GCC-3.4.6源代码学习笔记（48续）

4.2.10. 评估算术操作的代价

4.2.10.1. 创建伪函数上下文

回到backend_init，下来是init_dummy_function_start。它初始化了rtl的展开机制，因而我们可以进行一些简单序列的生成。

 

6518 void

6519 init_dummy_function_start (void)                                                                      in function.c

6520 {

6521   prepare_function_start (NULL);

6522 }

 

在这里调用了prepare_function_start来创建伪函数上下文cfun，因为下面的初始化函数需要产生rtl（在正式的代码编译中，只有完成代码解析及前端的优化后，才会生成rtl的表示形式，并交予后端）。

 

6482 static void

6483 prepare_function_start (tree fndecl)                                                            in function.c

6484 {

6485   if (fndecl && DECL_SAVED_INSNS (fndecl))

6486     cfun = DECL_SAVED_INSNS (fndecl);

6487   else

6488     allocate_struct_function (fndecl);

6489   init_emit ();

6490   init_varasm_status (cfun);

6491   init_expr ();

6492

6493   cse_not_expected = ! optimize;

6494

6495   /* Caller save not needed yet.  */

6496   caller_save_needed = 0;

6497

6498   /* We haven't done register allocation yet.  */

6499   reg_renumber = 0;

6500

6501   /* Indicate that we need to distinguish between the return value of the

6502     present function and the return value of a function being called.  */

6503   rtx_equal_function_value_matters = 1;

6504

6505   /* Indicate that we have not instantiated virtual registers yet.  */

6506   virtuals_instantiated = 0;

6507

6508   /* Indicate that we want CONCATs now.  */

6509   generating_concat_p = 1;

6510

6511   /* Indicate we have no need of a frame pointer yet.  */

6512   frame_pointer_needed = 0;

6513 }

 

结构体function保存了，描述当前被编译函数状态的所有重要的，全局及静态变量。

 

176    struct function GTY(())                                                                            in function.h

177    {

178      struct eh_status *eh;

179      struct stmt_status *stmt;

180      struct expr_status *expr;

181      struct emit_status *emit;

182      struct varasm_status *varasm;

183   

184      /* For function.c.  */

185   

186      /* Points to the FUNCTION_DECL of this function.  */

187      tree decl;

188   

189      /* Function containing this function, if any.  */

190      struct function *outer;

191   

192     /* Number of bytes of args popped by function being compiled on its return.

193        Zero if no bytes are to be popped.

194        May affect compilation of return insn or of function epilogue.  */

195      int pops_args;

196   

197      /* If function's args have a fixed size, this is that size, in bytes.

198        Otherwise, it is -1.

199        May affect compilation of return insn or of function epilogue.  */

200      int args_size;

201   

202      /* # bytes the prologue should push and pretend that the caller pushed them.

203         The prologue must do this, but only if parms can be passed in

204         registers.  */

205      int pretend_args_size;

206   

207      /* # of bytes of outgoing arguments. If ACCUMULATE_OUTGOING_ARGS is

208         defined, the needed space is pushed by the prologue.  */

209      int outgoing_args_size;

210   

211       /* This is the offset from the arg pointer to the place where the first

212        anonymous arg can be found, if there is one.  */

213      rtx arg_offset_rtx;

214   

215      /* Quantities of various kinds of registers

216        used for the current function's args.  */

217      CUMULATIVE_ARGS args_info;

218   

219      /* If nonzero, an RTL expression for the location at which the current

220        function returns its result. If the current function returns its

221        result in a register, current_function_return_rtx will always be

222        the hard register containing the result.  */

223      rtx return_rtx;

224   

225     /* The arg pointer hard register, or the pseudo into which it was copied.  */

226      rtx internal_arg_pointer;

227   

228      /* Language-specific reason why the current function cannot be made

229        inline.  */

230      const char *cannot_inline;

231   

232      /* Opaque pointer used by get_hard_reg_initial_val and

233        has_hard_reg_initial_val (see integrate.[hc]).  */

234      struct initial_value_struct *hard_reg_initial_vals;

235   

236      /* Number of function calls seen so far in current function.  */

237      int x_function_call_count;

238   

239     /* List (chain of TREE_LIST) of LABEL_DECLs for all nonlocal labels

240        (labels to which there can be nonlocal gotos from nested functions)

241        in this function.  */

242      tree x_nonlocal_labels;

243   

244      /* List (chain of EXPR_LIST) of stack slots that hold the current handlers

245        for nonlocal gotos. There is one for every nonlocal label in the

246        function; this list matches the one in nonlocal_labels.

247        Zero when function does not have nonlocal labels.  */

248      rtx x_nonlocal_goto_handler_slots;

249   

250      /* List (chain of EXPR_LIST) of labels heading the current handlers for

251        nonlocal gotos.  */

252      rtx x_nonlocal_goto_handler_labels;

253   

254      /* RTX for stack slot that holds the stack pointer value to restore

255        for a nonlocal goto.

256        Zero when function does not have nonlocal labels.  */

257      rtx x_nonlocal_goto_stack_level;

258   

259      /* Label that will go on parm cleanup code, if any.

260        Jumping to this label runs cleanup code for parameters, if

261        such code must be run. Following this code is the logical return

262        label.  */

263      rtx x_cleanup_label;

264   

265      /* Label that will go on function epilogue.

266        Jumping to this label serves as a "return" instruction

267        on machines which require execution of the epilogue on all returns.  */

268      rtx x_return_label;

269   

270      /* Label that will go on the end of function epilogue.

271        Jumping to this label serves as a "naked return" instruction

272        on machines which require execution of the epilogue on all returns.  */

273      rtx x_naked_return_label;

274   

275      /* Label and register for unswitching computed gotos.  */

276      rtx computed_goto_common_label;

277      rtx computed_goto_common_reg;

278   

279      /* List (chain of EXPR_LISTs) of pseudo-regs of SAVE_EXPRs.

280        So we can mark them all live at the end of the function, if nonopt.  */

281      rtx x_save_expr_regs;

282   

283      /* List (chain of EXPR_LISTs) of all stack slots in this function.

284        Made for the sake of unshare_all_rtl.  */

285      rtx x_stack_slot_list;

286   

287      /* Chain of all RTL_EXPRs that have insns in them.  */

288      tree x_rtl_expr_chain;

289   

290      /* Label to jump back to for tail recursion, or 0 if we have

291        not yet needed one for this function.  */

292      rtx x_tail_recursion_label;

293   

294      /* Place after which to insert the tail_recursion_label if we need one.  */

295      rtx x_tail_recursion_reentry;

296   

297      /* Location at which to save the argument pointer if it will need to be

298        referenced. There are two cases where this is done: if nonlocal gotos

299        exist, or if vars stored at an offset from the argument pointer will be

300        needed by inner routines.  */

301      rtx x_arg_pointer_save_area;

302   

303      /* If the function returns non-void, we will emit a clobber of the

304        return registers just in case the user fell off the end without

305        returning a proper value. This is that insn.  */

306      rtx x_clobber_return_insn;

307   

308      /* Offset to end of allocated area of stack frame.

309        If stack grows down, this is the address of the last stack slot allocated.

310        If stack grows up, this is the address for the next slot.  */

311       HOST_WIDE_INT x_frame_offset;

312   

313      /* List (chain of TREE_LISTs) of static chains for containing functions.

314        Each link has a FUNCTION_DECL in the TREE_PURPOSE and a reg rtx

315        in an RTL_EXPR in the TREE_VALUE.  */

316      tree x_context_display;

317   

318      /* List (chain of TREE_LISTs) of trampolines for nested functions.

319        The trampoline sets up the static chain and jumps to the function.

320        We supply the trampoline's address when the function's address is

321        requested.

322   

323        Each link has a FUNCTION_DECL in the TREE_PURPOSE and a reg rtx

324        in an RTL_EXPR in the TREE_VALUE.  */

325      tree x_trampoline_list;

326   

327      /* Insn after which register parms and SAVE_EXPRs are born, if nonopt.  */

328      rtx x_parm_birth_insn;

329   

330      /* Last insn of those whose job was to put parms into their nominal

331        homes.  */

332      rtx x_last_parm_insn;

333   

334      /* 1 + last pseudo register number possibly used for loading a copy

335        of a parameter of this function.  */

336      unsigned int x_max_parm_reg;

337   

338      /* Vector indexed by REGNO, containing location on stack in which

339        to put the parm which is nominally in pseudo register REGNO,

340        if we discover that that parm must go in the stack. The highest

341        element in this vector is one less than MAX_PARM_REG, above.  */

342      rtx * GTY ((length ("%h.x_max_parm_reg"))) x_parm_reg_stack_loc;

343   

344      /* List of all temporaries allocated, both available and in use.  */

345      struct temp_slot *x_temp_slots;

346   

347     /* Current nesting level for temporaries.  */

348      int x_temp_slot_level;

349   

350     /* Current nesting level for variables in a block.  */

351      int x_var_temp_slot_level;

352   

353      /* When temporaries are created by TARGET_EXPRs, they are created at

354        this level of temp_slot_level, so that they can remain allocated

355        until no longer needed. CLEANUP_POINT_EXPRs define the lifetime

356        of TARGET_EXPRs.  */

357      int x_target_temp_slot_level;

358   

359     /* This slot is initialized as 0 and is added to

360        during the nested function.  */

361      struct var_refs_queue *fixup_var_refs_queue;

362   

363      /* For integrate.c.  */

364      int inlinable;

365      int no_debugging_symbols;

366      rtvec original_arg_vector;

367      tree original_decl_initial;

368      /* Last insn of those whose job was to put parms into their nominal

369        homes.  */

370      rtx inl_last_parm_insn;

371      /* Highest label number in current function.  */

372      int inl_max_label_num;

373   

374      /* Function sequence number for profiling, debugging, etc.  */

375      int funcdef_no;

376   

377     /* For md files.  */

378   

379      /* tm.h can use this to store whatever it likes.  */

380      struct machine_function * GTY ((maybe_undef (""))) machine;

381      /* The largest alignment of slot allocated on the stack.  */

382      int stack_alignment_needed;

383      /* Preferred alignment of the end of stack frame.  */

384      int preferred_stack_boundary;

385      /* Set when the call to function itself has been emit.  */

386      bool recursive_call_emit;

387   

388     /* Language-specific code can use this to store whatever it likes.  */

389      struct language_function * language;

390   

391     /* For reorg.  */

392   

393      /* If some insns can be deferred to the delay slots of the epilogue, the

394        delay list for them is recorded here.  */

395      rtx epilogue_delay_list;

396   

397      /* How commonly executed the function is. Initialized during branch

398        probabilities pass.  */

399      enum function_frequency {

400        /* This function most likely won't be executed at all.

401          (set only when profile feedback is available).  */

402        FUNCTION_FREQUENCY_UNLIKELY_EXECUTED,

403        /* The default value.  */

404        FUNCTION_FREQUENCY_NORMAL,

405        /* Optimize this function hard

406          (set only when profile feedback is available).  */

407        FUNCTION_FREQUENCY_HOT

408      } function_frequency;

409   

410     /* Maximal number of entities in the single jumptable. Used to estimate

411         final flowgraph size.  */

412      int max_jumptable_ents;

413   

414      /* Collected bit flags.  */

415   

416      /* Nonzero if function being compiled needs to be given an address

417        where the value should be stored.  */

418      unsigned int returns_struct : 1;

419   

420      /* Nonzero if function being compiled needs to

421        return the address of where it has put a structure value.  */

422      unsigned int returns_pcc_struct : 1;

423   

424      /* Nonzero if the current function returns a pointer type.  */

425      unsigned int returns_pointer : 1;

426   

427      /* Nonzero if function being compiled needs to be passed a static chain.  */

428      unsigned int needs_context : 1;

429   

430      /* Nonzero if function being compiled can call setjmp.  */

431      unsigned int calls_setjmp : 1;

432   

433      /* Nonzero if function being compiled can call longjmp.  */

434      unsigned int calls_longjmp : 1;

435   

436      /* Nonzero if function being compiled can call alloca,

437        either as a subroutine or builtin.  */

438      unsigned int calls_alloca : 1;

439   

440      /* Nonzero if the function calls __builtin_eh_return.  */

441      unsigned int calls_eh_return : 1;

442   

443      /* Nonzero if the function calls __builtin_constant_p.  */

444      unsigned int calls_constant_p : 1;

445   

446      /* Nonzero if function being compiled receives nonlocal gotos

447        from nested functions.  */

448      unsigned int has_nonlocal_label : 1;

449   

450      /* Nonzero if function being compiled has nonlocal gotos to parent

451        function.  */

452      unsigned int has_nonlocal_goto : 1;

453   

454     /* Nonzero if function being compiled contains nested functions.  */

455      unsigned int contains_functions : 1;

456   

457      /* Nonzero if the function being compiled issues a computed jump.  */

458      unsigned int has_computed_jump : 1;

459   

460      /* Nonzero if the current function is a thunk, i.e., a lightweight

461        function implemented by the output_mi_thunk hook) that just

462        adjusts one of its arguments and forwards to another

463        function.  */

464      unsigned int is_thunk : 1;

465   

466      /* This bit is used by the exception handling logic. It is set if all

467        calls (if any) are sibling calls. Such functions do not have to

468        have EH tables generated, as they cannot throw. A call to such a

469        function, however, should be treated as throwing if any of its callees

470        can throw.  */

471      unsigned int all_throwers_are_sibcalls : 1;

472   

473      /* Nonzero if instrumentation calls for function entry and exit should be

474        generated.  */

475      unsigned int instrument_entry_exit : 1;

476   

477      /* Nonzero if profiling code should be generated.  */

478      unsigned int profile : 1;

479   

480      /* Nonzero if stack limit checking should be enabled in the current

481        function.  */

482      unsigned int limit_stack : 1;

483   

484      /* Nonzero if current function uses stdarg.h or equivalent.  */

485      unsigned int stdarg : 1;

486   

487      /* Nonzero if this function is being processed in function-at-a-time

488        mode. In other words, if all tree structure for this function,

489        including the BLOCK tree, is created before RTL generation

490        commences.  */

491      unsigned int x_whole_function_mode_p : 1;

492   

493      /* Nonzero if the back-end should not keep track of expressions that

494        determine the size of variable-sized objects. Normally, such

495        expressions are saved away, and then expanded when the next

496        function is started. For example, if a parameter has a

497        variable-sized type, then the size of the parameter is computed

498        when the function body is entered. However, some front-ends do

499        not desire this behavior.  */

500      unsigned int x_dont_save_pending_sizes_p : 1;

501   

502      /* Nonzero if the current function uses the constant pool.  */

503      unsigned int uses_const_pool : 1;

504   

505      /* Nonzero if the current function uses pic_offset_table_rtx.  */

506      unsigned int uses_pic_offset_table : 1;

507   

508      /* Nonzero if the current function needs an lsda for exception handling.  */

509      unsigned int uses_eh_lsda : 1;

510   

511      /* Nonzero if code to initialize arg_pointer_save_area has been emitted.  */

512      unsigned int arg_pointer_save_area_init : 1;

513   

514      /* Flag for use by ther rtl inliner, to tell if the function has been

515        processed at least once.  */

516      unsigned int rtl_inline_init : 1;

517   

518      /* Nonzero if the rtl inliner has saved the function for inlining.  */

519      unsigned int saved_for_inline : 1;

520    };

 

其结构中前6个成员是非常重要的部分。其中，eh_status控制了函数的异常处理；stmt_status记录了语句对栈，数据流的累积效应，这些效应将影响当前语句的翻译；expr_status则记录了函数被调用及退出的行为；emit_status控制相应的rtx形式指令的生成；varasm_status提供了记录函数中常量的方法。

 

6434 void

6435 allocate_struct_function (tree fndecl)                                                          in function.c

6436 {

6437   tree result;

6438

6439   cfun = ggc_alloc_cleared (sizeof (struct function));

6440

6441   max_parm_reg = LAST_VIRTUAL_REGISTER + 1;

6442

6443   cfun->stack_alignment_needed = STACK_BOUNDARY;

6444   cfun->preferred_stack_boundary = STACK_BOUNDARY;

6445

6446   current_function_funcdef_no = funcdef_no++;

6447

6448   cfun->function_frequency = FUNCTION_FREQUENCY_NORMAL;

6449

6450   init_stmt_for_function ();

6451   init_eh_for_function ();

6452

6453   (*lang_hooks.function.init) (cfun);

6454   if (init_machine_status)

6455     cfun->machine = (*init_machine_status) ();

6456

6457   if (fndecl == NULL)

6458     return;

  …

6477 }

 

在上面，current_function_funcdef_no，current_function_returns_pcc_struct，current_function_returns_struct，current_function_returns_pointer，max_parm_reg，current_function_needs_context都是用于从function结构体中选择特定成员的宏。例如。

 

558    #define current_function_funcdef_no (cfun->funcdef_no)                                   in function.h

567    #define max_parm_reg (cfun->x_max_parm_reg)

 

在6441行，对于x86机器，LAST_VIRTUAL_REGISTER被定义成FIRST_PSEUDO_REGISTER +4。其值是57，它定义了可被用于rtx形式代码的最大寄存器数目。

在6443及6444行，STACK_BOUNDARY定义了对栈的对齐边界要求。对于x86，这是32（4字节）。

在6446行，funcdef_no为用于概要分析（profiling）及调试等目的的标签赋予唯一的号码。在function 结构体相应的域则记录这个值。

接下来，在6450及6451行，函数init_stmt_for_function，init_eh_for_function仅是分配了结构体中的eh及stmt域，它们分别用于异常处理及语句。

在6453行，钩子lang_hooks.function.init对于C++语言是cxx_push_function_context。在当前上下文环境中，它仅为function对象分配language_function实例。结构体language_function将保存及恢复那些记录当前函数编译进度的变量。它被用于嵌套函数。

在6454行，对于x86机器，init_machine_status指向ix86_init_machine_status。它为function 对象创建了machine_function的实例，这个结构保存了与函数处理相关的机器信息。

在6457行，在这次的调用中，fndecl为null，函数在此返回。