template compiled with "make noinline=1" now parses, runs, passes

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 6857]

4 files changed, 28 insertions, 8 deletions

diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l
index 0ab7522..0d66bfd 100644
--- a/src/cuda-sim/ptx.l
+++ b/src/cuda-sim/ptx.l
@@ -174,8 +174,8 @@ xor     TC; ptx_lval.int_value = XOR_OP; return OPCODE;
 \.const TC; return CONST_DIRECTIVE;
 \.entry TC; return ENTRY_DIRECTIVE;
 \.extern TC; return EXTERN_DIRECTIVE;
-\.file	 TC; return FILE_DIRECTIVE;
-\.func   TC; return FUNC_DIRECTIVE;
+\.file	 TC; BEGIN(INITIAL); return FILE_DIRECTIVE;
+\.func   TC; BEGIN(NOT_OPCODE); return FUNC_DIRECTIVE;
 \.global TC; return GLOBAL_DIRECTIVE;
 \.local  TC; return LOCAL_DIRECTIVE;
 \.loc    TC; return LOC_DIRECTIVE;
@@ -220,8 +220,8 @@ xor     TC; ptx_lval.int_value = XOR_OP; return OPCODE;
 "%warpid"  TC; ptx_lval.int_value = WARPID_REG; return SPECIAL_REGISTER;
 "WARP_SZ"  TC; ptx_lval.int_value = WARPSZ_REG; return SPECIAL_REGISTER;
 
-[a-zA-Z][a-zA-Z0-9_$]*  TC; ptx_lval.string_value = strdup(yytext); return IDENTIFIER;
-[_$%][a-zA-Z0-9_$]+  TC; ptx_lval.string_value = strdup(yytext); return IDENTIFIER;
+[a-zA-Z_][a-zA-Z0-9_$]*  TC; ptx_lval.string_value = strdup(yytext); return IDENTIFIER;
+[$%][a-zA-Z0-9_$]+  TC; ptx_lval.string_value = strdup(yytext); return IDENTIFIER;
 
 [0-9]+\.[0-9]+ 	 TC; sscanf(yytext,"%lf", &ptx_lval.double_value); return DOUBLE_OPERAND;
 	
@@ -340,7 +340,7 @@ xor     TC; ptx_lval.int_value = XOR_OP; return OPCODE;
 ";"	TC; BEGIN(INITIAL); return SEMI_COLON;
 "!"	TC; return EXCLAMATION;
 "=" 	TC; return EQUALS;
-"{"	TC; return LEFT_BRACE;
+"{"	TC; BEGIN(INITIAL); return LEFT_BRACE;
 "}"	TC; return RIGHT_BRACE;
 \.	TC; return PERIOD;
 
diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y
index dd69fe5..9ddb214 100644
--- a/src/cuda-sim/ptx.y
+++ b/src/cuda-sim/ptx.y
@@ -217,7 +217,7 @@ input:	/* empty */
 	| input function_decl
 	;
 
-function_defn: function_decl { set_symtab($1); } LEFT_BRACE statement_list RIGHT_BRACE { end_function(); }
+function_defn: function_decl { set_symtab($1); start_function_definition(); } LEFT_BRACE statement_list RIGHT_BRACE { end_function(); }
 	| function_decl { set_symtab($1); } block_spec LEFT_BRACE statement_list RIGHT_BRACE { end_function(); }
 	;
 
@@ -255,6 +255,7 @@ directive_statement: variable_declaration SEMI_COLON
 	| TARGET_DIRECTIVE IDENTIFIER 
 	| FILE_DIRECTIVE INT_OPERAND STRING { add_file($2,$3); } 
 	| LOC_DIRECTIVE INT_OPERAND INT_OPERAND INT_OPERAND 
+	| PRAGMA_DIRECTIVE STRING SEMI_COLON { add_pragma($2); }
 	;
 
 variable_declaration: variable_spec identifier_list { add_variables(); }
diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc
index 542426a..c981802 100644
--- a/src/cuda-sim/ptx_ir.cc
+++ b/src/cuda-sim/ptx_ir.cc
@@ -181,6 +181,14 @@ void start_function( int entry_point )
    g_entry_func_param_index=0;
 }
 
+static bool g_in_function_definition = false;
+
+void start_function_definition()
+{
+   DPRINTF("start_function_definition");
+   g_in_function_definition = true;
+}
+
 char *g_add_identifier_cached__identifier = NULL;
 int g_add_identifier_cached__array_dim;
 int g_add_identifier_cached__array_ident;
@@ -222,12 +230,14 @@ void end_function()
 {
    DPRINTF("end_function");
 
+   g_in_function_definition = false;
    init_directive_state();
    init_instruction_state();
    g_max_regs_per_thread = mymax( g_max_regs_per_thread, (g_current_symbol_table->next_reg_num()-1)); 
    g_func_info->add_inst( g_instructions );
    g_instructions.clear();
    gpgpu_ptx_assemble( g_func_info->get_name(), g_func_info );
+   g_current_symbol_table = g_global_symbol_table;
 
    DPRINTF("function %s, PC = %d\n", g_func_info->get_name().c_str(), g_func_info->get_start_PC());
 }
@@ -472,8 +482,10 @@ void add_identifier( const char *identifier, int array_dim, unsigned array_ident
 
 
    if ( ti.is_param() ) {
-      g_func_info->add_param_name_and_type(g_entry_func_param_index,identifier, ti.scalar_type() );
-      g_entry_func_param_index++;
+      if( !g_in_function_definition ) {
+         g_func_info->add_param_name_and_type(g_entry_func_param_index,identifier, ti.scalar_type() );
+         g_entry_func_param_index++;
+      }
    }
 }
 
@@ -1268,6 +1280,11 @@ extern "C" void set_symtab(void*symtab)
    g_current_symbol_table = (symbol_table*)symtab;
 }
 
+extern "C" void add_pragma( const char *str )
+{
+   printf("GPGPU-Sim: Warning -- ignoring pragma '%s'\n", str );
+}
+
 unsigned ptx_kernel_shmem_size( void *kernel_impl )
 {
    function_info *f = (function_info*)kernel_impl;
diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h
index 82d5b4c..aca3556 100644
--- a/src/cuda-sim/ptx_ir.h
+++ b/src/cuda-sim/ptx_ir.h
@@ -1176,6 +1176,7 @@ extern "C" {
 #endif 
 
    void start_function( int entry_point );
+   void start_function_definition();
    void add_function_name( const char *fname );
    void init_directive_state();
    void add_directive(); 
@@ -1210,6 +1211,7 @@ extern "C" {
    void add_file( unsigned num, const char *filename );
    void *reset_symtab();
    void set_symtab(void*);
+   void add_pragma( const char *str );
 
 
 #define NON_ARRAY_IDENTIFIER 1