyright_notice ===========================*/
+// REQUIRES: regkeys
+// windows unsupported due to issues on 32bit build, to be debugged.
+// UNSUPPORTED: system-windows
+
+// checking the asm dump file
+// RUN: ocloc compile -file %s -options " -g -igc_opts 'VISAOptions=-asmToConsole'" -device pvc 2>&1 | FileCheck %s --check-prefix=CHECK-ASM
+// checking the llvm-IR after EmitVISAPass
+// RUN: ocloc compile -file %s -options " -g -igc_opts 'PrintToConsole=1 PrintMDBeforeModule=1 PrintAfter=EmitPass'" -device pvc 2>&1 | FileCheck %s --check-prefix=CHECK-LLVM
+// RUN: ocloc compile -file %s -options " -g -igc_opts 'PrintToConsole=1 PrintMDBeforeModule=1 PrintAfter=EmitPass'" -device pvc 2>&1 | FileCheck %s --check-prefix=CHECK-LLVM2
+// RUN: ocloc compile -file %s -options " -g -igc_opts 'PrintToConsole=1 PrintMDBeforeModule=1 PrintAfter=EmitPass'" -device pvc 2>&1 | FileCheck %s --check-prefix=CHECK-LLVM3
+// RUN: ocloc compile -file %s -options " -g -igc_opts 'PrintToConsole=1 PrintMDBeforeModule=1 PrintAfter=EmitPass'" -device pvc 2>&1 | FileCheck %s --check-prefix=CHECK-LLVM4
+
+// Looking for the comment attached to the instructions ex:
+// (W)     store.ugm .... //  address space: private; ; $193
+// CHECK-ASM: address space: private;
+
+// Looking for the MD attached to the instructions ex:
+// store <4 x i32> %39, <4 x i32>* %40, align 4, !user_as_priv !348   ; visa id: 82
+// %55 = load i32, i32* %54, align 4, !user_as_priv !348   ; visa id: 127
+// Looking for the reason of not promoting the user variable to grf
+
+
+kernel void test_simple(global int* output)
+{
+  // To force spill of the variable private_array.
+  // This will generate the metadata !user_as_priv in LLVM-IR
+  // and comment in asm dump "address space: private;"
+  int private_array[1024];
+  int private_array_2[1024];
+
+  for(int i = 0; i < 1024; ++i)
+  {
+// Look first for the debugInfo mark which indetify the asign place
+// CHECK-LLVM: [[PA1:![0-9]+]] = !DILocation(line: [[@LINE+7]], column: 22,
+// Look if we have any load/store pointing to the dbg loc which don't have
+// md attribute !user_as_priv
+// regex in POSIX: !([^u]|u[^s]|us[^e]|use[^r]|user[^_]|user_[^a]|user_a[^s]|user_as[^_]|user_as_[^p])
+// equals negative lookahead !(?!user_as_p)
+// CHECK-LLVM-NOT: {{^ store .*, !dbg }}[[PA1]]{{, !([^u]|u[^s]|us[^e]|use[^r]|user[^_]|user_[^a]|user_a[^s]|user_as[^_]|user_as_[^p])}}
+// CHECK-LLVM-NOT: {{^ .* = load .*, !dbg }}[[PA1]]{{, !([^u]|u[^s]|us[^e]|use[^r]|user[^_]|user_[^a]|user_a[^s]|user_as[^_]|user_as_[^p])}}
+    private_array[i] = i;
+  }
+
+  for(int i = 1023; i >= 0 ; --i)
+  {
+// CHECK-LLVM2: [[PA2:![0-9]+]] = !DILocation(line: [[@LINE+3]], column: 21,
+// CHECK-LLVM2-NOT: {{^ store .*, !dbg }}[[PA2]]{{, !([^u]|u[^s]|us[^e]|use[^r]|user[^_]|user_[^a]|user_a[^s]|user_as[^_]|user_as_[^p])}}
+// CHECK-LLVM2-NOT: {{^ .* = load .*, !dbg }}[[PA2]]{{, !([^u]|u[^s]|us[^e]|use[^r]|user[^_]|user_[^a]|user_a[^s]|user_as[^_]|user_as_[^p])}}
+    const int tmp = private_array[i];
+// CHECK-LLVM3: [[PA3:![0-9]+]] = !DILocation(line: [[@LINE+3]], column: 24,
+// CHECK-LLVM3-NOT: {{^ store .*, !dbg }}[[PA3]]{{, !([^u]|u[^s]|us[^e]|use[^r]|user[^_]|user_[^a]|user_a[^s]|user_as[^_]|user_as_[^p])}}
+// CHECK-LLVM3-NOT: {{^ .* = load .*, !dbg }}[[PA3]]{{, !([^u]|u[^s]|us[^e]|use[^r]|user[^_]|user_[^a]|user_a[^s]|user_as[^_]|user_as_[^p])}}
+    private_array_2[i] = tmp;
+  }
+
+  for(int i = 0; i < 1024; ++i)
+  {
+// CHECK-LLVM4: [[PA4:![0-9]+]] = !DILocation(line: [[@LINE+3]], column: 21,
+// CHECK-LLVM4-NOT: {{^ store .*, !dbg }}[[PA4]]{{, !([^u]|u[^s]|us[^e]|use[^r]|user[^_]|user_[^a]|user_a[^s]|user_as[^_]|user_as_[^p])}}
+// CHECK-LLVM4-NOT: {{^ .* = load .*, !dbg }}[[PA4]]{{, !([^u]|u[^s]|us[^e]|use[^r]|user[^_]|user_[^a]|user_a[^s]|user_as[^_]|user_as_[^p])}}
+    const int tmp = private_array_2[i];
+    output[i] = tmp;
+  }
+}