3 * implement do-loop -> bdnz transform
4 * implement powerpc-64 for darwin
6 ===-------------------------------------------------------------------------===
8 Use the stfiwx instruction for:
10 void foo(float a, int *b) { *b = a; }
12 ===-------------------------------------------------------------------------===
14 unsigned short foo(float a) { return a; }
26 rlwinm r3, r2, 0, 16, 31
29 ===-------------------------------------------------------------------------===
31 Support 'update' load/store instructions. These are cracked on the G5, but are
34 ===-------------------------------------------------------------------------===
36 Should hint to the branch select pass that it doesn't need to print the second
37 unconditional branch, so we don't end up with things like:
38 b .LBBl42__2E_expand_function_8_674 ; loopentry.24
39 b .LBBl42__2E_expand_function_8_42 ; NewDefault
40 b .LBBl42__2E_expand_function_8_42 ; NewDefault
44 ===-------------------------------------------------------------------------===
49 if (X == 0x12345678) bar();
65 ===-------------------------------------------------------------------------===
67 Lump the constant pool for each function into ONE pic object, and reference
68 pieces of it as offsets from the start. For functions like this (contrived
69 to have lots of constants obviously):
71 double X(double Y) { return (Y*1.23 + 4.512)*2.34 + 14.38; }
76 lis r2, ha16(.CPI_X_0)
77 lfd f0, lo16(.CPI_X_0)(r2)
78 lis r2, ha16(.CPI_X_1)
79 lfd f2, lo16(.CPI_X_1)(r2)
81 lis r2, ha16(.CPI_X_2)
82 lfd f1, lo16(.CPI_X_2)(r2)
83 lis r2, ha16(.CPI_X_3)
84 lfd f2, lo16(.CPI_X_3)(r2)
88 It would be better to materialize .CPI_X into a register, then use immediates
89 off of the register to avoid the lis's. This is even more important in PIC
92 Note that this (and the static variable version) is discussed here for GCC:
93 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
95 ===-------------------------------------------------------------------------===
97 PIC Code Gen IPO optimization:
99 Squish small scalar globals together into a single global struct, allowing the
100 address of the struct to be CSE'd, avoiding PIC accesses (also reduces the size
101 of the GOT on targets with one).
103 Note that this is discussed here for GCC:
104 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
106 ===-------------------------------------------------------------------------===
108 Implement Newton-Rhapson method for improving estimate instructions to the
109 correct accuracy, and implementing divide as multiply by reciprocal when it has
110 more than one use. Itanium will want this too.
112 ===-------------------------------------------------------------------------===
114 #define ARRAY_LENGTH 16
119 unsigned int field0 : 6;
120 unsigned int field1 : 6;
121 unsigned int field2 : 6;
122 unsigned int field3 : 6;
123 unsigned int field4 : 3;
124 unsigned int field5 : 4;
125 unsigned int field6 : 1;
127 unsigned int field6 : 1;
128 unsigned int field5 : 4;
129 unsigned int field4 : 3;
130 unsigned int field3 : 6;
131 unsigned int field2 : 6;
132 unsigned int field1 : 6;
133 unsigned int field0 : 6;
142 typedef struct program_t {
143 union bitfield array[ARRAY_LENGTH];
149 void AdjustBitfields(program* prog, unsigned int fmt1)
151 prog->array[0].bitfields.field0 = fmt1;
152 prog->array[0].bitfields.field1 = fmt1 + 1;
155 We currently generate:
160 rlwinm r2, r2, 0, 0, 19
161 rlwinm r5, r5, 6, 20, 25
162 rlwimi r2, r4, 0, 26, 31
167 We should teach someone that or (rlwimi, rlwinm) with disjoint masks can be
168 turned into rlwimi (rlwimi)
170 The better codegen would be:
181 ===-------------------------------------------------------------------------===
185 int %f1(int %a, int %b) {
186 %tmp.1 = and int %a, 15 ; <int> [#uses=1]
187 %tmp.3 = and int %b, 240 ; <int> [#uses=1]
188 %tmp.4 = or int %tmp.3, %tmp.1 ; <int> [#uses=1]
192 without a copy. We make this currently:
195 rlwinm r2, r4, 0, 24, 27
196 rlwimi r2, r3, 0, 28, 31
200 The two-addr pass or RA needs to learn when it is profitable to commute an
201 instruction to avoid a copy AFTER the 2-addr instruction. The 2-addr pass
202 currently only commutes to avoid inserting a copy BEFORE the two addr instr.
204 ===-------------------------------------------------------------------------===
206 Compile offsets from allocas:
209 %X = alloca { int, int }
210 %Y = getelementptr {int,int}* %X, int 0, uint 1
214 into a single add, not two:
221 --> important for C++.
223 ===-------------------------------------------------------------------------===
225 int test3(int a, int b) { return (a < 0) ? a : 0; }
227 should be branch free code. LLVM is turning it into < 1 because of the RHS.
229 ===-------------------------------------------------------------------------===
231 No loads or stores of the constants should be needed:
233 struct foo { double X, Y; };
234 void xxx(struct foo F);
235 void bar() { struct foo R = { 1.0, 2.0 }; xxx(R); }
237 ===-------------------------------------------------------------------------===
239 Darwin Stub LICM optimization:
245 Have to go through an indirect stub if bar is external or linkonce. It would
246 be better to compile it as:
251 which only computes the address of bar once (instead of each time through the
252 stub). This is Darwin specific and would have to be done in the code generator.
253 Probably not a win on x86.
255 ===-------------------------------------------------------------------------===
257 PowerPC i1/setcc stuff (depends on subreg stuff):
259 Check out the PPC code we get for 'compare' in this testcase:
260 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19672
262 oof. on top of not doing the logical crnand instead of (mfcr, mfcr,
263 invert, invert, or), we then have to compare it against zero instead of
264 using the value already in a CR!
266 that should be something like
270 bne cr0, LBB_compare_4
278 rlwinm r7, r7, 30, 31, 31
279 rlwinm r8, r8, 30, 31, 31
285 bne cr0, LBB_compare_4 ; loopexit
287 FreeBench/mason has a basic block that looks like this:
289 %tmp.130 = seteq int %p.0__, 5 ; <bool> [#uses=1]
290 %tmp.134 = seteq int %p.1__, 6 ; <bool> [#uses=1]
291 %tmp.139 = seteq int %p.2__, 12 ; <bool> [#uses=1]
292 %tmp.144 = seteq int %p.3__, 13 ; <bool> [#uses=1]
293 %tmp.149 = seteq int %p.4__, 14 ; <bool> [#uses=1]
294 %tmp.154 = seteq int %p.5__, 15 ; <bool> [#uses=1]
295 %bothcond = and bool %tmp.134, %tmp.130 ; <bool> [#uses=1]
296 %bothcond123 = and bool %bothcond, %tmp.139 ; <bool>
297 %bothcond124 = and bool %bothcond123, %tmp.144 ; <bool>
298 %bothcond125 = and bool %bothcond124, %tmp.149 ; <bool>
299 %bothcond126 = and bool %bothcond125, %tmp.154 ; <bool>
300 br bool %bothcond126, label %shortcirc_next.5, label %else.0
302 This is a particularly important case where handling CRs better will help.
304 ===-------------------------------------------------------------------------===
306 Simple IPO for argument passing, change:
307 void foo(int X, double Y, int Z) -> void foo(int X, int Z, double Y)
309 the Darwin ABI specifies that any integer arguments in the first 32 bytes worth
310 of arguments get assigned to r3 through r10. That is, if you have a function
311 foo(int, double, int) you get r3, f1, r6, since the 64 bit double ate up the
312 argument bytes for r4 and r5. The trick then would be to shuffle the argument
313 order for functions we can internalize so that the maximum number of
314 integers/pointers get passed in regs before you see any of the fp arguments.
316 Instead of implementing this, it would actually probably be easier to just
317 implement a PPC fastcc, where we could do whatever we wanted to the CC,
318 including having this work sanely.
320 ===-------------------------------------------------------------------------===
322 Fix Darwin FP-In-Integer Registers ABI
324 Darwin passes doubles in structures in integer registers, which is very very
325 bad. Add something like a BIT_CONVERT to LLVM, then do an i-p transformation
326 that percolates these things out of functions.
328 Check out how horrible this is:
329 http://gcc.gnu.org/ml/gcc/2005-10/msg01036.html
331 This is an extension of "interprocedural CC unmunging" that can't be done with
334 ===-------------------------------------------------------------------------===
336 Generate lwbrx and other byteswapping load/store instructions when reasonable.
338 ===-------------------------------------------------------------------------===
340 Implement TargetConstantVec, and set up PPC to custom lower ConstantVec into
341 TargetConstantVec's if it's one of the many forms that are algorithmically
342 computable using the spiffy altivec instructions.
344 ===-------------------------------------------------------------------------===
348 double %test(double %X) {
349 %Y = cast double %X to long
350 %Z = cast long %Y to double
367 without the lwz/stw's.
369 ===-------------------------------------------------------------------------===
376 return b * 3; // ignore the fact that this is always 3.
382 into something not this:
387 rlwinm r2, r2, 29, 31, 31
389 bgt cr0, LBB1_2 ; UnifiedReturnBlock
391 rlwinm r2, r2, 0, 31, 31
394 LBB1_2: ; UnifiedReturnBlock
398 In particular, the two compares (marked 1) could be shared by reversing one.
399 This could be done in the dag combiner, by swapping a BR_CC when a SETCC of the
400 same operands (but backwards) exists. In this case, this wouldn't save us
401 anything though, because the compares still wouldn't be shared.
403 ===-------------------------------------------------------------------------===
405 The legalizer should lower this:
407 bool %test(ulong %x) {
408 %tmp = setlt ulong %x, 4294967296
412 into "if x.high == 0", not:
428 noticed in 2005-05-11-Popcount-ffs-fls.c.
431 ===-------------------------------------------------------------------------===
433 We should custom expand setcc instead of pretending that we have it. That
434 would allow us to expose the access of the crbit after the mfcr, allowing
435 that access to be trivially folded into other ops. A simple example:
437 int foo(int a, int b) { return (a < b) << 4; }
444 rlwinm r2, r2, 29, 31, 31
448 ===-------------------------------------------------------------------------===
450 Fold add and sub with constant into non-extern, non-weak addresses so this:
453 void bar(int b) { a = b; }
454 void foo(unsigned char *c) {
471 lbz r2, lo16(_a+3)(r2)
475 ===-------------------------------------------------------------------------===
477 We generate really bad code for this:
479 int f(signed char *a, _Bool b, _Bool c) {