Create Patches for Function Pointer Symbol and Relocation Tables

dlei6g · gfxbot · commit 81d467882a26 · 2019-03-21T05:53:58.000-07:00
Change-Id: I29ce6ed33d8f74caef9b818e452ff770419e9647
diff --git a/IGC/AdaptorOCL/OCL/Patch/patch_parser.cpp b/IGC/AdaptorOCL/OCL/Patch/patch_parser.cpp
@@ -1414,29 +1414,62 @@ void DebugPatchList(
             break;
 
         case iOpenCL::PATCH_TOKEN_GTPIN_FREE_GRF_INFO:
-        {
-            const iOpenCL::SPatchGtpinFreeGRFInfo* pPatchItem =
-                (const iOpenCL::SPatchGtpinFreeGRFInfo*)pHeader;
+            {
+                const iOpenCL::SPatchGtpinFreeGRFInfo* pPatchItem =
+                    (const iOpenCL::SPatchGtpinFreeGRFInfo*)pHeader;
 
-            ICBE_DPF_STR(output, GFXDBG_HARDWARE,
-                "PATCH_TOKEN_GTPIN_FREE_GRF_INFO (%08X) (size = %d)\n",
-                pPatchItem->Token,
-                pPatchItem->Size);
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                    "PATCH_TOKEN_GTPIN_FREE_GRF_INFO (%08X) (size = %d)\n",
+                    pPatchItem->Token,
+                    pPatchItem->Size);
 
-            ICBE_DPF_STR(output, GFXDBG_HARDWARE,
-                "\tBufferSize = %d\n",
-                pPatchItem->BufferSize);
-        }
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                    "\tBufferSize = %d\n",
+                    pPatchItem->BufferSize);
+            }
+            break;
         case iOpenCL::PATCH_TOKEN_GTPIN_INFO:
-        {
-            const iOpenCL::SPatchItemHeader* pPatchItem = pHeader;
-            
-            ICBE_DPF_STR(output, GFXDBG_HARDWARE,
-                         "PATCH_TOKEN_GTPIN_INFO (%08X) (size = %d)\n",
-                         pPatchItem->Token,
-                         pPatchItem->Size);
-        }
-        break;
+            {
+                const iOpenCL::SPatchItemHeader* pPatchItem = pHeader;
+
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                             "PATCH_TOKEN_GTPIN_INFO (%08X) (size = %d)\n",
+                             pPatchItem->Token,
+                             pPatchItem->Size);
+            }
+            break;
+        case iOpenCL::PATCH_TOKEN_FUNCTION_SYMBOL_TABLE:
+            {
+                const iOpenCL::SPatchFunctionTableInfo* pPatchItem =
+                    (const iOpenCL::SPatchFunctionTableInfo*)pHeader;
+
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                             "PATCH_TOKEN_FUNCTION_SYMBOL_TABLE (%08X) (size = %d)\n",
+                             pPatchItem->Token,
+                             pPatchItem->Size);
+
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                             "\tNumEntries = %d\n",
+                             pPatchItem->NumEntries);
+
+            }
+            break;
+        case iOpenCL::PATCH_TOKEN_FUNCTION_RELOCATION_TABLE:
+            {
+                const iOpenCL::SPatchFunctionTableInfo* pPatchItem =
+                    (const iOpenCL::SPatchFunctionTableInfo*)pHeader;
+
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                             "PATCH_TOKEN_FUNCTION_RELOCATION_TABLE (%08X) (size = %d)\n",
+                             pPatchItem->Token,
+                             pPatchItem->Size);
+
+                ICBE_DPF_STR(output, GFXDBG_HARDWARE,
+                             "\tNumEntries = %d\n",
+                             pPatchItem->NumEntries);
+
+            }
+            break;
 
         default:
             {
diff --git a/IGC/AdaptorOCL/OCL/sp/sp_g8.cpp b/IGC/AdaptorOCL/OCL/sp/sp_g8.cpp
@@ -2104,6 +2104,98 @@ RETVAL CGen8OpenCLStateProcessor::CreatePatchList(
         }
     }
 
+    // Patch for function symbol table
+    if (retValue.Success)
+    {
+        iOpenCL::SPatchFunctionTableInfo patch;
+        memset(&patch, 0, sizeof(patch));
+
+        patch.Token = PATCH_TOKEN_FUNCTION_SYMBOL_TABLE;
+        uint32_t size = 0;
+        uint32_t entries = 0;
+        void* buffer = nullptr;
+        const IGC::SKernelProgram* program = &(annotations.m_kernelProgram);
+        if (annotations.m_executionEnivronment.CompiledSIMDSize == 8)
+        {
+            buffer = program->simd8.m_funcSymbolTable;
+            size = program->simd8.m_funcSymbolTableSize;
+            entries = program->simd8.m_funcSymbolTableEntries;
+        }
+        else if (annotations.m_executionEnivronment.CompiledSIMDSize == 16)
+        {
+            buffer = program->simd16.m_funcSymbolTable;
+            size = program->simd16.m_funcSymbolTableSize;
+            entries = program->simd16.m_funcSymbolTableEntries;
+        }
+        else if (annotations.m_executionEnivronment.CompiledSIMDSize == 32)
+        {
+            buffer = program->simd32.m_funcSymbolTable;
+            size = program->simd32.m_funcSymbolTableSize;
+            entries = program->simd32.m_funcSymbolTableEntries;
+        }
+
+        if (size > 0)
+        {
+            patch.Size = sizeof(patch) + size;
+            patch.NumEntries = entries;
+
+            retValue = AddPatchItem(patch, membuf);
+
+            if (!membuf.Write((const char*)buffer, size))
+            {
+                retValue.Success = false;
+                return retValue;
+            }
+            free(buffer);
+        }
+    }
+
+    // Patch for function relocation table
+    if (retValue.Success)
+    {
+        iOpenCL::SPatchFunctionTableInfo patch;
+        memset(&patch, 0, sizeof(patch));
+
+        patch.Token = PATCH_TOKEN_FUNCTION_RELOCATION_TABLE;
+        uint32_t size = 0;
+        uint32_t entries = 0;
+        void* buffer = nullptr;
+        const IGC::SKernelProgram* program = &(annotations.m_kernelProgram);
+        if (annotations.m_executionEnivronment.CompiledSIMDSize == 8)
+        {
+            buffer = program->simd8.m_funcRelocationTable;
+            size = program->simd8.m_funcRelocationTableSize;
+            entries = program->simd8.m_funcRelocationTableEntries;
+        }
+        else if (annotations.m_executionEnivronment.CompiledSIMDSize == 16)
+        {
+            buffer = program->simd16.m_funcRelocationTable;
+            size = program->simd16.m_funcRelocationTableSize;
+            entries = program->simd16.m_funcRelocationTableEntries;
+        }
+        else if (annotations.m_executionEnivronment.CompiledSIMDSize == 32)
+        {
+            buffer = program->simd32.m_funcRelocationTable;
+            size = program->simd32.m_funcRelocationTableSize;
+            entries = program->simd32.m_funcRelocationTableEntries;
+        }
+
+        if (size > 0)
+        {
+            patch.Size = sizeof(patch) + size;
+            patch.NumEntries = entries;
+
+            retValue = AddPatchItem(patch, membuf);
+
+            if (!membuf.Write((const char*)buffer, size))
+            {
+                retValue.Success = false;
+                return retValue;
+            }
+            freeBlock(buffer);
+        }
+    }
+
     return retValue;
 }
 
diff --git a/IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_list.h b/IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_list.h
@@ -34,7 +34,7 @@ Abstract:   Contains common patch structure definitions
 namespace iOpenCL
 {
 
-const uint32_t CURRENT_ICBE_VERSION = 1057;
+const uint32_t CURRENT_ICBE_VERSION = 1058;
 
 const uint32_t MAGIC_CL = 0x494E5443;      // 'I', 'N', 'T', 'C'
 const uint32_t INVALID_INDEX = 0xFFFFFFFF;
@@ -147,13 +147,15 @@ enum PATCH_TOKEN
     PATCH_TOKEN_CONSTRUCTOR_DESTRUCTOR_KERNEL_PROGRAM_BINARY_INFO, // 49  - (Unused)
     PATCH_TOKEN_INLINE_VME_SAMPLER_INFO,                           // 50  - (Unused)
     PATCH_TOKEN_GTPIN_FREE_GRF_INFO,                               // 51  @SPatchGtpinFreeGRFInfo@
-    PATCH_TOKEN_GTPIN_INFO,
+    PATCH_TOKEN_GTPIN_INFO,                                        // 52  @SPatchItemHeader@
+    PATCH_TOKEN_FUNCTION_SYMBOL_TABLE,                             // 53  @SPatchFunctionTableInfo@
+    PATCH_TOKEN_FUNCTION_RELOCATION_TABLE,                         // 54  @SPatchFunctionTableInfo@
 
     NUM_PATCH_TOKENS
 };
 
 // Update CURRENT_ICBE_VERSION when modifying the patch list
-static_assert( NUM_PATCH_TOKENS == 53, "NUM_PATCH_TOKENS has invalid value");
+static_assert( NUM_PATCH_TOKENS == 55, "NUM_PATCH_TOKENS has invalid value");
 
 /*****************************************************************************\
 ENUM: IMAGE_MEMORY_OBJECT_TYPE
diff --git a/IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_shared.h b/IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_shared.h
@@ -540,6 +540,18 @@ struct SPatchGtpinFreeGRFInfo :
 // Update CURRENT_ICBE_VERSION when modifying the patch list
 static_assert(sizeof(SPatchGtpinFreeGRFInfo) == (4 + sizeof(SPatchItemHeader)), "The size of SPatchGtpinFreeGRFInfo is not what is expected");
 
+/*****************************************************************************\
+ STRUCT: SPatchFunctionTableInfo
+ \*****************************************************************************/
+struct SPatchFunctionTableInfo :
+    SPatchItemHeader
+{
+    uint32_t   NumEntries;
+};
+
+// Update CURRENT_ICBE_VERSION when modifying the patch list
+static_assert(sizeof(SPatchFunctionTableInfo) == (4 + sizeof(SPatchItemHeader)), "The size of SPatchFunctionTableInfo is not what is expected");
+
 
 } // namespace
 #pragma pack( pop )
diff --git a/IGC/Compiler/CISACodeGen/CISABuilder.cpp b/IGC/Compiler/CISACodeGen/CISABuilder.cpp
@@ -39,6 +39,7 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #include "common/shaderOverride.hpp"
 #include "common/CompilerStatsUtils.hpp"
 #include "inc/common/sku_wa.h"
+#include "inc/common/RelocationInfo.h"
 #include <iStdLib/utility.h>
 
 #if !defined(_WIN32)
@@ -4214,6 +4215,75 @@ bool CEncoder::AvoidRetryOnSmallSpill() const
         context->m_retryManager.IsFirstTry();
 }
 
+void CEncoder::CreateFunctionSymbolTable(void*& buffer, unsigned& bufferSize, unsigned& tableEntries)
+{
+    buffer = nullptr;
+    bufferSize = 0;
+    tableEntries = 0;
+
+    if (IGC_IS_FLAG_ENABLED(EnableFunctionPointer))
+    {
+        Module* pModule = m_program->GetContext()->getModule();
+        std::vector<Function*> funcsToExport;
+        for (auto &F : pModule->getFunctionList())
+        {
+            // Find all functions in the module we need to export as symbols
+            if (F.hasFnAttribute("AsFunctionPointer"))
+            {
+                if (!F.isDeclaration() || F.getNumUses() > 0)
+                    funcsToExport.push_back(&F);
+            }
+        }
+
+        if (funcsToExport.empty())
+            return;
+
+        // Allocate buffer to store symbol table entries
+        tableEntries = funcsToExport.size();
+        bufferSize = sizeof(IGC::GenSymEntry) * tableEntries;
+        buffer = (void*) malloc(bufferSize);
+        assert(buffer && "Function Symbol Table not allocated");
+        IGC::GenSymEntry* entry_ptr = (IGC::GenSymEntry*) buffer;
+
+        for (auto pFunc : funcsToExport)
+        {
+            assert(pFunc->getName().size() <= IGC::MAX_SYMBOL_NAME_LENGTH);
+            strcpy(entry_ptr->s_name, pFunc->getName().str().c_str());
+
+            if (pFunc->isDeclaration())
+            {
+                // If the function is only declared, set as undefined type
+                entry_ptr->s_type = IGC::GenSymType::S_UNDEF;
+                entry_ptr->s_offset = 0;
+            }
+            else
+            {
+                auto Iter = stackFuncMap.find(pFunc);
+                assert(Iter != stackFuncMap.end() && "vISA function not found");
+
+                // Query vISA for the function's byte offset within the compiled module
+                VISAFunction* visaFunc = Iter->second;
+                entry_ptr->s_type = IGC::GenSymType::S_FUNC;
+                entry_ptr->s_offset = (uint32_t) visaFunc->getGenOffset();
+            }
+            entry_ptr++;
+        }
+    }
+}
+void CEncoder::CreateFunctionRelocationTable(void*& buffer, unsigned& bufferSize, unsigned& tableEntries)
+{
+    buffer = nullptr;
+    bufferSize = 0;
+    tableEntries = 0;
+
+    if (IGC_IS_FLAG_ENABLED(EnableFunctionPointer))
+    {
+        // vISA will directly return the buffer with GenRelocEntry layout
+        V(vMainKernel->GetGenRelocEntryBuffer(buffer, bufferSize, tableEntries));
+        assert(sizeof(IGC::GenRelocEntry) * tableEntries == bufferSize);
+    }
+}
+
 void CEncoder::Compile()
 {
     COMPILER_TIME_START(m_program->GetContext(), TIME_CG_vISAEmitPass);
@@ -4450,6 +4520,13 @@ void CEncoder::Compile()
 
     vMainKernel->GetGTPinBuffer(pOutput->m_gtpinBuffer, pOutput->m_gtpinBufferSize);
 
+    CreateFunctionSymbolTable(pOutput->m_funcSymbolTable,
+                              pOutput->m_funcSymbolTableSize,
+                              pOutput->m_funcSymbolTableEntries);
+    CreateFunctionRelocationTable(pOutput->m_funcRelocationTable,
+                                  pOutput->m_funcRelocationTableSize,
+                                  pOutput->m_funcRelocationTableEntries);
+
     if (jitInfo->isSpill == true)
     {
         pOutput->m_scratchSpaceUsedBySpills = jitInfo->spillMemUsed;
diff --git a/IGC/Compiler/CISACodeGen/CISABuilder.hpp b/IGC/Compiler/CISACodeGen/CISABuilder.hpp
@@ -489,6 +489,9 @@ class CEncoder
     // save compile time by avoiding retry if the amount of spill is (very) small
     bool AvoidRetryOnSmallSpill() const;
 
+    void CreateFunctionSymbolTable(void*& buffer, unsigned& bufferSize, unsigned& tableEntries);
+    void CreateFunctionRelocationTable(void*& buffer, unsigned& bufferSize, unsigned& tableEntries);
+
 protected:
     // encoder states
     SEncoderState m_encoderState;
diff --git a/IGC/Compiler/CISACodeGen/CShader.cpp b/IGC/Compiler/CISACodeGen/CShader.cpp
@@ -533,10 +533,8 @@ void  CShader::CreateConstantBufferOutput(SKernelProgram *pKernelProgram)
 
 void CShader::CreateFuncSymbolToRegisterMap(llvm::Function* pFunc)
 {
+    // Functions with uses in this module requires relocation
     CVariable* funcAddr = GetSymbol(pFunc);
-    //CVariable* funcAddr32bit = GetNewVariable(1, ISA_TYPE_UD, EALIGN_GRF, true);
-    //encoder.Cast(funcAddr32bit, funcAddr);
-    //encoder.Push();
     encoder.AddFunctionSymbol(pFunc, funcAddr);
     encoder.Push();
 }
@@ -2232,7 +2230,8 @@ CVariable* CShader::GetSymbol(llvm::Value *value, bool fromConstantPool)
     if (Constant *C = llvm::dyn_cast<llvm::Constant>(value))
     {
         // Function Pointer
-        if (isa<GlobalValue>(value) &&
+        if (IGC_IS_FLAG_ENABLED(EnableFunctionPointer) &&
+            isa<GlobalValue>(value) &&
             value->getType()->isPointerTy() &&
             value->getType()->getPointerElementType()->isFunctionTy())
         {
diff --git a/IGC/Compiler/CISACodeGen/EmitVISAPass.cpp b/IGC/Compiler/CISACodeGen/EmitVISAPass.cpp
@@ -414,7 +414,7 @@ bool EmitPass::runOnFunction(llvm::Function &F)
                     // Creates a mapping of the function symbol to a register.
                     // Any user function used by the kernel, including function declarations,
                     // should have a register allocated to store it's physical address
-                    if (F.getNumUses() > 0 && F.hasFnAttribute("AsFunctionPointer"))
+                    if (F.hasFnAttribute("AsFunctionPointer") && F.getNumUses() > 0)
                     {
                         m_currShader->CreateFuncSymbolToRegisterMap(&F);
                     }
diff --git a/IGC/Compiler/CodeGenPublic.h b/IGC/Compiler/CodeGenPublic.h
@@ -102,6 +102,12 @@ namespace IGC
         unsigned int    m_InstructionCount;
         void*           m_gtpinBuffer;              // Will be populated by VISA only when special switch is passed by gtpin
         unsigned int    m_gtpinBufferSize;
+        void*           m_funcSymbolTable;
+        unsigned int    m_funcSymbolTableSize;
+        unsigned int    m_funcSymbolTableEntries;
+        void*           m_funcRelocationTable;
+        unsigned int    m_funcRelocationTableSize;
+        unsigned int    m_funcRelocationTableEntries;
 
         void Destroy()
         {

Original file line number	Diff line number	Diff line change
`@@ -533,10 +533,8 @@ void CShader::CreateConstantBufferOutput(SKernelProgram *pKernelProgram)`
`533`	`533`
`534`	`534`	`void CShader::CreateFuncSymbolToRegisterMap(llvm::Function* pFunc)`
`535`	`535`	`{`
	`536`	`+ // Functions with uses in this module requires relocation`
`536`	`537`	`CVariable* funcAddr = GetSymbol(pFunc);`
`537`		`- //CVariable* funcAddr32bit = GetNewVariable(1, ISA_TYPE_UD, EALIGN_GRF, true);`
`538`		`- //encoder.Cast(funcAddr32bit, funcAddr);`
`539`		`- //encoder.Push();`
`540`	`538`	`encoder.AddFunctionSymbol(pFunc, funcAddr);`
`541`	`539`	`encoder.Push();`
`542`	`540`	`}`
`@@ -2232,7 +2230,8 @@ CVariable* CShader::GetSymbol(llvm::Value *value, bool fromConstantPool)`
`2232`	`2230`	`if (Constant *C = llvm::dyn_cast<llvm::Constant>(value))`
`2233`	`2231`	`{`
`2234`	`2232`	`// Function Pointer`
`2235`		`- if (isa<GlobalValue>(value) &&`
	`2233`	`+ if (IGC_IS_FLAG_ENABLED(EnableFunctionPointer) &&`
	`2234`	`+ isa<GlobalValue>(value) &&`
`2236`	`2235`	`value->getType()->isPointerTy() &&`
`2237`	`2236`	`value->getType()->getPointerElementType()->isFunctionTy())`
`2238`	`2237`	`{`
Original file line number	Diff line number	Diff line change
`@@ -414,7 +414,7 @@ bool EmitPass::runOnFunction(llvm::Function &F)`
`414`	`414`	`// Creates a mapping of the function symbol to a register.`
`415`	`415`	`// Any user function used by the kernel, including function declarations,`
`416`	`416`	`// should have a register allocated to store it's physical address`
`417`		`- if (F.getNumUses() > 0 && F.hasFnAttribute("AsFunctionPointer"))`
	`417`	`+ if (F.hasFnAttribute("AsFunctionPointer") && F.getNumUses() > 0)`
`418`	`418`	`{`
`419`	`419`	`m_currShader->CreateFuncSymbolToRegisterMap(&F);`
`420`	`420`	`}`