17. Passes and Files of the Compiler

The overall control structure of the compiler is in `toplev.c'. This file is responsible for initialization, decoding arguments, opening and closing files, and sequencing the passes.

The parsing pass is invoked only once, to parse the entire input. A high level tree representation is then generated from the input, one function at a time. This tree code is then transformed into RTL intermediate code, and processed. The files involved in transforming the trees into RTL are `expr.c', `expmed.c', and `stmt.c'. The order of trees that are processed, is not necessarily the same order they are generated from the input, due to deferred inlining, and other considerations.

Each time the parsing pass reads a complete function definition or top-level declaration, it calls either the function rest_of_compilation, or the function rest_of_decl_compilation in `toplev.c', which are responsible for all further processing necessary, ending with output of the assembler language. All other compiler passes run, in sequence, within rest_of_compilation. When that function returns from compiling a function definition, the storage used for that function definition's compilation is entirely freed, unless it is an inline function, or was deferred for some reason (this can occur in templates, for example). (see section An Inline Function is As Fast As a Macro).

Here is a list of all the passes of the compiler and their source files. Also included is a description of where debugging dumps can be requested with `-d' options.

• Parsing. This pass reads the entire text of a function definition, constructing a high level tree representation. (Because of the semantic analysis that takes place during this pass, it does more than is formally considered to be parsing.)

  The tree representation does not entirely follow C syntax, because it is intended to support other languages as well.

  Language-specific data type analysis is also done in this pass, and every tree node that represents an expression has a data type attached. Variables are represented as declaration nodes.

  The language-independent source files for parsing are `stor-layout.c', `fold-const.c', and `tree.c'. There are also header files `tree.h' and `tree.def' which define the format of the tree representation.

  C Preprocessing, for language front ends, that want or require it, is performed by cpplib, which is covered in separate documentation. In particular, the internals are covered in See section `Cpplib internals' in Cpplib Internals.

  The source files to parse C are `c-aux-info.c', `c-convert.c', `c-decl.c', `c-errors.c', `c-lang.c', `c-parse.in', and `c-typeck.c', along with a header file `c-tree.h' and some files shared with Objective-C and C++.

  The source files for parsing C++ are in `cp/'. They are `parse.y', `class.c', `cvt.c', `decl.c', `decl2.c', `except.c', `expr.c', `init.c', `lex.c', `method.c', `ptree.c', `search.c', `spew.c', `semantics.c', `tree.c', `typeck2.c', and `typeck.c', along with header files `cp-tree.def', `cp-tree.h', and `decl.h'.

  The special source files for parsing Objective C are in `objc/'. They are `objc-parse.y', `objc-act.c', `objc-tree.def', and `objc-act.h'. Certain C-specific files are used for this as well.

  The files `c-common.c', `c-common.def', `c-dump.c', `c-format.c', `c-lex.c', `c-pragma.c', and `c-semantics.c', along with header files `c-common.h', `c-dump.h', `c-lex.h', and `c-pragma.h', are also used for all of the above languages.

• Tree optimization. This is the optimization of the tree representation, before converting into RTL code.

  Currently, the main optimization performed here is tree-based inlining. This is implemented for C++ in `cp/optimize.c'. Note that tree based inlining turns off rtx based inlining (since it's more powerful, it would be a waste of time to do rtx based inlining in addition). The C front end currently does not perform tree based inlining.

  Constant folding and some arithmetic simplifications are also done during this pass, on the tree representation. The routines that perform these tasks are located in `fold-const.c'.

• RTL generation. This is the conversion of syntax tree into RTL code.

  This is where the bulk of target-parameter-dependent code is found, since often it is necessary for strategies to apply only when certain standard kinds of instructions are available. The purpose of named instruction patterns is to provide this information to the RTL generation pass.

  Optimization is done in this pass for if-conditions that are comparisons, boolean operations or conditional expressions. Tail recursion is detected at this time also. Decisions are made about how best to arrange loops and how to output switch statements.

  The source files for RTL generation include `stmt.c', `calls.c', `expr.c', `explow.c', `expmed.c', `function.c', `optabs.c' and `emit-rtl.c'. Also, the file `insn-emit.c', generated from the machine description by the program genemit, is used in this pass. The header file `expr.h' is used for communication within this pass.

  The header files `insn-flags.h' and `insn-codes.h', generated from the machine description by the programs genflags and gencodes, tell this pass which standard names are available for use and which patterns correspond to them.

  Aside from debugging information output, none of the following passes refers to the tree structure representation of the function (only part of which is saved).

  The decision of whether the function can and should be expanded inline in its subsequent callers is made at the end of rtl generation. The function must meet certain criteria, currently related to the size of the function and the types and number of parameters it has. Note that this function may contain loops, recursive calls to itself (tail-recursive functions can be inlined!), gotos, in short, all constructs supported by GCC. The file `integrate.c' contains the code to save a function's rtl for later inlining and to inline that rtl when the function is called. The header file `integrate.h' is also used for this purpose.

  The option `-dr' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.rtl' to the input file name.

• Sibiling call optimization. This pass performs tail recursion elimination, and tail and sibling call optimizations. The purpose of these optimizations is to reduce the overhead of function calls, whenever possible.

  The source file of this pass is `sibcall.c'

  The option `-di' causes a debugging dump of the RTL code after this pass is run. This dump file's name is made by appending `.sibling' to the input file name.

• Jump optimization. This pass simplifies jumps to the following instruction, jumps across jumps, and jumps to jumps. It deletes unreferenced labels and unreachable code, except that unreachable code that contains a loop is not recognized as unreachable in this pass. (Such loops are deleted later in the basic block analysis.) It also converts some code originally written with jumps into sequences of instructions that directly set values from the results of comparisons, if the machine has such instructions.

  Jump optimization is performed two or three times. The first time is immediately following RTL generation. The second time is after CSE, but only if CSE says repeated jump optimization is needed. The last time is right before the final pass. That time, cross-jumping and deletion of no-op move instructions are done together with the optimizations described above.

  The source file of this pass is `jump.c'.

  The option `-dj' causes a debugging dump of the RTL code after this pass is run for the first time. This dump file's name is made by appending `.jump' to the input file name.

• Register scan. This pass finds the first and last use of each register, as a guide for common subexpression elimination. Its source is in `regclass.c'.

• Jump threading. This pass detects a condition jump that branches to an identical or inverse test. Such jumps can be `threaded' through the second conditional test. The source code for this pass is in `jump.c'. This optimization is only performed if `-fthread-jumps' is enabled.

• Common subexpression elimination. This pass also does constant propagation. Its source files are `cse.c', and `cselib.c'. If constant propagation causes conditional jumps to become unconditional or to become no-ops, jump optimization is run again when CSE is finished.

  The option `-ds' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.cse' to the input file name.

• Static Single Assignment (SSA) based optimization passes. The SSA conversion passes (to/from) are turned on by the `-fssa' option (it is also done automatically if you enable an SSA optimization pass). These passes utilize a form called Static Single Assignment. In SSA form, each variable (pseudo register) is only set once, giving you def-use and use-def chains for free, and enabling a lot more optimization passes to be run in linear time. Conversion to and from SSA form is handled by functions in `ssa.c'.

  The option `-de' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.ssa' to the input file name.

  • Dead Code Elimination. Turned on by the `-fdce' option. This pass performs elimination of code considered unnecessary because it is never executed. It operates in linear time.

    The option `-dX' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.dce' to the input file name.

• Global common subexpression elimination. This pass performs two different types of GCSE depending on whether you are optimizing for size or not (LCM based GCSE tends to increase code size for a gain in speed, while Morel-Renvoise based GCSE does not). When optimizing for size, GCSE is done using Morel-Renvoise Partial Redundancy Elimination, with the exception that it does not try to move invariants out of loops--that is left to the loop optimization pass. If MR PRE GCSE is done, code hoisting (aka unification) is also done, as well as load motion. If you are optimizing for speed, LCM (lazy code motion) based GCSE is done. LCM is based on the work of Knoop, Ruthing, and Steffen. LCM based GCSE also does loop invariant code motion. We also perform load and store motion when optimizing for speed. Regardless of which type of GCSE is used, the GCSE pass also performs global constant and copy propagation.

  The source file for this pass is `gcse.c', and the LCM routines are in `lcm.c'.

  The option `-dG' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.gcse' to the input file name.

• Loop optimization. This pass moves constant expressions out of loops, and optionally does strength-reduction and loop unrolling as well. Its source files are `loop.c' and `unroll.c', plus the header `loop.h' used for communication between them. Loop unrolling uses some functions in `integrate.c' and the header `integrate.h'. Loop dependency analysis routines are contained in `dependence.c'.

  The option `-dL' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.loop' to the input file name.

• If `-frerun-cse-after-loop' was enabled, a second common subexpression elimination pass is performed after the loop optimization pass. Jump threading is also done again at this time if it was specified.

  The option `-dt' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.cse2' to the input file name.

• Data flow analysis (`flow.c'). This pass divides the program into basic blocks (and in the process deletes unreachable loops); then it computes which pseudo-registers are live at each point in the program, and makes the first instruction that uses a value point at the instruction that computed the value.

  This pass also deletes computations whose results are never used, and combines memory references with add or subtract instructions to make autoincrement or autodecrement addressing.

  The option `-df' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.flow' to the input file name. If stupid register allocation is in use, this dump file reflects the full results of such allocation.

• Instruction combination (`combine.c'). This pass attempts to combine groups of two or three instructions that are related by data flow into single instructions. It combines the RTL expressions for the instructions by substitution, simplifies the result using algebra, and then attempts to match the result against the machine description.

  The option `-dc' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.combine' to the input file name.

• If-conversion is a transformation that transforms control dependencies into data dependencies (IE it transforms conditional code into a single control stream). It is implemented in the file `ifcvt.c'.

  The option `-dE' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.ce' to the input file name.

• Register movement (`regmove.c'). This pass looks for cases where matching constraints would force an instruction to need a reload, and this reload would be a register to register move. It then attempts to change the registers used by the instruction to avoid the move instruction.

  The option `-dN' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.regmove' to the input file name.

• Instruction scheduling (`sched.c'). This pass looks for instructions whose output will not be available by the time that it is used in subsequent instructions. (Memory loads and floating point instructions often have this behavior on RISC machines). It re-orders instructions within a basic block to try to separate the definition and use of items that otherwise would cause pipeline stalls.

  Instruction scheduling is performed twice. The first time is immediately after instruction combination and the second is immediately after reload.

  The option `-dS' causes a debugging dump of the RTL code after this pass is run for the first time. The dump file's name is made by appending `.sched' to the input file name.

• Register class preferencing. The RTL code is scanned to find out which register class is best for each pseudo register. The source file is `regclass.c'.

• Local register allocation (`local-alloc.c'). This pass allocates hard registers to pseudo registers that are used only within one basic block. Because the basic block is linear, it can use fast and powerful techniques to do a very good job.

  The option `-dl' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.lreg' to the input file name.

• Global register allocation (`global.c'). This pass allocates hard registers for the remaining pseudo registers (those whose life spans are not contained in one basic block).

• Reloading. This pass renumbers pseudo registers with the hardware registers numbers they were allocated. Pseudo registers that did not get hard registers are replaced with stack slots. Then it finds instructions that are invalid because a value has failed to end up in a register, or has ended up in a register of the wrong kind. It fixes up these instructions by reloading the problematical values temporarily into registers. Additional instructions are generated to do the copying.

  The reload pass also optionally eliminates the frame pointer and inserts instructions to save and restore call-clobbered registers around calls.

  Source files are `reload.c' and `reload1.c', plus the header `reload.h' used for communication between them.

  The option `-dg' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.greg' to the input file name.

• Instruction scheduling is repeated here to try to avoid pipeline stalls due to memory loads generated for spilled pseudo registers.

  The option `-dR' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.sched2' to the input file name.

• Basic block reordering. This pass implements profile guided code positioning. If profile information is not available, various types of static analysis are performed to make the predictions normally coming from the profile feedback (IE execution frequency, branch probability, etc). It is implemented in the file `bb-reorder.c', and the various prediction routines are in `predict.c'.

  The option `-dB' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.bbro' to the input file name.

• Jump optimization is repeated, this time including cross-jumping and deletion of no-op move instructions.

  The option `-dJ' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.jump2' to the input file name.

• Delayed branch scheduling. This optional pass attempts to find instructions that can go into the delay slots of other instructions, usually jumps and calls. The source file name is `reorg.c'.

  The option `-dd' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.dbr' to the input file name.

• Branch shortening. On many RISC machines, branch instructions have a limited range. Thus, longer sequences of instructions must be used for long branches. In this pass, the compiler figures out what how far each instruction will be from each other instruction, and therefore whether the usual instructions, or the longer sequences, must be used for each branch.

• Conversion from usage of some hard registers to usage of a register stack may be done at this point. Currently, this is supported only for the floating-point registers of the Intel 80387 coprocessor. The source file name is `reg-stack.c'.

  The options `-dk' causes a debugging dump of the RTL code after this pass. This dump file's name is made by appending `.stack' to the input file name.

• Final. This pass outputs the assembler code for the function. It is also responsible for identifying spurious test and compare instructions. Machine-specific peephole optimizations are performed at the same time. The function entry and exit sequences are generated directly as assembler code in this pass; they never exist as RTL.

  The source files are `final.c' plus `insn-output.c'; the latter is generated automatically from the machine description by the tool `genoutput'. The header file `conditions.h' is used for communication between these files.

• Debugging information output. This is run after final because it must output the stack slot offsets for pseudo registers that did not get hard registers. Source files are `dbxout.c' for DBX symbol table format, `sdbout.c' for SDB symbol table format, `dwarfout.c' for DWARF symbol table format, and the files `dwarf2out.c' and `dwarf2asm.c' for DWARF2 symbol table format.

Some additional files are used by all or many passes:

• Every pass uses `machmode.def' and `machmode.h' which define the machine modes.

• Several passes use `real.h', which defines the default representation of floating point constants and how to operate on them.

• All the passes that work with RTL use the header files `rtl.h' and `rtl.def', and subroutines in file `rtl.c'. The tools gen* also use these files to read and work with the machine description RTL.

• Several passes refer to the header file `insn-config.h' which contains a few parameters (C macro definitions) generated automatically from the machine description RTL by the tool genconfig.

• Several passes use the instruction recognizer, which consists of `recog.c' and `recog.h', plus the files `insn-recog.c' and `insn-extract.c' that are generated automatically from the machine description by the tools `genrecog' and `genextract'.

• Several passes use the header files `regs.h' which defines the information recorded about pseudo register usage, and `basic-block.h' which defines the information recorded about basic blocks.

• `hard-reg-set.h' defines the type HARD_REG_SET, a bit-vector with a bit for each hard register, and some macros to manipulate it. This type is just int if the machine has few enough hard registers; otherwise it is an array of int and some of the macros expand into loops.

• Several passes use instruction attributes. A definition of the attributes defined for a particular machine is in file `insn-attr.h', which is generated from the machine description by the program `genattr'. The file `insn-attrtab.c' contains subroutines to obtain the attribute values for insns. It is generated from the machine description by the program `genattrtab'.