No change

No change

Author: pratikashar

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -4488,6 +4488,15 @@ namespace IGC
             SaveOption(vISA_TotalGRFNum, context->getNumGRFPerThread());
         }
 
+        if (m_program->HasStackCalls())
+        {
+            bool ZEBinEnabled = IGC_IS_FLAG_ENABLED(EnableZEBinary) || context->getCompilerOption().EnableZEBinary;
+
+            // pass higher ABI version when using ZEBinary
+            if (ZEBinEnabled)
+                SaveOption(vISA_StackCallABIVer, (uint32_t)2);
+        }
+
         //
         // Setting number of GRF and threads per EU is restricted to OCL only
         // Number of threads can be set by:

IGC/ZEBinWriter/zebin/source/ZEELFObjectBuilder.cpp

@@ -815,8 +815,15 @@ void ELFWriter::writeHeader()
     // e_ident[EI_VERSION]
     m_W.OS << char(ELF::EV_CURRENT);
 
+    // e_ident[EI_OSABI]
+    m_W.OS << char(ELF::ELFOSABI_NONE);
+
+    // ABI version is hard coded right now. In future, if more ABI versions exist then revisit this.
+    // e_ident[EI_ABIVERSION]
+    m_W.OS << char(1);
+
     // e_ident padding
-    m_W.OS.write_zeros(ELF::EI_NIDENT - ELF::EI_OSABI);
+    m_W.OS.write_zeros(ELF::EI_NIDENT - ELF::EI_PAD);
 
     // e_type: Currently IGC always emits a relocatable file
     m_W.write<uint16_t>(ELF::ET_REL);

visa/G4_Kernel.cpp

@@ -863,7 +863,7 @@ KernelDebugInfo* G4_Kernel::getKernelDebugInfo()
 
 unsigned G4_Kernel::getStackCallStartReg() const
 {
-    // Last 3 GRFs to be used as scratch
+    // Last 3 (or 2) GRFs reserved for stack call purpose
     unsigned totalGRFs = getNumRegTotal();
     unsigned startReg = totalGRFs - numReservedABIGRF();
     return startReg;

visa/G4_Kernel.hpp

@@ -344,6 +344,7 @@ class G4_Kernel
         return getKernelAttrs()->getInt32KernelAttr(aID);
     }
     bool getOption(vISAOptions opt) const { return m_options->getOption(opt); }
+    uint32_t getuInt32Option(vISAOptions opt) const { return m_options->getuInt32Option(opt); }
     void computeChannelSlicing();
     void calculateSimdSize();
     G4_ExecSize getSimdSize() { return simdSize; }
@@ -388,10 +389,26 @@ class G4_Kernel
     unsigned calleeSaveStart() const;
     unsigned getNumCalleeSaveRegs() const;
 
+    enum StackCallABIVersion
+    {
+        VER_1 = 1, // This version is used pre-zebin
+        VER_2 = 2, // This version is used for zebin
+    };
+
     // return the number of reserved GRFs for stack call ABI
     // the reserved registers are at the end of the GRF file (e.g., r125-r127)
+    // for some architectures/ABI version, we dont need a copy of r0 and
+    // instructions can directly refer to r0 in code (eg, sends).
     uint32_t numReservedABIGRF() const {
-        return 3;
+        if (getuInt32Option(vISA_StackCallABIVer) == VER_1)
+            return 3;
+        else
+        {
+            // for ABI version > 1,
+            if (getOption(vISA_PreserveR0InR0))
+                return 2;
+            return 3;
+        }
     }
 
     // purpose of the GRFs reserved for stack call ABI
@@ -400,15 +417,30 @@ class G4_Kernel
     const int ThreadHeaderGRF = 2;
 
     uint32_t getFPSPGRF() const{
-        return getStackCallStartReg() + FPSPGRF;
+        // For ABI V1 return r125.
+        // For ABI V2 return r127.
+        if(getuInt32Option(vISA_StackCallABIVer) == VER_1)
+            return getStackCallStartReg() + FPSPGRF;
+        else
+            return (getNumRegTotal() - 1) - FPSPGRF;
     }
 
     uint32_t getSpillHeaderGRF() const{
-        return getStackCallStartReg() + SpillHeaderGRF;
+        // For ABI V1 return r126.
+        // For ABI V2 return r126.
+        if (getuInt32Option(vISA_StackCallABIVer) == VER_1)
+            return getStackCallStartReg() + SpillHeaderGRF;
+        else
+            return (getNumRegTotal() - 1) - SpillHeaderGRF;
     }
 
     uint32_t getThreadHeaderGRF() const{
-        return getStackCallStartReg() + ThreadHeaderGRF;
+        // For ABI V1 return r127.
+        // For ABI V2 return r125.
+        if (getuInt32Option(vISA_StackCallABIVer) == VER_1)
+            return getStackCallStartReg() + ThreadHeaderGRF;
+        else
+            return (getNumRegTotal() - 1) - ThreadHeaderGRF;
     }
 
     void renameAliasDeclares();

visa/include/VISAOptionsDefs.h

@@ -134,6 +134,7 @@ DEF_VISA_OPTION(vISA_CodePatch,   ET_INT32, "-codePatch",        UNUSED, 0)
 DEF_VISA_OPTION(vISA_Linker,      ET_INT32, "-linker",        UNUSED, 0)
 DEF_VISA_OPTION(vISA_lscEnableImmOffsFor,   ET_INT32, "-lscEnableImmOffsFor", UNUSED, 0x3001E)
 DEF_VISA_OPTION(vISA_PreserveR0InR0,        ET_BOOL, "-preserver0", UNUSED, false)
+DEF_VISA_OPTION(vISA_StackCallABIVer,       ET_INT32, "-abiver", UNUSED, 1)
 
 //=== RA options ===
 DEF_VISA_OPTION(vISA_RoundRobin,            ET_BOOL, "-noroundrobin",    UNUSED, true)