From deee9038d3d67e60f106776be3dd0a846dd11df9 Mon Sep 17 00:00:00 2001 From: Tor Aamodt Date: Sun, 1 Apr 2018 17:31:14 -0700 Subject: fix regressions -- move call to pre_decode into do_pdom --- src/cuda-sim/cuda-sim.cc | 4 ++-- src/cuda-sim/ptx_ir.cc | 57 ++++++++++++++++++++++++++++-------------------- src/cuda-sim/ptx_ir.h | 2 ++ 3 files changed, 37 insertions(+), 26 deletions(-) diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 987e3f2..dce35ca 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -252,7 +252,7 @@ void function_info::ptx_assemble() target.set_type(label_t); } } - + m_n = n; printf(" done.\n"); fflush(stdout); @@ -282,7 +282,6 @@ void function_info::ptx_assemble() print_postdominators(); print_ipostdominators(); } -#endif printf("GPGPU-Sim PTX: pre-decoding instructions for \'%s\'...\n", m_name.c_str() ); for ( unsigned ii=0; ii < n; ii += m_instr_mem[ii]->inst_size() ) { // handle branch instructions @@ -293,6 +292,7 @@ void function_info::ptx_assemble() fflush(stdout); m_assembled = true; +#endif } addr_t shared_to_generic( unsigned smid, addr_t addr ) diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 17e91df..be25dbe 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -577,30 +577,39 @@ bool function_info::connect_break_targets() //connecting break instructions with return modified; } -void function_info::do_pdom() { - create_basic_blocks(); - connect_basic_blocks(); - bool modified = false; - do { - find_dominators(); - find_idominators(); - modified = connect_break_targets(); - } while (modified == true); - - if ( g_debug_execution>=50 ) { - print_basic_blocks(); - print_basic_block_links(); - print_basic_block_dot(); - } - if ( g_debug_execution>=2 ) { - print_dominators(); - } - find_postdominators(); - find_ipostdominators(); - if ( g_debug_execution>=50 ) { - print_postdominators(); - print_ipostdominators(); - } +void function_info::do_pdom() +{ + create_basic_blocks(); + connect_basic_blocks(); + bool modified = false; + do { + find_dominators(); + find_idominators(); + modified = connect_break_targets(); + } while (modified == true); + + if ( g_debug_execution>=50 ) { + print_basic_blocks(); + print_basic_block_links(); + print_basic_block_dot(); + } + if ( g_debug_execution>=2 ) { + print_dominators(); + } + find_postdominators(); + find_ipostdominators(); + if ( g_debug_execution>=50 ) { + print_postdominators(); + print_ipostdominators(); + } + printf("GPGPU-Sim PTX: pre-decoding instructions for \'%s\'...\n", m_name.c_str() ); + for ( unsigned ii=0; ii < m_n; ii += m_instr_mem[ii]->inst_size() ) { // handle branch instructions + ptx_instruction *pI = m_instr_mem[ii]; + pI->pre_decode(); + } + printf("GPGPU-Sim PTX: ... done pre-decoding instructions for \'%s\'.\n", m_name.c_str() ); + fflush(stdout); + m_assembled = true; } void intersect( std::set &A, const std::set &B ) { diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 26a2839..85b2a3b 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1308,6 +1308,8 @@ private: //parameter size for device kernels int m_args_aligned_size; + + addr_t m_n; // offset in m_instr_mem (used in do_pdom) }; class arg_buffer_t { -- cgit v1.3