|
Open64 (mfef90, whirl2f, and IR tools)
TAG: version-openad; SVN changeset: 916
|
|
Open64 (mfef90, whirl2f, and IR tools)
TAG: version-openad; SVN changeset: 916
|
00001 /* 00002 00003 Copyright (C) 2000, 2001 Silicon Graphics, Inc. All Rights Reserved. 00004 00005 This program is free software; you can redistribute it and/or modify it 00006 under the terms of version 2 of the GNU General Public License as 00007 published by the Free Software Foundation. 00008 00009 This program is distributed in the hope that it would be useful, but 00010 WITHOUT ANY WARRANTY; without even the implied warranty of 00011 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. 00012 00013 Further, this software is distributed without any warranty that it is 00014 free of the rightful claim of any third person regarding infringement 00015 or the like. Any license provided herein, whether implied or 00016 otherwise, applies only to this software file. Patent licenses, if 00017 any, provided herein do not apply to combinations of this program with 00018 other software, or any other product whatsoever. 00019 00020 You should have received a copy of the GNU General Public License along 00021 with this program; if not, write the Free Software Foundation, Inc., 59 00022 Temple Place - Suite 330, Boston MA 02111-1307, USA. 00023 00024 Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pky, 00025 Mountain View, CA 94043, or: 00026 00027 http://www.sgi.com 00028 00029 For further information regarding this notice, see: 00030 00031 http://oss.sgi.com/projects/GenInfo/NoticeExplan 00032 00033 */ 00034 00035 00036 /* ==================================================================== 00037 * ==================================================================== 00038 * 00039 * 00040 * Revision history: 00041 * 11-Nov-96 - Original Version. 00042 * 00043 * Description: 00044 * 00045 * Configure the -LNO option group (included in config.c). 00046 * 00047 * ==================================================================== 00048 * 00049 * TODO: at a convenient point: 00050 * 00051 * ==================================================================== 00052 * ==================================================================== 00053 */ 00054 00055 #ifdef _KEEP_RCS_ID 00056 /*REFERENCED*/ 00057 #endif /* _KEEP_RCS_ID */ 00058 00059 /* This file is included in config.c, so it doesn't need its own set of 00060 * standard includes -- only the following: 00061 */ 00062 #include "config_lno.h" 00063 #include "config_cache.h" 00064 00065 /* ==================================================================== 00066 * 00067 * -LNO: option group 00068 * 00069 * Define the defaults for -LNO option group flags. 00070 * 00071 * NOTE: There are two copies of this defined -- one to copy when 00072 * initializing a new struct to the defaults, and one to serve as the 00073 * initial bottom of the Current_LNO stack. Fix both of them when 00074 * adding/changing things. 00075 * 00076 * WARNING: The current memory model can handle four levels. In order 00077 * to change it, it is necessary to change the parameter in the file 00078 * ... as well as the initial value lists of the structs below, and the 00079 * _DUP option descriptor macros below. 00080 * 00081 * ==================================================================== 00082 */ 00083 00084 #define DEFAULT_UNROLL_PROD_MAX 16 00085 #define DEFAULT_UNROLL_MAX 10 00086 00087 BOOL Run_autopar = FALSE; 00088 00089 static LNO_FLAGS Default_LNO = { 00090 NULL, /* next */ 00091 NULL, /* _mhd */ 00092 00093 TRUE, /* Aequiv */ 00094 FALSE, /* Automatic Data Distribution */ 00095 1, /* Run_AP */ 00096 FALSE, /* Ill_xform_directives */ 00097 TRUE, /* Backward_substitution */ 00098 FALSE, /* Blind_loop_reversal */ 00099 TRUE, /* Blocking */ 00100 0, /* Blocking_Size */ 00101 TRUE, /* Cache_model_edge_effects */ 00102 TRUE, /* Coupled_opts */ 00103 TRUE, /* Cse */ 00104 FALSE, /* Fancy_tile */ 00105 FALSE, /* Run_fiz_fuse */ 00106 1, /* Fission */ 00107 0, /* Fission_inner_register_limit */ 00108 TRUE, /* Forward_substitution */ 00109 1, /* Fusion */ 00110 5, /* Fusion_peeling_limit */ 00111 1, /* Gather_Scatter */ 00112 0xfffe, /* Graph_capacity */ 00113 TRUE, /* Hoist_messy_bounds */ 00114 FALSE, /* Ignore_pragmas */ 00115 TRUE, /* Interchange */ 00116 FALSE, FALSE, /* Run_lego */ 00117 TRUE, /* Run_lego_localizer */ 00118 TRUE, /* Loop_finalization */ 00119 8, /* Max_do_loop_depth_strict */ 00120 FALSE, /* Mem_sim */ 00121 TRUE, /* Minvar */ 00122 1, /* Opt */ 00123 2, /* Cache_model */ 00124 TRUE, /* Run_outer */ 00125 0, /* Outer_unroll */ 00126 TRUE, /* Outer_unroll_deep */ 00127 6, /* Outer_unroll_min_for_further_unroll */ 00128 0, /* Outer_unroll_max */ 00129 FALSE, /* Outer_unroll_model_only */ 00130 0, /* Outer_unroll_prod_max */ 00131 TRUE, /* Outer_unroll_unity */ 00132 2, /* ou_aggre */ 00133 0, /* Run_p3 */ 00134 TRUE, /* Pseudo_lower */ 00135 0, FALSE, /* Run_prefetch */ 00136 2, /* Prefetch_ahead */ 00137 2, /* Prefetch_iters_ahead */ 00138 2, /* Prefetch_cache_factor */ 00139 FALSE, /* Prefetch_indirect */ 00140 FALSE, FALSE, /* Run_prefetch_manual */ 00141 TRUE, /* Power_of_two_hack */ 00142 TRUE, /* Sclrze */ 00143 0, /* SE_tile_size */ 00144 1, /* Split_tiles */ 00145 64, /* Split_tiles_size */ 00146 FALSE, /* Run_test */ 00147 FALSE, /* Test_dump */ 00148 TRUE, /* Trapezoidal_outer_unroll */ 00149 FALSE, /* Use_malloc */ 00150 TRUE, /* Use_parm */ 00151 FALSE, /* Verbose */ 00152 TRUE, /* Version_mp_loops */ 00153 TRUE, /* Run_vintr */ 00154 TRUE, /* Run_oinvar */ 00155 1, /* Run_doacross */ 00156 0, /* Preferred_doacross_tile_size */ 00157 2600, /* Parallel_overhead */ 00158 FALSE, /* Prompl */ 00159 TRUE, /* IfMinMax */ 00160 FALSE, /* Run_call_info */ 00161 TRUE, /* Shackle */ 00162 FALSE, /* Cross_loop */ 00163 FALSE, /* IPA_Enabled */ 00164 100, /* Num_Iters */ 00165 1, /* Pure_Level */ 00166 5, /* Small_trip_count */ 00167 (UINT32)-1, /* Local_pad_size */ 00168 0, /* Full_unrolling */ 00169 { 0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0 } /* buffer[16] */ 00170 }; 00171 00172 LNO_FLAGS Initial_LNO = { 00173 NULL, /* next */ 00174 &Mhd_Options, /* _mhd */ 00175 00176 TRUE, /* Aequiv */ 00177 FALSE, /* Automatic data distribution */ 00178 1, /* Run_AP */ 00179 FALSE, /* Ill_xform_directives */ 00180 TRUE, /* Backward_substitution */ 00181 FALSE, /* Blind_loop_reversal */ 00182 TRUE, /* Blocking */ 00183 0, /* Blocking_Size */ 00184 TRUE, /* Cache_model_edge_effects */ 00185 TRUE, /* Coupled_opts */ 00186 TRUE, /* Cse */ 00187 FALSE, /* Fancy_tile */ 00188 FALSE, /* Run_fiz_fuse */ 00189 1, /* Fission */ 00190 0, /* Fission_inner_register_limit */ 00191 TRUE, /* Forward_substitution */ 00192 1, /* Fusion */ 00193 5, /* Fusion_peeling_limit */ 00194 1, /* Gather_Scatter */ 00195 0xfffe, /* Graph_capacity */ 00196 TRUE, /* Hoist_messy_bounds */ 00197 FALSE, /* Ignore_pragmas */ 00198 TRUE, /* Interchange */ 00199 FALSE, FALSE, /* Run_lego */ 00200 TRUE, /* Run_lego_localizer */ 00201 TRUE, /* Loop_finalization */ 00202 8, /* Max_do_loop_depth_strict */ 00203 FALSE, /* Mem_sim */ 00204 TRUE, /* Minvar */ 00205 1, /* Opt */ 00206 2, /* Cache_model */ 00207 TRUE, /* Run_outer */ 00208 0, /* Outer_unroll */ 00209 TRUE, /* Outer_unroll_deep */ 00210 6, /* Outer_unroll_min_for_further_unroll */ 00211 0, /* Outer_unroll_max */ 00212 FALSE, /* Outer_unroll_model_only */ 00213 0, /* Outer_unroll_prod_max */ 00214 TRUE, /* Outer_unroll_unity */ 00215 2, /* ou_aggre */ 00216 0, /* Run_p3 */ 00217 TRUE, /* Pseudo_lower */ 00218 0, FALSE, /* Run_prefetch */ 00219 2, /* Prefetch_ahead */ 00220 2, /* Prefetch_iters_ahead */ 00221 2, /* Prefetch_cache_factor */ 00222 FALSE, /* Prefetch_indirect */ 00223 FALSE, FALSE, /* Run_prefetch_manual */ 00224 TRUE, /* Power_of_two_hack */ 00225 TRUE, /* Sclrze */ 00226 0, /* SE_tile_size */ 00227 1, /* Split_tiles */ 00228 64, /* Split_tiles_size */ 00229 FALSE, /* Run_test */ 00230 FALSE, /* Test_dump */ 00231 TRUE, /* Trapezoidal_outer_unroll */ 00232 FALSE, /* Use_malloc */ 00233 TRUE, /* Use_parm */ 00234 FALSE, /* Verbose */ 00235 TRUE, /* Version_mp_loops */ 00236 TRUE, /* Run_vintr */ 00237 TRUE, /* Run_oinvar */ 00238 1, /* Run_doacross */ 00239 0, /* Preferred_doacross_tile_size */ 00240 2600, /* Parallel_overhead */ 00241 FALSE, /* Prompl */ 00242 TRUE, /* IfMinMax */ 00243 FALSE, /* Run_call_info */ 00244 TRUE, /* Shackle */ 00245 FALSE, /* Cross_loop */ 00246 FALSE, /* IPA array section information available */ 00247 100, /* Num_Iters */ 00248 1, /* Pure_Level */ 00249 5, /* Small_trip_count */ 00250 (UINT32)-1, /* Local_pad_size */ 00251 0, /* Full_unrolling */ 00252 { 0,0,0,0, 0,0,0,0, 0,0,0,0, 0,0,0,0 } /* buffer[16] */ 00253 }; 00254 00255 LNO_FLAGS *Current_LNO = &Initial_LNO; 00256 MHD *Current_MHD = &Mhd_Options; 00257 00258 /* ==================================================================== 00259 * Descriptor for the -LNO option group. 00260 * ==================================================================== 00261 */ 00262 00263 /* Macros for defining primary -LNO group options: */ 00264 #define IL Initial_LNO 00265 #define LNOPT_U32(name,abbr,def,lb,ub,var) \ 00266 { OVK_UINT32, OV_VISIBLE, TRUE, name, abbr, def, lb, ub, &IL.var, NULL } 00267 #define LNOPT_U32_SET(name,abbr,def,lb,ub,var,var_set) \ 00268 { OVK_UINT32, OV_VISIBLE, TRUE, name, abbr, def, lb, ub, &IL.var, &IL.var_set } 00269 #define LNOPT_BOOL(name,abbr,var) \ 00270 { OVK_BOOL, OV_VISIBLE, TRUE, name, abbr, 1, 0, 1, &IL.var, NULL } 00271 #define LNOPT_BOOL_SET(name,abbr,var,var_set) \ 00272 { OVK_BOOL, OV_VISIBLE, TRUE, name, abbr, 1, 0, 1, &IL.var, &IL.var_set } 00273 00274 /* Macros for defining memory hierarchy -LNO group options: */ 00275 #define MHO Mhd_Options 00276 #define MHOPT_I32(name,abbr,def,lb,ub,var) \ 00277 { OVK_INT32, OV_VISIBLE, TRUE, name, abbr, def, lb, ub, &MHO.var, NULL } 00278 #define MHOPT_I32_SET(name,abbr,def,lb,ub,var,var_set) \ 00279 { OVK_INT32, OV_VISIBLE, TRUE, name, abbr, def, lb, ub, &MHO.var, &MHO.var_set } 00280 #define MHOPT_I32_DUP(name,def,lb,ub,var) \ 00281 MHOPT_I32 ( name "1", NULL, def, lb, ub, L[0].var ), \ 00282 MHOPT_I32 ( name "2", NULL, def, lb, ub, L[1].var ), \ 00283 MHOPT_I32 ( name "3", NULL, def, lb, ub, L[2].var ), \ 00284 MHOPT_I32 ( name "4", NULL, def, lb, ub, L[3].var ) 00285 #define MHOPT_I32_SET_DUP(name,def,lb,ub,var,vset) \ 00286 MHOPT_I32_SET ( name "1", NULL, def, lb, ub, L[0].var, L[0].vset ), \ 00287 MHOPT_I32_SET ( name "2", NULL, def, lb, ub, L[1].var, L[1].vset ), \ 00288 MHOPT_I32_SET ( name "3", NULL, def, lb, ub, L[2].var, L[2].vset ), \ 00289 MHOPT_I32_SET ( name "4", NULL, def, lb, ub, L[3].var, L[3].vset ) 00290 #define MHOPT_NAME(name,abbr,var) \ 00291 { OVK_NAME, OV_VISIBLE, TRUE, name, abbr, 0, 0, 0, &MHO.var, NULL } 00292 #define MHOPT_NAME_DUP(name,var) \ 00293 MHOPT_NAME ( name "1", NULL, L[0].var ), \ 00294 MHOPT_NAME ( name "2", NULL, L[1].var ), \ 00295 MHOPT_NAME ( name "3", NULL, L[2].var ), \ 00296 MHOPT_NAME ( name "4", NULL, L[3].var ) 00297 #define MHOPT_BOOL(name,abbr,var) \ 00298 { OVK_BOOL, OV_VISIBLE, TRUE, name, abbr, 0, 0, 1, &MHO.var, NULL } 00299 #define MHOPT_BOOL_SET(name,abbr,var,var_set) \ 00300 { OVK_BOOL, OV_VISIBLE, TRUE, name, abbr, 0, 0, 1, &MHO.var, &MHO.var_set } 00301 #define MHOPT_BOOL_DUP(name,var) \ 00302 MHOPT_BOOL ( name "1", NULL, L[0].var ), \ 00303 MHOPT_BOOL ( name "2", NULL, L[1].var ), \ 00304 MHOPT_BOOL ( name "3", NULL, L[2].var ), \ 00305 MHOPT_BOOL ( name "4", NULL, L[3].var ) 00306 #define MHOPT_BOOL_SET_DUP(name,var,vset) \ 00307 MHOPT_BOOL_SET ( name "1", NULL, L[0].var, L[0].vset ), \ 00308 MHOPT_BOOL_SET ( name "2", NULL, L[1].var, L[1].vset ), \ 00309 MHOPT_BOOL_SET ( name "3", NULL, L[2].var, L[2].vset ), \ 00310 MHOPT_BOOL_SET ( name "4", NULL, L[3].var, L[3].vset ) 00311 00312 static OPTION_DESC Options_LNO[] = { 00313 00314 LNOPT_BOOL ( "aequiv", NULL, Aequiv ), 00315 LNOPT_U32 ( "ap", NULL, 1,0,2, Run_AP ), 00316 LNOPT_BOOL ( "apply_illegal_xform_directives", "", Ill_xform_directives ), 00317 MHOPT_I32_DUP ( "associativity", 0,0,8192, Associativity ), 00318 MHOPT_I32_DUP ( "assoc", 0,0,8192, Associativity ), 00319 LNOPT_BOOL ( "auto_dist", NULL, Autodist ), 00320 LNOPT_BOOL ( "ad", NULL, Autodist ), 00321 LNOPT_BOOL ( "backward_substitution", NULL, Backward_substitution ), 00322 LNOPT_BOOL ( "blind_loop_reversal", NULL, Blind_loop_reversal ), 00323 LNOPT_BOOL ( "blocking", NULL, Blocking ), 00324 LNOPT_U32 ( "blocking_size", NULL, 0,0,99999, Blocking_Size ), 00325 LNOPT_BOOL ( "cache_edge_effects", NULL, Cache_model_edge_effects ), 00326 MHOPT_NAME_DUP ( "cache_size", CS_string ), 00327 MHOPT_NAME_DUP ( "cs", CS_string ), 00328 MHOPT_I32_SET_DUP ( "clean_miss_penalty", 0,0,1000000000, 00329 Clean_Miss_Penalty, CMP_Set ), 00330 MHOPT_I32_SET_DUP ( "cmp", 0,0,1000000000, 00331 Clean_Miss_Penalty, CMP_Set ), 00332 LNOPT_BOOL ( "cse", NULL, Cse ), 00333 MHOPT_I32_SET_DUP ( "dirty_miss_penalty", 0,0,1000000000, 00334 Dirty_Miss_Penalty, DMP_Set ), 00335 MHOPT_I32_SET_DUP ( "dmp", 0,0,1000000000, 00336 Dirty_Miss_Penalty, DMP_Set ), 00337 LNOPT_BOOL ( "fancy_tile", NULL, Fancy_tile ), 00338 LNOPT_BOOL ( "ff", NULL, Run_fiz_fuse ), 00339 LNOPT_U32 ( "fission", "fis", 1,0,2, Fission ), 00340 LNOPT_U32 ( "fission_inner_register_limit", NULL, 32,0,99999, 00341 Fission_inner_register_limit ), 00342 LNOPT_BOOL ( "forward_substitution", NULL, Forward_substitution ), 00343 LNOPT_U32 ( "fusion", "fus", 1,0,2, Fusion ), 00344 LNOPT_U32 ( "fusion_peeling_limit", NULL, 5,0,99999, 00345 Fusion_peeling_limit ), 00346 LNOPT_U32 ( "gather_scatter", "gath", 1,0,100, Gather_Scatter ), 00347 LNOPT_U32 ( "gc", NULL, 0xfffe, 0, 0xfffe, 00348 Graph_capacity ), 00349 LNOPT_BOOL ( "hmb", NULL, Hoist_messy_bounds ), 00350 LNOPT_BOOL ( "ignore_pragmas", NULL, Ignore_pragmas ), 00351 LNOPT_BOOL ( "interchange", NULL, Interchange ), 00352 MHOPT_BOOL_SET_DUP ( "is_memory_level", 00353 Is_Mem_Level, Is_Mem_Level_Set ), 00354 MHOPT_BOOL_SET_DUP ( "is_mem", Is_Mem_Level, Is_Mem_Level_Set ), 00355 LNOPT_BOOL_SET ( "lego", NULL, Run_lego, Run_lego_set ), 00356 LNOPT_BOOL ( "lego_local", NULL, Run_lego_localizer ), 00357 MHOPT_I32_DUP ( "line_size", 0,0,8192, Line_Size ), 00358 MHOPT_I32_DUP ( "ls", 0,0,8192, Line_Size ), 00359 MHOPT_I32 ( "loop_overhead_base", NULL, 1,0,1000, Loop_Overhead_Base ), 00360 MHOPT_I32 ( "lob", NULL, 1,0,1000, Loop_Overhead_Base ), 00361 MHOPT_I32 ( "loop_overhead_memref", NULL, 1,0,1000, 00362 Loop_Overhead_Memref ), 00363 MHOPT_I32 ( "lom", NULL, 1,0,1000, 00364 Loop_Overhead_Memref ), 00365 LNOPT_U32 ( "max_depth", NULL, 10,0,16, 00366 Max_do_loop_depth_strict ), 00367 LNOPT_BOOL ( "mem_sim", NULL, Mem_sim ), 00368 LNOPT_BOOL ( "minvariant", "", Minvar ), 00369 MHOPT_I32_SET_DUP ( "miss_penalty", 0,0,1000000000, 00370 Miss_Penalty, Miss_Penalty_Set ), 00371 MHOPT_I32_SET_DUP ( "mp", 0,0,1000000000, 00372 Miss_Penalty, Miss_Penalty_Set ), 00373 MHOPT_BOOL ( "non_blocking_loads", NULL, Non_Blocking_Loads ), 00374 MHOPT_BOOL ( "nbl", NULL, Non_Blocking_Loads ), 00375 LNOPT_U32 ( "opt", NULL, 1,0,1, Opt ), 00376 LNOPT_U32 ( "optimize_cache", NULL, 2,0,2, Cache_model ), 00377 LNOPT_U32 ( "optc", NULL, 2,0,2, Cache_model ), 00378 LNOPT_BOOL ( "outer", NULL, Run_outer ), 00379 LNOPT_U32 ( "outer_unroll", NULL, 0,1,32, Outer_unroll ), 00380 LNOPT_U32 ( "ou", NULL, 0,1,32, Outer_unroll ), 00381 LNOPT_BOOL ( "outer_unroll_deep", NULL, Outer_unroll_deep ), 00382 LNOPT_BOOL ( "ou_deep", NULL, Outer_unroll_deep ), 00383 LNOPT_U32 ( "outer_unroll_further", NULL, 6,0,99999, 00384 Outer_unroll_min_for_further_unroll ), 00385 LNOPT_U32 ( "ou_further", NULL, 6,0,99999, 00386 Outer_unroll_min_for_further_unroll ), 00387 LNOPT_U32 ( "outer_unroll_max", NULL, 0,1,32, Outer_unroll_max ), 00388 LNOPT_U32 ( "ou_max", NULL, 0,1,32, Outer_unroll_max ), 00389 LNOPT_BOOL ( "outer_unroll_model_only", NULL, Outer_unroll_model_only ), 00390 LNOPT_BOOL ( "ou_model_only", NULL, Outer_unroll_model_only ), 00391 LNOPT_U32 ( "outer_unroll_prod_max", NULL, 0,1,64, Outer_unroll_prod_max ), 00392 LNOPT_U32 ( "ou_prod_max", NULL, 0,1,64, Outer_unroll_prod_max ), 00393 LNOPT_BOOL ( "outer_unroll_unity", NULL, Outer_unroll_unity ), 00394 LNOPT_BOOL ( "ou_unity", NULL, Outer_unroll_unity ), 00395 LNOPT_U32 ( "ou_aggre", NULL, 2,0,3, Outer_unroll_aggre ), 00396 LNOPT_U32 ( "p3", NULL, 0,0,2, Run_p3 ), 00397 MHOPT_I32_DUP ( "page_size", 0,0,8192, Page_Size ), 00398 MHOPT_I32_DUP ( "ps", 0,0,8192, Page_Size ), 00399 MHOPT_I32_DUP ( "pct_xwrites_nonhidable", 0,0,100, 00400 Pct_Excess_Writes_Nonhidable ), 00401 MHOPT_I32_DUP ( "xwn", 0,0,100, 00402 Pct_Excess_Writes_Nonhidable ), 00403 LNOPT_BOOL ( "plower", NULL, Pseudo_lower ), 00404 LNOPT_BOOL ( "plower_mp", NULL, Pseudo_lower ), 00405 LNOPT_U32_SET ( "prefetch", "pref", 0,0,2, Run_prefetch, 00406 Run_prefetch_set ), 00407 LNOPT_U32 ( "prefetch_ahead", NULL, 2,0,50, Prefetch_ahead ), 00408 LNOPT_U32 ( "pf_ahead", NULL, 2,0,50, Prefetch_ahead ), 00409 LNOPT_U32 ( "prefetch_iters_ahead", NULL, 2,0,50, Prefetch_iters_ahead ), 00410 LNOPT_U32 ( "pf_it_ahead", NULL, 2,0,50, Prefetch_iters_ahead ), 00411 LNOPT_U32 ( "prefetch_cache_factor", NULL, 1,1,50, Prefetch_cache_factor), 00412 LNOPT_U32 ( "pf_cf", NULL, 1,1,50, Prefetch_cache_factor), 00413 LNOPT_BOOL ( "prefetch_indirect", "", Prefetch_indirect ), 00414 MHOPT_BOOL_DUP ( "prefetch_level", Prefetch_Level ), 00415 MHOPT_BOOL_DUP ( "pf", Prefetch_Level ), 00416 LNOPT_BOOL_SET ( "prefetch_manual", "", Run_prefetch_manual, 00417 Run_prefetch_manual_set ), 00418 LNOPT_BOOL_SET ( "pf_manual", NULL, Run_prefetch_manual, 00419 Run_prefetch_manual_set ), 00420 LNOPT_BOOL ( "pwr2", NULL, Power_of_two_hack ), 00421 LNOPT_BOOL ( "sclrze", NULL, Sclrze ), 00422 LNOPT_U32 ( "se_tile_size", NULL, 0,0,99999, SE_tile_size ), 00423 LNOPT_U32 ( "split_tile", NULL, 1,0,2, Split_tiles ), 00424 LNOPT_U32 ( "split_tile_size", NULL, 0,0,99999, Split_tiles_size ), 00425 LNOPT_BOOL ( "test", NULL, Run_test ), 00426 LNOPT_BOOL ( "testdump", NULL, Test_dump ), 00427 MHOPT_I32_SET_DUP ( "tlb_clean_miss_penalty", 0,0,1000000000, 00428 TLB_Clean_Miss_Penalty, TLB_CMP_Set ), 00429 MHOPT_I32_SET_DUP ( "tlbcmp", 0,0,1000000000, 00430 TLB_Clean_Miss_Penalty, TLB_CMP_Set ), 00431 MHOPT_I32_SET_DUP ( "tlb_dirty_miss_penalty", 0,0,1000000000, 00432 TLB_Dirty_Miss_Penalty, TLB_DMP_Set ), 00433 MHOPT_I32_SET_DUP ( "tlbdmp", 0,0,1000000000, 00434 TLB_Dirty_Miss_Penalty, TLB_DMP_Set ), 00435 MHOPT_I32_DUP ( "tlb_entries", 0,0,8192, TLB_Entries ), 00436 MHOPT_I32_DUP ( "tlb", 0,0,8192, TLB_Entries ), 00437 MHOPT_I32_SET_DUP ( "tlb_miss_penalty", 0,0,1000000000, 00438 TLB_Miss_Penalty, TLB_MP_Set ), 00439 MHOPT_I32_SET_DUP ( "tlbmp", 0,0,1000000000, 00440 TLB_Miss_Penalty, TLB_MP_Set ), 00441 MHOPT_BOOL ( "tlb_noblocking_model", NULL, TLB_NoBlocking_Model ), 00442 MHOPT_I32 ( "tlb_trustworthiness", NULL, 1,0,1000, 00443 TLB_Trustworthiness ), 00444 LNOPT_BOOL ( "trapezoidal_outer_unroll", NULL, 00445 Trapezoidal_outer_unroll ), 00446 LNOPT_BOOL ( "use_malloc", NULL, Use_malloc ), 00447 LNOPT_BOOL ( "use_parm", NULL, Use_parm ), 00448 LNOPT_BOOL ( "version_mp_loops", NULL, Version_mp_loops ), 00449 LNOPT_BOOL ( "vintr", NULL, Run_vintr ), 00450 LNOPT_BOOL ( "oinvar", NULL, Run_oinvar ), 00451 LNOPT_U32 ( "doacross", NULL, 1,0,4,Run_doacross), 00452 LNOPT_U32 ( "preferred_doacross_tile_size", 00453 NULL, 0,0,999999, 00454 Preferred_doacross_tile_size ), 00455 LNOPT_U32 ( "parallel_overhead", NULL, 2600,0,0x7fffffff, 00456 Parallel_overhead), 00457 LNOPT_BOOL ( "prompl", NULL, Prompl ), 00458 LNOPT_BOOL ( "ifminmax", NULL, IfMinMax ), 00459 LNOPT_BOOL ( "call_info", NULL, Run_call_info ), 00460 LNOPT_BOOL ( "loop_finalize", NULL, Loop_finalization), 00461 LNOPT_BOOL ( "shackle", NULL, Shackle), 00462 LNOPT_BOOL ( "cross_loop", NULL, Cross_loop), 00463 LNOPT_BOOL ( "ipa", NULL, IPA_Enabled), 00464 LNOPT_U32 ( "num_iters", NULL, 100,0,UINT32_MAX,Num_Iters), 00465 LNOPT_U32 ( "pure", NULL, 1,0,2,Pure_Level), 00466 LNOPT_U32 ( "small_trip_count", NULL, 5,0,10,Small_trip_count), 00467 LNOPT_U32 ( "local_pad_size", NULL, 0,0,1000,Local_pad_size), 00468 LNOPT_U32 ( "full_unroll", "fu", 0,0,100,Full_unrolling), 00469 00470 { OVK_COUNT } /* List terminator -- must be last */ 00471 }; 00472 00473 #undef IL 00474 00475 /* ==================================================================== 00476 * 00477 * LNO_Init_Config 00478 * 00479 * Initialize the current top of stack to defaults, by copying the 00480 * Default_LNO struct to it. Be careful not to lose the next 00481 * pointer. 00482 * 00483 * WARNING: We are always picking up the default Mhd_Options struct 00484 * here. 00485 * 00486 * ==================================================================== 00487 */ 00488 00489 void 00490 LNO_Init_Config ( void ) 00491 { 00492 LNO_FLAGS *next = LNO_FLAGS_next(Current_LNO); 00493 00494 *Current_LNO = Default_LNO; 00495 LNO_FLAGS_next(Current_LNO) = next; 00496 LNO_FLAGS_mhd (Current_LNO) = &Mhd_Options; 00497 } 00498 00499 /* ==================================================================== 00500 * 00501 * LNO_Push_Config 00502 * 00503 * Push a new struct on top of stack, either a copy of the current 00504 * TOS, or the defaults. 00505 * 00506 * TODO: To make the option group printing functionality work once the 00507 * configuration structs start changing, this function must also update 00508 * the addresses in the group descriptor to point to the current 00509 * top-of-stack structure. 00510 * 00511 * ==================================================================== 00512 */ 00513 00514 void 00515 LNO_Push_Config ( BOOL use_default ) 00516 { 00517 LNO_FLAGS *new_flags = 00518 (LNO_FLAGS *) malloc ( sizeof(LNO_FLAGS) ); 00519 00520 if ( new_flags == NULL ) { 00521 ErrMsg ( EC_No_Mem, "LNO_Push" ); 00522 } 00523 00524 *new_flags = use_default ? Default_LNO : *Current_LNO; 00525 LNO_FLAGS_next(new_flags) = Current_LNO; 00526 LNO_FLAGS_mhd (new_flags) = &Mhd_Options; 00527 Current_LNO = new_flags; 00528 } 00529 00530 00531 00532 /* ==================================================================== 00533 * 00534 * LNO_Pop_Config 00535 * 00536 * Pop a struct from top of stack and return TRUE if the old TOS was 00537 * not the original TOS, or do nothing and return FALSE. 00538 * 00539 * TODO: To make the option group printing functionality work once the 00540 * configuration structs start changing, this function must also update 00541 * the addresses in the group descriptor to point to the current 00542 * top-of-stack structure. 00543 * 00544 * ==================================================================== 00545 */ 00546 00547 BOOL 00548 LNO_Pop_Config ( void ) 00549 { 00550 if ( Current_LNO->next == NULL ) { 00551 /* This is the bottom of the stack: */ 00552 return FALSE; 00553 } else { 00554 /* Deallocate the top element and pop it: */ 00555 LNO_FLAGS *new_flags = LNO_FLAGS_next(Current_LNO); 00556 free ( Current_LNO ); 00557 Current_LNO = new_flags; 00558 return TRUE; 00559 } 00560 } 00561 00562 /* ==================================================================== 00563 * 00564 * LNO_Configure 00565 * 00566 * Configure the current top of stack struct. 00567 * 00568 * ==================================================================== 00569 */ 00570 00571 void 00572 LNO_Configure ( void ) 00573 { 00574 INT i; 00575 00576 #ifdef COMPILER_LICENSING 00577 /* check licensing for automatic parallelization */ 00578 if (Run_autopar) { 00579 get_license("auto_pp", FALSE, TRUE); /* soft licensing == FALSE; tier licensing == TRUE */ 00580 } 00581 #endif 00582 00583 /* This option is actually a trace flag: */ 00584 LNO_Verbose = Get_Trace ( TP_LNOPT, 0x0004 ); 00585 00586 /* Unrolling: of LNO_Outer_Unroll, LNO_Outer_Unroll_Max, and 00587 * LNO_Outer_Unroll_Prod_Max, only the first OR either or both of the 00588 * last two may be non-zero (i.e. meaningful). 00589 */ 00590 if ( LNO_Outer_Unroll != 0 && LNO_Outer_Unroll_Max != 0 ) { 00591 ErrMsg ( EC_Opt_Conflict, "-LNO:outer_unroll", 00592 "-LNO:outer_unroll_max", "-LNO:outer_unroll" ); 00593 LNO_Outer_Unroll_Max = 0; 00594 } 00595 if ( LNO_Outer_Unroll != 0 && LNO_Outer_Unroll_Prod_Max != 0 ) { 00596 ErrMsg ( EC_Opt_Conflict, "-LNO:outer_unroll", 00597 "-LNO:outer_unroll_prod_max", "-LNO:outer_unroll" ); 00598 LNO_Outer_Unroll_Prod_Max = 0; 00599 } 00600 if ( LNO_Outer_Unroll == 0 00601 && LNO_Outer_Unroll_Max == 0 00602 && LNO_Outer_Unroll_Prod_Max == 0 ) 00603 { 00604 LNO_Outer_Unroll_Prod_Max = DEFAULT_UNROLL_PROD_MAX; 00605 } 00606 if ( LNO_Outer_Unroll == 0 && LNO_Outer_Unroll_Max == 0 ) { 00607 LNO_Outer_Unroll_Max = DEFAULT_UNROLL_MAX; 00608 } 00609 00610 /* Cache parameters */ 00611 00612 for ( i = 0; i < 4; i++ ) { 00613 if ( Mhd_Options.L[i].CS_string != NULL 00614 && ! Atoi_KMG ( Mhd_Options.L[i].CS_string, 00615 &Mhd_Options.L[i].Size, TRUE ) ) 00616 { 00617 char buffer[256]; 00618 00619 sprintf ( buffer, "-LNO:cache_size%d=%s", 00620 i, Mhd_Options.L[i].CS_string ); 00621 ErrMsg ( EC_Inv_Opt_Val, buffer ); 00622 Mhd_Options.L[i].Size = 0; 00623 } 00624 00625 if ( Mhd_Options.L[i].Is_Mem_Level == 1) 00626 Mhd_Options.L[i].Type = MHD_TYPE_MEM; 00627 else if ( Mhd_Options.L[i].Is_Mem_Level == 0 ) 00628 Mhd_Options.L[i].Type = MHD_TYPE_CACHE; 00629 00630 if ( Mhd_Options.L[i].Miss_Penalty_Set ) { 00631 if ( Mhd_Options.L[i].CMP_Set ) { 00632 ErrMsg ( EC_Opt_Conflict, "-LNO:miss_penalty", 00633 "-LNO:clean_miss_penalty", "-LNO:miss_penalty" ); 00634 } 00635 if ( Mhd_Options.L[i].DMP_Set ) { 00636 ErrMsg ( EC_Opt_Conflict, "-LNO:miss_penalty", 00637 "-LNO:dirty_miss_penalty", "-LNO:miss_penalty" ); 00638 } 00639 Mhd_Options.L[i].Clean_Miss_Penalty = Mhd_Options.L[i].Miss_Penalty; 00640 Mhd_Options.L[i].Dirty_Miss_Penalty = Mhd_Options.L[i].Miss_Penalty; 00641 } 00642 if (Mhd_Options.L[i].TLB_Miss_Penalty != -1) { 00643 if ( Mhd_Options.L[i].TLB_CMP_Set ) { 00644 ErrMsg ( EC_Opt_Conflict, "-LNO:tlb_miss_penalty", 00645 "-LNO:tlb_clean_miss_penalty", "-LNO:tlb_miss_penalty" ); 00646 } 00647 if ( Mhd_Options.L[i].TLB_DMP_Set ) { 00648 ErrMsg ( EC_Opt_Conflict, "-LNO:tlb_miss_penalty", 00649 "-LNO:tlb_dirty_miss_penalty", "-LNO:tlb_miss_penalty" ); 00650 } 00651 Mhd_Options.L[i].TLB_Clean_Miss_Penalty = 00652 Mhd_Options.L[i].TLB_Miss_Penalty; 00653 Mhd_Options.L[i].TLB_Dirty_Miss_Penalty = 00654 Mhd_Options.L[i].TLB_Miss_Penalty; 00655 } 00656 } 00657 } 00658
1.7.6.1 
