![[BACK]](/cvsweb/icons/back.gif){kind=icon}
Return to PROJECTS CVS log ![[TXT]](/cvsweb/icons/text.gif){kind=icon}
![[DIR]](/cvsweb/icons/dir.gif){kind=icon}
Up to [GCC 2.x] / gcc|
Return to PROJECTS CVS log | Up to [GCC 2.x] / gcc |
1.1 root 1: 0. Improved efficiency.
2:
3: * Parse and output array initializers an element at a time, freeing
4: storage after each, instead of parsing the whole initializer first and
5: then outputting. This would reduce memory usage for large
6: initializers.
7:
8: * See if the techniques describe in Oct 1991 SIGPLAN Notices
9: (Frazer and Hanson) are applicable to GCC.
10:
11: 1. Better optimization.
12:
13: * Constants in unused inline functions
14:
15: It would be nice to delay output of string constants so that string
16: constants mentioned in unused inline functions are never generated.
17: Perhaps this would also take care of string constants in dead code.
18:
19: The difficulty is in finding a clean way for the RTL which refers
20: to the constant (currently, only by an assembler symbol name)
21: to point to the constant and cause it to be output.
22:
23: * More cse
24:
25: The techniques for doing full global cse are described in the red
26: dragon book, or (a different version) in Frederick Chow's thesis from
27: Stanford. It is likely to be slow and use a lot of memory, but it
28: might be worth offering as an additional option.
29:
30: It is probably possible to extend cse to a few very frequent cases
31: without so much expense.
32:
33: For example, it is not very hard to handle cse through if-then
34: statements with no else clauses. Here's how to do it. On reaching a
35: label, notice that the label's use-count is 1 and that the last
36: preceding jump jumps conditionally to this label. Now you know it
37: is a simple if-then statement. Remove from the hash table
38: all the expressions that were entered since that jump insn
39: and you can continue with cse.
40:
41: It is probably not hard to handle cse from the end of a loop
42: around to the beginning, and a few loops would be greatly sped
43: up by this.
44:
45: * Optimize a sequence of if statements whose conditions are exclusive.
46:
47: It is possible to optimize
48:
49: if (x == 1) ...;
50: if (x == 2) ...;
51: if (x == 3) ...;
52:
53: into
54:
55: if (x == 1) ...;
56: else if (x == 2) ...;
57: else if (x == 3) ...;
58:
59: provided that x is not altered by the contents of the if statements.
60:
61: It's not certain whether this is worth doing. Perhaps programmers
62: nearly always write the else's themselves, leaving few opportunities
63: to improve anything.
64:
65: * Un-cse.
66:
67: Perhaps we should have an un-cse step right after cse, which tries to
68: replace a reg with its value if the value can be substituted for the
69: reg everywhere, if that looks like an improvement. Which is if the
70: reg is used only a few times. Use rtx_cost to determine if the
71: change is really an improvement.
72:
73: * Support more general tail-recursion among different functions.
74:
75: This might be possible under certain circumstances, such as when
76: the argument lists of the functions have the same lengths.
77: Perhaps it could be done with a special declaration.
78:
79: You would need to verify in the calling function that it does not
80: use the addresses of any local variables and does not use setjmp.
81:
82: * Put short statics vars at low addresses and use short addressing mode?
83:
84: Useful on the 68000/68020 and perhaps on the 32000 series,
85: provided one has a linker that works with the feature.
86: This is said to make a 15% speedup on the 68000.
87:
88: * Keep global variables in registers.
89:
90: Here is a scheme for doing this. A global variable, or a local variable
91: whose address is taken, can be kept in a register for an entire function
92: if it does not use non-constant memory addresses and (for globals only)
93: does not call other functions. If the entire function does not meet
94: this criterion, a loop may.
95:
96: The VAR_DECL for such a variable would have to have two RTL expressions:
97: the true home in memory, and the pseudo-register used temporarily.
98: It is necessary to emit insns to copy the memory location into the
99: pseudo-register at the beginning of the function or loop, and perhaps
100: back out at the end. These insns should have REG_EQUIV notes so that,
101: if the pseudo-register does not get a hard register, it is spilled into
102: the memory location which exists in any case.
103:
104: The easiest way to set up these insns is to modify the routine
105: put_var_into_stack so that it does not apply to the entire function
106: (sparing any loops which contain nothing dangerous) and to call it at
107: the end of the function regardless of where in the function the
108: address of a local variable is taken. It would be called
109: unconditionally at the end of the function for all relevant global
110: variables.
111:
112: For debugger output, the thing to do is to invent a new binding level
113: around the appropriate loop and define the variable name as a register
114: variable with that scope.
115:
116: * Live-range splitting.
117:
118: Currently a variable is allocated a hard register either for the full
119: extent of its use or not at all. Sometimes it would be good to
120: allocate a variable a hard register for just part of a function; for
121: example, through a particular loop where the variable is mostly used,
122: or outside of a particular loop where the variable is not used. (The
123: latter is nice because it might let the variable be in a register most
124: of the time even though the loop needs all the registers.)
125:
126: It might not be very hard to do this in global-alloc.c when a variable
127: fails to get a hard register for its entire life span.
128:
129: The first step is to find a loop in which the variable is live, but
130: which is not the whole life span or nearly so. It's probably best to
131: use a loop in which the variable is heavily used.
132:
133: Then create a new pseudo-register to represent the variable in that loop.
134: Substitute this for the old pseudo-register there, and insert move insns
135: to copy between the two at the loop entry and all exits. (When several
136: such moves are inserted at the same place, some new feature should be
137: added to say that none of those registers conflict merely because of
138: overlap between the new moves. And the reload pass should reorder them
139: so that a store precedes a load, for any given hard register.)
140:
141: After doing this for all the reasonable candidates, run global-alloc
142: over again. With luck, one of the two pseudo-registers will be fit
143: somewhere. It may even have a much higher priority due to its reduced
144: life span.
145:
146: There will be no room in general for the new pseudo-registers in
147: basic_block_live_at_start, so there will need to be a second such
148: matrix exclusively for the new ones. Various other vectors indexed by
149: register number will have to be made bigger, or there will have to be
150: secondary extender vectors just for global-alloc.
151:
152: A simple new feature could arrange that both pseudo-registers get the
153: same stack slot if they both fail to get hard registers.
154:
155: Other compilers split live ranges when they are not connected, or
156: try to split off pieces `at the edge'. I think splitting around loops
157: will provide more speedup.
158:
159: Creating a fake binding block and a new like-named variable with
160: shorter life span and different address might succeed in describing
161: this technique for the debugger.
162:
163: * Detect dead stores into memory?
164:
165: A store into memory is dead if it is followed by another store into
166: the same location; and, in between, there is no reference to anything
167: that might be that location (including no reference to a variable
168: address).
169:
170: * Loop optimization.
171:
172: Strength reduction and iteration variable elimination could be
173: smarter. They should know how to decide which iteration variables are
174: not worth making explicit because they can be computed as part of an
175: address calculation. Based on this information, they should decide
176: when it is desirable to eliminate one iteration variable and create
177: another in its place.
178:
179: It should be possible to compute what the value of an iteration
180: variable will be at the end of the loop, and eliminate the variable
181: within the loop by computing that value at the loop end.
182:
183: When a loop has a simple increment that adds 1,
184: instead of jumping in after the increment,
185: decrement the loop count and jump to the increment.
186: This allows aob insns to be used.
187:
188: * Using constraints on values.
189:
190: Many operations could be simplified based on knowledge of the
191: minimum and maximum possible values of a register at any particular time.
192: These limits could come from the data types in the tree, via rtl generation,
193: or they can be deduced from operations that are performed. For example,
194: the result of an `and' operation one of whose operands is 7 must be in
195: the range 0 to 7. Compare instructions also tell something about the
196: possible values of the operand, in the code beyond the test.
197:
198: Value constraints can be used to determine the results of a further
199: comparison. They can also indicate that certain `and' operations are
200: redundant. Constraints might permit a decrement and branch
201: instruction that checks zeroness to be used when the user has
202: specified to exit if negative.
203:
204: * Smarter reload pass.
205:
206: The reload pass as currently written can reload values only into registers
207: that are reserved for reloading. This means that in order to use a
208: register for reloading it must spill everything out of that register.
209:
210: It would be straightforward, though complicated, for reload1.c to keep
211: track, during its scan, of which hard registers were available at each
212: point in the function, and use for reloading even registers that were
213: free only at the point they were needed. This would avoid much spilling
214: and make better code.
215:
216: * Change the type of a variable.
217:
218: Sometimes a variable is declared as `int', it is assigned only once
219: from a value of type `char', and then it is used only by comparison
220: against constants. On many machines, better code would result if
221: the variable had type `char'. If the compiler could detect this
222: case, it could change the declaration of the variable and change
223: all the places that use it.
224:
225: * Better handling for very sparse switches.
226:
227: There may be cases where it would be better to compile a switch
228: statement to use a fixed hash table rather than the current
229: combination of jump tables and binary search.
230:
231: * Order of subexpressions.
232:
233: It might be possible to make better code by paying attention
234: to the order in which to generate code for subexpressions of an expression.
235:
236: * More code motion.
237:
238: Consider hoisting common code up past conditional branches or
239: tablejumps.
240:
241: * Trace scheduling.
242:
243: This technique is said to be able to figure out which way a jump
244: will usually go, and rearrange the code to make that path the
245: faster one.
246:
247: * Distributive law.
248:
249: The C expression *(X + 4 * (Y + C)) compiles better on certain
250: machines if rewritten as *(X + 4*C + 4*Y) because of known addressing
251: modes. It may be tricky to determine when, and for which machines, to
252: use each alternative.
253:
254: Some work has been done on this, in combine.c.
255:
256: * Can optimize by changing if (x) y; else z; into z; if (x) y;
257: if z and x do not interfere and z has no effects not undone by y.
258: This is desirable if z is faster than jumping.
259:
260: * For a two-insn loop on the 68020, such as
261: foo: movb a2@+,a3@+
262: jne foo
263: it is better to insert dbeq d0,foo before the jne.
264: d0 can be a junk register. The challenge is to fit this into
265: a portable framework: when can you detect this situation and
266: still be able to allocate a junk register?
267:
268: 2. Simpler porting.
269:
270: Right now, describing the target machine's instructions is done
271: cleanly, but describing its addressing mode is done with several
272: ad-hoc macro definitions. Porting would be much easier if there were
273: an RTL description for addressing modes like that for instructions.
274: Tools analogous to genflags and genrecog would generate macros from
275: this description.
276:
277: There would be one pattern in the address-description file for each
278: kind of addressing, and this pattern would have:
279:
280: * the RTL expression for the address
281: * C code to verify its validity (since that may depend on
282: the exact data).
283: * C code to print the address in assembler language.
284: * C code to convert the address into a valid one, if it is not valid.
285: (This would replace LEGITIMIZE_ADDRESS).
286: * Register constraints for all indeterminates that appear
287: in the RTL expression.
288:
289: 3. Other languages.
290:
291: Front ends for Pascal, Fortran, Algol, Cobol, Modula-2 and Ada are
292: desirable.
293:
294: Pascal, Modula-2 and Ada require the implementation of functions
295: within functions. Some of the mechanisms for this already exist.
296:
297: 4. More extensions.
298:
299: * Generated unique labels. Have some way of generating distinct labels
300: for use in extended asm statements. I don't know what a good syntax would
301: be.
302:
303: * A way of defining a structure containing a union, in which the choice of
304: union alternative is controlled by a previous structure component.
305:
306: Here is a possible syntax for this.
307:
308: struct foo {
309: enum { INT, DOUBLE } code;
310: auto union { case INT: int i; case DOUBLE: double d;} value : code;
311: };
312:
313: * Allow constructor expressions as lvalues, like this:
314:
315: (struct foo) {a, b, c} = foo();
316:
317: This would call foo, which returns a structure, and then store the
318: several components of the structure into the variables a, b, and c.
319:
320: 5. Generalize the machine model.
321:
322: * Some new compiler features may be needed to do a good job on machines
323: where static data needs to be addressed using base registers.
324:
325: * Some machines have two stacks in different areas of memory, one used
326: for scalars and another for large objects. The compiler does not
327: now have a way to understand this.
328:
329: 6. Useful warnings.
330:
331: * Warn about statements that are undefined because the order of
332: evaluation of increment operators makes a big difference. Here is an
333: example:
334:
335: *foo++ = hack (*foo);
336:
337: 7. Better documentation of how GCC works and how to port it.
338:
339: Here is an outline proposed by Allan Adler.
340:
341: I. Overview of this document
342: II. The machines on which GCC is implemented
343: A. Prose description of those characteristics of target machines and
344: their operating systems which are pertinent to the implementation
345: of GCC.
346: i. target machine characteristics
347: ii. comparison of this system of machine characteristics with
348: other systems of machine specification currently in use
349: B. Tables of the characteristics of the target machines on which
350: GCC is implemented.
351: C. A priori restrictions on the values of characteristics of target
352: machines, with special reference to those parts of the source code
353: which entail those restrictions
354: i. restrictions on individual characteristics
355: ii. restrictions involving relations between various characteristics
356: D. The use of GCC as a cross-compiler
357: i. cross-compilation to existing machines
358: ii. cross-compilation to non-existent machines
359: E. Assumptions which are made regarding the target machine
360: i. assumptions regarding the architecture of the target machine
361: ii. assumptions regarding the operating system of the target machine
362: iii. assumptions regarding software resident on the target machine
363: iv. where in the source code these assumptions are in effect made
364: III. A systematic approach to writing the files tm.h and xm.h
365: A. Macros which require special care or skill
366: B. Examples, with special reference to the underlying reasoning
367: IV. A systematic approach to writing the machine description file md
368: A. Minimal viable sets of insn descriptions
369: B. Examples, with special reference to the underlying reasoning
370: V. Uses of the file aux-output.c
371: VI. Specification of what constitutes correct performance of an
372: implementation of GCC
373: A. The components of GCC
374: B. The itinerary of a C program through GCC
375: C. A system of benchmark programs
376: D. What your RTL and assembler should look like with these benchmarks
377: E. Fine tuning for speed and size of compiled code
378: VII. A systematic procedure for debugging an implementation of GCC
379: A. Use of GDB
380: i. the macros in the file .gdbinit for GCC
381: ii. obstacles to the use of GDB
382: a. functions implemented as macros can't be called in GDB
383: B. Debugging without GDB
384: i. How to turn off the normal operation of GCC and access specific
385: parts of GCC
386: C. Debugging tools
387: D. Debugging the parser
388: i. how machine macros and insn definitions affect the parser
389: E. Debugging the recognizer
390: i. how machine macros and insn definitions affect the recognizer
391:
392: ditto for other components
393:
394: VIII. Data types used by GCC, with special reference to restrictions not
395: specified in the formal definition of the data type
396: IX. References to the literature for the algorithms used in GCC
397:
This archive runs on limited infrastructure. Preserving old code on modern bandwidth. Automated agents are requested to crawl responsibly.