1 //===- README.txt - Notes for improving PowerPC-specific code gen ---------===//
4 * lmw/stmw pass a la arm load store optimizer for prolog/epilog
6 ===-------------------------------------------------------------------------===
10 unsigned add32carry(unsigned sum, unsigned x) {
17 Should compile to something like:
27 rlwinm r4, r4, 29, 31, 31
32 ===-------------------------------------------------------------------------===
34 We compile the hottest inner loop of viterbi to:
45 bne cr0, LBB1_83 ;bb420.i
47 The CBE manages to produce:
58 This could be much better (bdnz instead of bdz) but it still beats us. If we
59 produced this with bdnz, the loop would be a single dispatch group.
61 ===-------------------------------------------------------------------------===
63 Lump the constant pool for each function into ONE pic object, and reference
64 pieces of it as offsets from the start. For functions like this (contrived
65 to have lots of constants obviously):
67 double X(double Y) { return (Y*1.23 + 4.512)*2.34 + 14.38; }
72 lis r2, ha16(.CPI_X_0)
73 lfd f0, lo16(.CPI_X_0)(r2)
74 lis r2, ha16(.CPI_X_1)
75 lfd f2, lo16(.CPI_X_1)(r2)
77 lis r2, ha16(.CPI_X_2)
78 lfd f1, lo16(.CPI_X_2)(r2)
79 lis r2, ha16(.CPI_X_3)
80 lfd f2, lo16(.CPI_X_3)(r2)
84 It would be better to materialize .CPI_X into a register, then use immediates
85 off of the register to avoid the lis's. This is even more important in PIC
88 Note that this (and the static variable version) is discussed here for GCC:
89 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
91 Here's another example (the sgn function):
92 double testf(double a) {
93 return a == 0.0 ? 0.0 : (a > 0.0 ? 1.0 : -1.0);
96 it produces a BB like this:
99 lfs f0, lo16(LCPI1_0)(r2)
100 lis r2, ha16(LCPI1_1)
101 lis r3, ha16(LCPI1_2)
102 lfs f2, lo16(LCPI1_2)(r3)
103 lfs f3, lo16(LCPI1_1)(r2)
108 ===-------------------------------------------------------------------------===
110 PIC Code Gen IPO optimization:
112 Squish small scalar globals together into a single global struct, allowing the
113 address of the struct to be CSE'd, avoiding PIC accesses (also reduces the size
114 of the GOT on targets with one).
116 Note that this is discussed here for GCC:
117 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
119 ===-------------------------------------------------------------------------===
121 Fold add and sub with constant into non-extern, non-weak addresses so this:
124 void bar(int b) { a = b; }
125 void foo(unsigned char *c) {
142 lbz r2, lo16(_a+3)(r2)
146 ===-------------------------------------------------------------------------===
148 We should compile these two functions to the same thing:
151 void f(int a, int b, int *P) {
152 *P = (a-b)>=0?(a-b):(b-a);
154 void g(int a, int b, int *P) {
158 Further, they should compile to something better than:
164 bgt cr0, LBB2_2 ; entry
181 ... which is much nicer.
183 This theoretically may help improve twolf slightly (used in dimbox.c:142?).
185 ===-------------------------------------------------------------------------===
188 define i32 @clamp0g(i32 %a) {
190 %cmp = icmp slt i32 %a, 0
191 %sel = select i1 %cmp, i32 0, i32 %a
195 Is compile to this with the PowerPC (32-bit) backend:
207 This could be reduced to the much simpler:
214 ===-------------------------------------------------------------------------===
216 int foo(int N, int ***W, int **TK, int X) {
219 for (t = 0; t < N; ++t)
220 for (i = 0; i < 4; ++i)
221 W[t / X][i][t % X] = TK[i][t];
226 We generate relatively atrocious code for this loop compared to gcc.
228 We could also strength reduce the rem and the div:
229 http://www.lcs.mit.edu/pubs/pdf/MIT-LCS-TM-600.pdf
231 ===-------------------------------------------------------------------------===
233 We generate ugly code for this:
235 void func(unsigned int *ret, float dx, float dy, float dz, float dw) {
237 if(dx < -dw) code |= 1;
238 if(dx > dw) code |= 2;
239 if(dy < -dw) code |= 4;
240 if(dy > dw) code |= 8;
241 if(dz < -dw) code |= 16;
242 if(dz > dw) code |= 32;
246 ===-------------------------------------------------------------------------===
248 %struct.B = type { i8, [3 x i8] }
250 define void @bar(%struct.B* %b) {
252 %tmp = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
253 %tmp = load i32* %tmp ; <uint> [#uses=1]
254 %tmp3 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
255 %tmp4 = load i32* %tmp3 ; <uint> [#uses=1]
256 %tmp8 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=2]
257 %tmp9 = load i32* %tmp8 ; <uint> [#uses=1]
258 %tmp4.mask17 = shl i32 %tmp4, i8 1 ; <uint> [#uses=1]
259 %tmp1415 = and i32 %tmp4.mask17, 2147483648 ; <uint> [#uses=1]
260 %tmp.masked = and i32 %tmp, 2147483648 ; <uint> [#uses=1]
261 %tmp11 = or i32 %tmp1415, %tmp.masked ; <uint> [#uses=1]
262 %tmp12 = and i32 %tmp9, 2147483647 ; <uint> [#uses=1]
263 %tmp13 = or i32 %tmp12, %tmp11 ; <uint> [#uses=1]
264 store i32 %tmp13, i32* %tmp8
274 rlwimi r2, r4, 0, 0, 0
278 We could collapse a bunch of those ORs and ANDs and generate the following
283 rlwinm r4, r2, 1, 0, 0
288 ===-------------------------------------------------------------------------===
290 Consider a function like this:
292 float foo(float X) { return X + 1234.4123f; }
294 The FP constant ends up in the constant pool, so we need to get the LR register.
295 This ends up producing code like this:
304 addis r2, r2, ha16(.CPI_foo_0-"L00000$pb")
305 lfs f0, lo16(.CPI_foo_0-"L00000$pb")(r2)
311 This is functional, but there is no reason to spill the LR register all the way
312 to the stack (the two marked instrs): spilling it to a GPR is quite enough.
314 Implementing this will require some codegen improvements. Nate writes:
316 "So basically what we need to support the "no stack frame save and restore" is a
317 generalization of the LR optimization to "callee-save regs".
319 Currently, we have LR marked as a callee-save reg. The register allocator sees
320 that it's callee save, and spills it directly to the stack.
322 Ideally, something like this would happen:
324 LR would be in a separate register class from the GPRs. The class of LR would be
325 marked "unspillable". When the register allocator came across an unspillable
326 reg, it would ask "what is the best class to copy this into that I *can* spill"
327 If it gets a class back, which it will in this case (the gprs), it grabs a free
328 register of that class. If it is then later necessary to spill that reg, so be
331 ===-------------------------------------------------------------------------===
335 return X ? 524288 : 0;
343 beq cr0, LBB1_2 ;entry
356 This sort of thing occurs a lot due to globalopt.
358 ===-------------------------------------------------------------------------===
362 define i32 @bar(i32 %x) nounwind readnone ssp {
364 %0 = icmp eq i32 %x, 0 ; <i1> [#uses=1]
365 %neg = sext i1 %0 to i32 ; <i32> [#uses=1]
377 it would be better to produce:
384 ===-------------------------------------------------------------------------===
386 We generate horrible ppc code for this:
398 addi r5, r5, 1 ;; Extra IV for the exit value compare.
402 xoris r6, r5, 30 ;; This is due to a large immediate.
403 cmplwi cr0, r6, 33920
406 //===---------------------------------------------------------------------===//
410 inline std::pair<unsigned, bool> full_add(unsigned a, unsigned b)
411 { return std::make_pair(a + b, a + b < a); }
412 bool no_overflow(unsigned a, unsigned b)
413 { return !full_add(a, b).second; }
430 rlwinm r2, r2, 29, 31, 31
434 //===---------------------------------------------------------------------===//
436 We compile some FP comparisons into an mfcr with two rlwinms and an or. For
439 int test(double x, double y) { return islessequal(x, y);}
440 int test2(double x, double y) { return islessgreater(x, y);}
441 int test3(double x, double y) { return !islessequal(x, y);}
443 Compiles into (all three are similar, but the bits differ):
448 rlwinm r3, r2, 29, 31, 31
449 rlwinm r2, r2, 31, 31, 31
453 GCC compiles this into:
462 which is more efficient and can use mfocr. See PR642 for some more context.
464 //===---------------------------------------------------------------------===//
466 void foo(float *data, float d) {
468 for (i = 0; i < 8000; i++)
471 void foo2(float *data, float d) {
474 for (i = 0; i < 8000; i++) {
487 cmplwi cr0, r4, 32000
496 cmplwi cr0, r4, 32000
501 The 'mr' could be eliminated to folding the add into the cmp better.
503 //===---------------------------------------------------------------------===//
504 Codegen for the following (low-probability) case deteriorated considerably
505 when the correctness fixes for unordered comparisons went in (PR 642, 58871).
506 It should be possible to recover the code quality described in the comments.
508 ; RUN: llvm-as < %s | llc -march=ppc32 | grep or | count 3
509 ; This should produce one 'or' or 'cror' instruction per function.
511 ; RUN: llvm-as < %s | llc -march=ppc32 | grep mfcr | count 3
514 define i32 @test(double %x, double %y) nounwind {
516 %tmp3 = fcmp ole double %x, %y ; <i1> [#uses=1]
517 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
521 define i32 @test2(double %x, double %y) nounwind {
523 %tmp3 = fcmp one double %x, %y ; <i1> [#uses=1]
524 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
528 define i32 @test3(double %x, double %y) nounwind {
530 %tmp3 = fcmp ugt double %x, %y ; <i1> [#uses=1]
531 %tmp34 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
535 //===---------------------------------------------------------------------===//
536 for the following code:
538 void foo (float *__restrict__ a, int *__restrict__ b, int n) {
542 we load b[n] to GPR, then move it VSX register and convert it float. We should
543 use vsx scalar integer load instructions to avoid direct moves
545 //===----------------------------------------------------------------------===//
546 ; RUN: llvm-as < %s | llc -march=ppc32 | not grep fneg
548 ; This could generate FSEL with appropriate flags (FSEL is not IEEE-safe, and
549 ; should not be generated except with -enable-finite-only-fp-math or the like).
550 ; With the correctness fixes for PR642 (58871) LowerSELECT_CC would need to
551 ; recognize a more elaborate tree than a simple SETxx.
553 define double @test_FNEG_sel(double %A, double %B, double %C) {
554 %D = fsub double -0.000000e+00, %A ; <double> [#uses=1]
555 %Cond = fcmp ugt double %D, -0.000000e+00 ; <i1> [#uses=1]
556 %E = select i1 %Cond, double %B, double %C ; <double> [#uses=1]
560 //===----------------------------------------------------------------------===//
561 The save/restore sequence for CR in prolog/epilog is terrible:
562 - Each CR subreg is saved individually, rather than doing one save as a unit.
563 - On Darwin, the save is done after the decrement of SP, which means the offset
564 from SP of the save slot can be too big for a store instruction, which means we
565 need an additional register (currently hacked in 96015+96020; the solution there
566 is correct, but poor).
567 - On SVR4 the same thing can happen, and I don't think saving before the SP
568 decrement is safe on that target, as there is no red zone. This is currently
569 broken AFAIK, although it's not a target I can exercise.
570 The following demonstrates the problem:
571 extern void bar(char *p);
575 __asm__("" ::: "cr2");
578 //===-------------------------------------------------------------------------===
579 Naming convention for instruction formats is very haphazard.
580 We have agreed on a naming scheme as follows:
582 <INST_form>{_<OP_type><OP_len>}+
585 INST_form is the instruction format (X-form, etc.)
586 OP_type is the operand type - one of OPC (opcode), RD (register destination),
587 RS (register source),
588 RDp (destination register pair),
589 RSp (source register pair), IM (immediate),
591 OP_len is the length of the operand in bits
593 VSX register operands would be of length 6 (split across two fields),
594 condition register fields of length 3.
595 We would not need denote reserved fields in names of instruction formats.
597 //===----------------------------------------------------------------------===//
599 Instruction fusion was introduced in ISA 2.06 and more opportunities added in
600 ISA 2.07. LLVM needs to add infrastructure to recognize fusion opportunities
601 and force instruction pairs to be scheduled together.
603 -----------------------------------------------------------------------------
605 More general handling of any_extend and zero_extend:
607 See https://reviews.llvm.org/D24924#555306