Fixing bugs 169, 170, 171

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 13761]

10 files changed, 76 insertions, 23 deletions

diff --git a/CHANGES b/CHANGES
index 4697bd6..33ceee2 100644
--- a/CHANGES
+++ b/CHANGES
@@ -10,6 +10,11 @@ Version 3.1.1+edits (development branch) versus 3.1.1
     - Adding support for atomic operations with generic memory space.  Before
       this, atomic operations can only work on global memory space.  (Bug #14
       external)
+    - Fixed a bug in conversion to ptxplus that was causing local memory store 
+      or load instruction to be ignored.
+    - Adding support for direct addressing using immediate values for the load
+      and store instructions.
+    - Fixed a bug that was causing inconsistency in local memory address calculations
 
 Version 3.1.1 versus 3.1.0
 - Add checks to top level makefile to ensure setup_environment is run and checks to
diff --git a/cuobjdump_to_ptxplus/cuobjdumpInst.cc b/cuobjdump_to_ptxplus/cuobjdumpInst.cc
index dd9bb99..22dbdd3 100644
--- a/cuobjdump_to_ptxplus/cuobjdumpInst.cc
+++ b/cuobjdump_to_ptxplus/cuobjdumpInst.cc
@@ -539,7 +539,8 @@ void cuobjdumpInst::printCuobjdumpOperand(std::string currentPiece, std::string
 					(m_base == "R2G") ||
 					(m_base == "GLD") ||
 					(m_base == "GST") ||
-					(m_base == "LST") ))) {
+					(m_base == "LST") ||
+					(m_base == "LLD")))) {
 			std::string modsub = mod.substr(1);
 			int regNumInt = atoi(modsub.c_str());
 			std::stringstream temp;
@@ -576,8 +577,7 @@ void cuobjdumpInst::printCuobjdumpOperand(std::string currentPiece, std::string
 		std::string modsub2;
 		std::string modsub3;
 		modsub = mod.c_str();
-		int localFlag = 0;
-
+		int const_sharedFlag =0;
 		if(mod.find("global14") != std::string::npos) {
 			//Those instructions don't need the dereferencing done by g [*]
 			if(	base == "GRED" ||
@@ -590,15 +590,23 @@ void cuobjdumpInst::printCuobjdumpOperand(std::string currentPiece, std::string
 		} else if(mod[0]=='g') {
 			//Shared memory
 			output("s[");
+			const_sharedFlag=1;
 		} else if(mod.find("local") !=  std::string::npos) {
-			modsub3 = modsub.substr(4, modsub.length()-10);
-			output(modsub3.c_str());
+			if((base=="LST")||
+				(base=="LLD"))
 			output("[");
-			localFlag = 1;
+		else
+			output("l[");
+			//modsub3 = modsub.substr(4, modsub.length()-10);
+			//output(modsub3.c_str());
+			//output("[");
+			//localFlag = 1;
 		} else if(mod.substr(0,9) == "constant1") {
 			output(modsub.substr(0, modsub.find_first_of("[]")+1).c_str());
+			const_sharedFlag=1;
 		} else if(mod.substr(0,9)=="constant0"){
 			output("constant0[");
+			const_sharedFlag=1;
 		} else {
 			printf("Unidentified modifier: %s\n", mod.c_str());
 			assert(0);
@@ -636,14 +644,17 @@ void cuobjdumpInst::printCuobjdumpOperand(std::string currentPiece, std::string
 				std::stringstream hexStringConvert;
 				hexStringConvert << std::hex << modsub2;
 				hexStringConvert >> addrValue;
-				unsigned chunksize = 4;
-				if (	this->m_typeModifiers->size()>0 &&
-						(	m_typeModifiers->back() == ".S16" ||
-							m_typeModifiers->back() == ".U16")) chunksize = 2;
-				if (	this->m_typeModifiers->size()>0 &&
-						(	m_typeModifiers->back() == ".U8" ||
-							m_typeModifiers->back() == ".S8")) chunksize = 1;
-				addrValue = addrValue*chunksize;
+				if(const_sharedFlag == 1)
+				{
+					unsigned chunksize = 4;
+					if (	this->m_typeModifiers->size()>0 &&
+							(	m_typeModifiers->back() == ".S16" ||
+								m_typeModifiers->back() == ".U16")) chunksize = 2;
+					if (	this->m_typeModifiers->size()>0 &&
+							(	m_typeModifiers->back() == ".U8" ||
+								m_typeModifiers->back() == ".S8")) chunksize = 1;
+					addrValue = addrValue*chunksize;
+				}
 				char outputHex[10];
 				sprintf(outputHex, "%x", addrValue);
 				std::stringstream outputhex;
@@ -669,7 +680,7 @@ void cuobjdumpInst::printCuobjdumpOperand(std::string currentPiece, std::string
 				hexStringConvert << std::hex << modsub;
 				hexStringConvert >> addrValue;
 
-				if(localFlag == 0)
+				if(const_sharedFlag == 1)
 				{
 					unsigned chunksize = 4;
 					if ( m_typeModifiers->size()>0 &&
diff --git a/cuobjdump_to_ptxplus/cuobjdumpInstList.cc b/cuobjdump_to_ptxplus/cuobjdumpInstList.cc
index fabc396..d3ebd9a 100644
--- a/cuobjdump_to_ptxplus/cuobjdumpInstList.cc
+++ b/cuobjdump_to_ptxplus/cuobjdumpInstList.cc
@@ -350,7 +350,7 @@ void cuobjdumpInstList::addCuobjdumpRegister(std::string reg, bool lo)
 		(typeModifiers->front() == ".S64") &&
 		((baseInst == "G2R")||(baseInst == "R2G")||
 		(baseInst == "GLD")||(baseInst == "GST")||
-		(baseInst == "LST"))))
+		(baseInst == "LST")|| (baseInst == "LLD"))))
 	{
 		vectorFlag = 64;
 	}
@@ -426,6 +426,7 @@ void cuobjdumpInstList::addCuobjdumpMemoryOperand(std::string mem, int memType)
 	}
 
 	// Local memory
+	/*
 	if(memType == 3) {
 		std::stringstream out;
 		printf("Trying to find lmem for: %s\n", m_entryList.back().m_entryName.c_str());
@@ -434,7 +435,7 @@ void cuobjdumpInstList::addCuobjdumpMemoryOperand(std::string mem, int memType)
 		out << "l" << kernellmemmap[m_entryList.back().m_entryName];// << mem;
 		mem = out.str();
 	}
-
+	*/
 	// Add the memory operand to instruction operand list
 	char* memName = new char [strlen(mem.c_str())+1];
 	strcpy(memName, mem.c_str());
diff --git a/cuobjdump_to_ptxplus/ptx_parser.h b/cuobjdump_to_ptxplus/ptx_parser.h
index c6d7027..ce77c2f 100644
--- a/cuobjdump_to_ptxplus/ptx_parser.h
+++ b/cuobjdump_to_ptxplus/ptx_parser.h
@@ -90,6 +90,7 @@ void add_pred( const char *a, int b, int c ) {DPRINTF(" ");}
 void add_scalar_operand( const char *a ) {DPRINTF("%s", a);}
 void add_neg_pred_operand( const char *a ) {DPRINTF(" ");}
 void add_address_operand( const char *a, int b ) {DPRINTF("%s", a);}
+void add_address_operand2( int b ) {DPRINTF(" ");}
 void change_operand_lohi( int a ) {DPRINTF(" ");}
 void change_double_operand_type( int a ) {DPRINTF(" ");}
 void change_operand_neg( ) {DPRINTF(" ");}
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index 4cad472..e6677fa 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -813,6 +813,9 @@ void ptx_instruction::pre_decode()
                   // TODO: first address register in $r1+=$r2 should be an output register as well
                }
             }
+            else if(o.is_immediate_address()){
+
+            }
             // Regular PTX operand
             else if (o.get_symbol()->type()->get_key().is_reg()) { // Memory operand contains a register
               ar1 = o.reg_num();
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index a7685aa..237bfd5 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -106,7 +106,9 @@ ptx_reg_t ptx_thread_info::get_operand_value( const operand_info &op, operand_in
             result = get_reg( op.get_symbol() );
          } else if ( op.is_builtin()) {
             result = get_builtin( op.get_int(), op.get_addr_offset() );
-         } else if ( op.is_memory_operand() ) {
+         } else  if(op.is_immediate_address()){
+    		 result = op.get_addr_offset();
+    	 } else if ( op.is_memory_operand() ) {
             // a few options here...
             const symbol *sym = op.get_symbol();
             const type_info *type = sym->type();
diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y
index 8b65796..2c01a70 100644
--- a/src/cuda-sim/ptx.y
+++ b/src/cuda-sim/ptx.y
@@ -527,6 +527,7 @@ address_expression: IDENTIFIER { add_address_operand($1,0); }
 	| IDENTIFIER LO_OPTION { add_address_operand($1,0); change_operand_lohi(1);}
 	| IDENTIFIER HI_OPTION { add_address_operand($1,0); change_operand_lohi(2); }
 	| IDENTIFIER PLUS INT_OPERAND { add_address_operand($1,$3); }
+	| INT_OPERAND { add_address_operand2($1); }
 	;
 
 %%
diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h
index 15c6743..4078e6b 100644
--- a/src/cuda-sim/ptx_ir.h
+++ b/src/cuda-sim/ptx_ir.h
@@ -342,6 +342,7 @@ public:
       m_const_mem_offset = 0;
       m_uid = get_uid();
       m_valid = false;
+      m_immediate_address=false;
    }
    operand_info( const symbol *addr )
    {
@@ -376,6 +377,7 @@ public:
       m_vector = false;
       m_neg_pred = false;
       m_is_return_var = false;
+      m_immediate_address=false;
    }
    operand_info( const symbol *addr1, const symbol *addr2 )
    {
@@ -396,6 +398,7 @@ public:
       m_vector = false;
       m_neg_pred = false;
       m_is_return_var = false;
+      m_immediate_address=false;
    }
    operand_info( int builtin_id, int dim_mod )
    {
@@ -412,6 +415,7 @@ public:
       m_addr_offset = dim_mod;
       m_neg_pred = false;
       m_is_return_var = false;
+      m_immediate_address=false;
    }
    operand_info( const symbol *addr, int offset )
    {
@@ -428,6 +432,7 @@ public:
       m_addr_offset = offset;
       m_neg_pred = false;
       m_is_return_var = false;
+      m_immediate_address=false;
    }
    operand_info( unsigned x )
    {
@@ -441,9 +446,10 @@ public:
       m_vector = false;
       m_type = unsigned_t;
       m_value.m_unsigned = x;
-      m_addr_offset = 0;
+      m_addr_offset = x;
       m_neg_pred = false;
       m_is_return_var = false;
+      m_immediate_address=true;
    }
    operand_info( int x )
    {
@@ -460,6 +466,7 @@ public:
       m_addr_offset = 0;
       m_neg_pred = false;
       m_is_return_var = false;
+      m_immediate_address=false;
    }
    operand_info( float x )
    {
@@ -476,6 +483,7 @@ public:
       m_addr_offset = 0;
       m_neg_pred = false;
       m_is_return_var = false;
+      m_immediate_address=false;
    }
    operand_info( double x )
    {
@@ -492,6 +500,7 @@ public:
       m_addr_offset = 0;
       m_neg_pred = false;
       m_is_return_var = false;
+      m_immediate_address=false;
    }
    operand_info( const symbol *s1, const symbol *s2, const symbol *s3, const symbol *s4 )
    {
@@ -512,11 +521,12 @@ public:
       m_addr_offset = 0;
       m_neg_pred = false;
       m_is_return_var = false;
+      m_immediate_address=false;
    }
 
    void make_memory_operand() { m_type = memory_t;}
    void set_return() { m_is_return_var = true; }
-
+   void set_immediate_addr() {m_immediate_address=true;}
    const std::string &name() const
    {
       assert( m_type == symbolic_t || m_type == reg_t || m_type == address_t || m_type == memory_t || m_type == label_t);
@@ -606,6 +616,10 @@ public:
       return (m_addr_space!=undefined_space); 
    }
 
+   bool is_immediate_address() const {
+	   return   m_immediate_address;
+   }
+
    bool is_literal() const { return m_type == int_t ||
       m_type == float_op_t ||
       m_type == double_op_t ||
@@ -678,7 +692,7 @@ private:
    bool m_valid;
    bool m_vector;
    enum operand_type m_type;
-
+   bool m_immediate_address;
    enum _memory_space_t m_addr_space;
    int m_operand_lohi;
    int m_double_operand_type;
diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc
index cc97b72..76d9a9c 100644
--- a/src/cuda-sim/ptx_parser.cc
+++ b/src/cuda-sim/ptx_parser.cc
@@ -689,8 +689,8 @@ void change_memory_addr_space(const char *identifier)
    if(!strcmp(l, "l"))
    {
        g_operands.back().set_addr_space(local_space);
-       parse_assert(g_current_symbol_table->lookup(identifier) != NULL, "Local memory segment was not defined.");
-       g_operands.back().set_const_mem_offset(g_current_symbol_table->lookup(identifier)->get_address());
+       //parse_assert(g_current_symbol_table->lookup(identifier) != NULL, "Local memory segment was not defined.");
+       //g_operands.back().set_const_mem_offset(g_current_symbol_table->lookup(identifier)->get_address());
        recognizedType = true;
    }
 
@@ -711,6 +711,13 @@ void change_operand_lohi( int lohi )
 
 }
 
+void set_immediate_operand_type ()
+{
+	 DPRINTF("set_immediate_operand_type");
+	 assert( !g_operands.empty() );
+	 g_operands.back().set_immediate_addr();
+}
+
 void change_double_operand_type( int operand_type )
 {
    /*
@@ -809,6 +816,12 @@ void add_address_operand( const char *identifier, int offset )
    g_operands.push_back( operand_info(s,offset) );
 }
 
+void add_address_operand2( int offset )
+{
+   DPRINTF("add_address_operand");
+   g_operands.push_back( operand_info((unsigned)offset) );
+}
+
 void add_array_initializer()
 {
    g_last_symbol->add_initializer(g_operands);
diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h
index 561fd74..57d763e 100644
--- a/src/cuda-sim/ptx_parser.h
+++ b/src/cuda-sim/ptx_parser.h
@@ -64,6 +64,7 @@ void add_literal_int( int value );
 void add_literal_float( float value );
 void add_literal_double( double value );
 void add_address_operand( const char *identifier, int offset );
+void add_address_operand2( int offset );
 void add_label( const char *idenfiier );
 void add_vector_spec(int spec );
 void add_space_spec( enum _memory_space_t spec, int value );
@@ -89,6 +90,7 @@ void change_memory_addr_space( const char *identifier );
 void change_operand_lohi( int lohi );
 void change_double_operand_type( int addr_type );
 void change_operand_neg( );
+void set_immediate_operand_type( );
 void version_header(double a);
 #ifdef __cplusplus 
 }