1 //===- README.txt - Notes for improving PowerPC-specific code gen ---------===//
5 * implement do-loop -> bdnz transform
6 * implement powerpc-64 for darwin
8 ===-------------------------------------------------------------------------===
10 Support 'update' load/store instructions. These are cracked on the G5, but are
13 ===-------------------------------------------------------------------------===
15 Teach the .td file to pattern match PPC::BR_COND to appropriate bc variant, so
16 we don't have to always run the branch selector for small functions.
18 ===-------------------------------------------------------------------------===
23 if (X == 0x12345678) bar();
39 ===-------------------------------------------------------------------------===
41 Lump the constant pool for each function into ONE pic object, and reference
42 pieces of it as offsets from the start. For functions like this (contrived
43 to have lots of constants obviously):
45 double X(double Y) { return (Y*1.23 + 4.512)*2.34 + 14.38; }
50 lis r2, ha16(.CPI_X_0)
51 lfd f0, lo16(.CPI_X_0)(r2)
52 lis r2, ha16(.CPI_X_1)
53 lfd f2, lo16(.CPI_X_1)(r2)
55 lis r2, ha16(.CPI_X_2)
56 lfd f1, lo16(.CPI_X_2)(r2)
57 lis r2, ha16(.CPI_X_3)
58 lfd f2, lo16(.CPI_X_3)(r2)
62 It would be better to materialize .CPI_X into a register, then use immediates
63 off of the register to avoid the lis's. This is even more important in PIC
66 Note that this (and the static variable version) is discussed here for GCC:
67 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
69 ===-------------------------------------------------------------------------===
71 PIC Code Gen IPO optimization:
73 Squish small scalar globals together into a single global struct, allowing the
74 address of the struct to be CSE'd, avoiding PIC accesses (also reduces the size
75 of the GOT on targets with one).
77 Note that this is discussed here for GCC:
78 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
80 ===-------------------------------------------------------------------------===
82 Implement Newton-Rhapson method for improving estimate instructions to the
83 correct accuracy, and implementing divide as multiply by reciprocal when it has
84 more than one use. Itanium will want this too.
86 ===-------------------------------------------------------------------------===
88 #define ARRAY_LENGTH 16
93 unsigned int field0 : 6;
94 unsigned int field1 : 6;
95 unsigned int field2 : 6;
96 unsigned int field3 : 6;
97 unsigned int field4 : 3;
98 unsigned int field5 : 4;
99 unsigned int field6 : 1;
101 unsigned int field6 : 1;
102 unsigned int field5 : 4;
103 unsigned int field4 : 3;
104 unsigned int field3 : 6;
105 unsigned int field2 : 6;
106 unsigned int field1 : 6;
107 unsigned int field0 : 6;
116 typedef struct program_t {
117 union bitfield array[ARRAY_LENGTH];
123 void AdjustBitfields(program* prog, unsigned int fmt1)
125 prog->array[0].bitfields.field0 = fmt1;
126 prog->array[0].bitfields.field1 = fmt1 + 1;
129 We currently generate:
134 rlwinm r2, r2, 0, 0, 19
135 rlwinm r5, r5, 6, 20, 25
136 rlwimi r2, r4, 0, 26, 31
141 We should teach someone that or (rlwimi, rlwinm) with disjoint masks can be
142 turned into rlwimi (rlwimi)
144 The better codegen would be:
155 ===-------------------------------------------------------------------------===
159 int %f1(int %a, int %b) {
160 %tmp.1 = and int %a, 15 ; <int> [#uses=1]
161 %tmp.3 = and int %b, 240 ; <int> [#uses=1]
162 %tmp.4 = or int %tmp.3, %tmp.1 ; <int> [#uses=1]
166 without a copy. We make this currently:
169 rlwinm r2, r4, 0, 24, 27
170 rlwimi r2, r3, 0, 28, 31
174 The two-addr pass or RA needs to learn when it is profitable to commute an
175 instruction to avoid a copy AFTER the 2-addr instruction. The 2-addr pass
176 currently only commutes to avoid inserting a copy BEFORE the two addr instr.
178 ===-------------------------------------------------------------------------===
180 Compile offsets from allocas:
183 %X = alloca { int, int }
184 %Y = getelementptr {int,int}* %X, int 0, uint 1
188 into a single add, not two:
195 --> important for C++.
197 ===-------------------------------------------------------------------------===
199 int test3(int a, int b) { return (a < 0) ? a : 0; }
201 should be branch free code. LLVM is turning it into < 1 because of the RHS.
203 ===-------------------------------------------------------------------------===
205 No loads or stores of the constants should be needed:
207 struct foo { double X, Y; };
208 void xxx(struct foo F);
209 void bar() { struct foo R = { 1.0, 2.0 }; xxx(R); }
211 ===-------------------------------------------------------------------------===
213 Darwin Stub LICM optimization:
219 Have to go through an indirect stub if bar is external or linkonce. It would
220 be better to compile it as:
225 which only computes the address of bar once (instead of each time through the
226 stub). This is Darwin specific and would have to be done in the code generator.
227 Probably not a win on x86.
229 ===-------------------------------------------------------------------------===
231 PowerPC i1/setcc stuff (depends on subreg stuff):
233 Check out the PPC code we get for 'compare' in this testcase:
234 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19672
236 oof. on top of not doing the logical crnand instead of (mfcr, mfcr,
237 invert, invert, or), we then have to compare it against zero instead of
238 using the value already in a CR!
240 that should be something like
244 bne cr0, LBB_compare_4
252 rlwinm r7, r7, 30, 31, 31
253 rlwinm r8, r8, 30, 31, 31
259 bne cr0, LBB_compare_4 ; loopexit
261 FreeBench/mason has a basic block that looks like this:
263 %tmp.130 = seteq int %p.0__, 5 ; <bool> [#uses=1]
264 %tmp.134 = seteq int %p.1__, 6 ; <bool> [#uses=1]
265 %tmp.139 = seteq int %p.2__, 12 ; <bool> [#uses=1]
266 %tmp.144 = seteq int %p.3__, 13 ; <bool> [#uses=1]
267 %tmp.149 = seteq int %p.4__, 14 ; <bool> [#uses=1]
268 %tmp.154 = seteq int %p.5__, 15 ; <bool> [#uses=1]
269 %bothcond = and bool %tmp.134, %tmp.130 ; <bool> [#uses=1]
270 %bothcond123 = and bool %bothcond, %tmp.139 ; <bool>
271 %bothcond124 = and bool %bothcond123, %tmp.144 ; <bool>
272 %bothcond125 = and bool %bothcond124, %tmp.149 ; <bool>
273 %bothcond126 = and bool %bothcond125, %tmp.154 ; <bool>
274 br bool %bothcond126, label %shortcirc_next.5, label %else.0
276 This is a particularly important case where handling CRs better will help.
278 ===-------------------------------------------------------------------------===
280 Simple IPO for argument passing, change:
281 void foo(int X, double Y, int Z) -> void foo(int X, int Z, double Y)
283 the Darwin ABI specifies that any integer arguments in the first 32 bytes worth
284 of arguments get assigned to r3 through r10. That is, if you have a function
285 foo(int, double, int) you get r3, f1, r6, since the 64 bit double ate up the
286 argument bytes for r4 and r5. The trick then would be to shuffle the argument
287 order for functions we can internalize so that the maximum number of
288 integers/pointers get passed in regs before you see any of the fp arguments.
290 Instead of implementing this, it would actually probably be easier to just
291 implement a PPC fastcc, where we could do whatever we wanted to the CC,
292 including having this work sanely.
294 ===-------------------------------------------------------------------------===
296 Fix Darwin FP-In-Integer Registers ABI
298 Darwin passes doubles in structures in integer registers, which is very very
299 bad. Add something like a BIT_CONVERT to LLVM, then do an i-p transformation
300 that percolates these things out of functions.
302 Check out how horrible this is:
303 http://gcc.gnu.org/ml/gcc/2005-10/msg01036.html
305 This is an extension of "interprocedural CC unmunging" that can't be done with
308 ===-------------------------------------------------------------------------===
310 Generate lwbrx and other byteswapping load/store instructions when reasonable.
312 ===-------------------------------------------------------------------------===
319 return b * 3; // ignore the fact that this is always 3.
325 into something not this:
330 rlwinm r2, r2, 29, 31, 31
332 bgt cr0, LBB1_2 ; UnifiedReturnBlock
334 rlwinm r2, r2, 0, 31, 31
337 LBB1_2: ; UnifiedReturnBlock
341 In particular, the two compares (marked 1) could be shared by reversing one.
342 This could be done in the dag combiner, by swapping a BR_CC when a SETCC of the
343 same operands (but backwards) exists. In this case, this wouldn't save us
344 anything though, because the compares still wouldn't be shared.
346 ===-------------------------------------------------------------------------===
348 The legalizer should lower this:
350 bool %test(ulong %x) {
351 %tmp = setlt ulong %x, 4294967296
355 into "if x.high == 0", not:
371 noticed in 2005-05-11-Popcount-ffs-fls.c.
374 ===-------------------------------------------------------------------------===
376 We should custom expand setcc instead of pretending that we have it. That
377 would allow us to expose the access of the crbit after the mfcr, allowing
378 that access to be trivially folded into other ops. A simple example:
380 int foo(int a, int b) { return (a < b) << 4; }
387 rlwinm r2, r2, 29, 31, 31
391 ===-------------------------------------------------------------------------===
393 Fold add and sub with constant into non-extern, non-weak addresses so this:
396 void bar(int b) { a = b; }
397 void foo(unsigned char *c) {
414 lbz r2, lo16(_a+3)(r2)
418 ===-------------------------------------------------------------------------===
420 We generate really bad code for this:
422 int f(signed char *a, _Bool b, _Bool c) {
428 ===-------------------------------------------------------------------------===
431 int test(unsigned *P) { return *P >> 24; }
446 ===-------------------------------------------------------------------------===
448 On the G5, logical CR operations are more expensive in their three
449 address form: ops that read/write the same register are half as expensive as
450 those that read from two registers that are different from their destination.
452 We should model this with two separate instructions. The isel should generate
453 the "two address" form of the instructions. When the register allocator
454 detects that it needs to insert a copy due to the two-addresness of the CR
455 logical op, it will invoke PPCInstrInfo::convertToThreeAddress. At this point
456 we can convert to the "three address" instruction, to save code space.
458 This only matters when we start generating cr logical ops.
460 ===-------------------------------------------------------------------------===
462 We should compile these two functions to the same thing:
465 void f(int a, int b, int *P) {
466 *P = (a-b)>=0?(a-b):(b-a);
468 void g(int a, int b, int *P) {
472 Further, they should compile to something better than:
478 bgt cr0, LBB2_2 ; entry
495 ... which is much nicer.
497 This theoretically may help improve twolf slightly (used in dimbox.c:142?).
499 ===-------------------------------------------------------------------------===
501 int foo(int N, int ***W, int **TK, int X) {
504 for (t = 0; t < N; ++t)
505 for (i = 0; i < 4; ++i)
506 W[t / X][i][t % X] = TK[i][t];
511 We generate relatively atrocious code for this loop compared to gcc.
513 We could also strength reduce the rem and the div:
514 http://www.lcs.mit.edu/pubs/pdf/MIT-LCS-TM-600.pdf
516 ===-------------------------------------------------------------------------===
518 float foo(float X) { return (int)(X); }
533 We could use a target dag combine to turn the lwz/extsw into an lwa when the
534 lwz has a single use. Since LWA is cracked anyway, this would be a codesize
537 ===-------------------------------------------------------------------------===
539 We generate ugly code for this:
541 void func(unsigned int *ret, float dx, float dy, float dz, float dw) {
543 if(dx < -dw) code |= 1;
544 if(dx > dw) code |= 2;
545 if(dy < -dw) code |= 4;
546 if(dy > dw) code |= 8;
547 if(dz < -dw) code |= 16;
548 if(dz > dw) code |= 32;
552 ===-------------------------------------------------------------------------===