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 ===-------------------------------------------------------------------------===
123 We still generate calls to foo$stub, and stubs, on Darwin. This is not
124 necessary when building with the Leopard (10.5) or later linker, as stubs are
125 generated by ld when necessary. Parameterizing this based on the deployment
126 target (-mmacosx-version-min) is probably enough. x86-32 does this right, see
129 ===-------------------------------------------------------------------------===
131 Darwin Stub LICM optimization:
137 Have to go through an indirect stub if bar is external or linkonce. It would
138 be better to compile it as:
143 which only computes the address of bar once (instead of each time through the
144 stub). This is Darwin specific and would have to be done in the code generator.
145 Probably not a win on x86.
147 ===-------------------------------------------------------------------------===
149 Simple IPO for argument passing, change:
150 void foo(int X, double Y, int Z) -> void foo(int X, int Z, double Y)
152 the Darwin ABI specifies that any integer arguments in the first 32 bytes worth
153 of arguments get assigned to r3 through r10. That is, if you have a function
154 foo(int, double, int) you get r3, f1, r6, since the 64 bit double ate up the
155 argument bytes for r4 and r5. The trick then would be to shuffle the argument
156 order for functions we can internalize so that the maximum number of
157 integers/pointers get passed in regs before you see any of the fp arguments.
159 Instead of implementing this, it would actually probably be easier to just
160 implement a PPC fastcc, where we could do whatever we wanted to the CC,
161 including having this work sanely.
163 ===-------------------------------------------------------------------------===
165 Fix Darwin FP-In-Integer Registers ABI
167 Darwin passes doubles in structures in integer registers, which is very very
168 bad. Add something like a BITCAST to LLVM, then do an i-p transformation that
169 percolates these things out of functions.
171 Check out how horrible this is:
172 http://gcc.gnu.org/ml/gcc/2005-10/msg01036.html
174 This is an extension of "interprocedural CC unmunging" that can't be done with
177 ===-------------------------------------------------------------------------===
179 Fold add and sub with constant into non-extern, non-weak addresses so this:
182 void bar(int b) { a = b; }
183 void foo(unsigned char *c) {
200 lbz r2, lo16(_a+3)(r2)
204 ===-------------------------------------------------------------------------===
206 We should compile these two functions to the same thing:
209 void f(int a, int b, int *P) {
210 *P = (a-b)>=0?(a-b):(b-a);
212 void g(int a, int b, int *P) {
216 Further, they should compile to something better than:
222 bgt cr0, LBB2_2 ; entry
239 ... which is much nicer.
241 This theoretically may help improve twolf slightly (used in dimbox.c:142?).
243 ===-------------------------------------------------------------------------===
246 define i32 @clamp0g(i32 %a) {
248 %cmp = icmp slt i32 %a, 0
249 %sel = select i1 %cmp, i32 0, i32 %a
253 Is compile to this with the PowerPC (32-bit) backend:
265 This could be reduced to the much simpler:
272 ===-------------------------------------------------------------------------===
274 int foo(int N, int ***W, int **TK, int X) {
277 for (t = 0; t < N; ++t)
278 for (i = 0; i < 4; ++i)
279 W[t / X][i][t % X] = TK[i][t];
284 We generate relatively atrocious code for this loop compared to gcc.
286 We could also strength reduce the rem and the div:
287 http://www.lcs.mit.edu/pubs/pdf/MIT-LCS-TM-600.pdf
289 ===-------------------------------------------------------------------------===
291 We generate ugly code for this:
293 void func(unsigned int *ret, float dx, float dy, float dz, float dw) {
295 if(dx < -dw) code |= 1;
296 if(dx > dw) code |= 2;
297 if(dy < -dw) code |= 4;
298 if(dy > dw) code |= 8;
299 if(dz < -dw) code |= 16;
300 if(dz > dw) code |= 32;
304 ===-------------------------------------------------------------------------===
306 %struct.B = type { i8, [3 x i8] }
308 define void @bar(%struct.B* %b) {
310 %tmp = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
311 %tmp = load i32* %tmp ; <uint> [#uses=1]
312 %tmp3 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
313 %tmp4 = load i32* %tmp3 ; <uint> [#uses=1]
314 %tmp8 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=2]
315 %tmp9 = load i32* %tmp8 ; <uint> [#uses=1]
316 %tmp4.mask17 = shl i32 %tmp4, i8 1 ; <uint> [#uses=1]
317 %tmp1415 = and i32 %tmp4.mask17, 2147483648 ; <uint> [#uses=1]
318 %tmp.masked = and i32 %tmp, 2147483648 ; <uint> [#uses=1]
319 %tmp11 = or i32 %tmp1415, %tmp.masked ; <uint> [#uses=1]
320 %tmp12 = and i32 %tmp9, 2147483647 ; <uint> [#uses=1]
321 %tmp13 = or i32 %tmp12, %tmp11 ; <uint> [#uses=1]
322 store i32 %tmp13, i32* %tmp8
332 rlwimi r2, r4, 0, 0, 0
336 We could collapse a bunch of those ORs and ANDs and generate the following
341 rlwinm r4, r2, 1, 0, 0
346 ===-------------------------------------------------------------------------===
348 Consider a function like this:
350 float foo(float X) { return X + 1234.4123f; }
352 The FP constant ends up in the constant pool, so we need to get the LR register.
353 This ends up producing code like this:
362 addis r2, r2, ha16(.CPI_foo_0-"L00000$pb")
363 lfs f0, lo16(.CPI_foo_0-"L00000$pb")(r2)
369 This is functional, but there is no reason to spill the LR register all the way
370 to the stack (the two marked instrs): spilling it to a GPR is quite enough.
372 Implementing this will require some codegen improvements. Nate writes:
374 "So basically what we need to support the "no stack frame save and restore" is a
375 generalization of the LR optimization to "callee-save regs".
377 Currently, we have LR marked as a callee-save reg. The register allocator sees
378 that it's callee save, and spills it directly to the stack.
380 Ideally, something like this would happen:
382 LR would be in a separate register class from the GPRs. The class of LR would be
383 marked "unspillable". When the register allocator came across an unspillable
384 reg, it would ask "what is the best class to copy this into that I *can* spill"
385 If it gets a class back, which it will in this case (the gprs), it grabs a free
386 register of that class. If it is then later necessary to spill that reg, so be
389 ===-------------------------------------------------------------------------===
393 return X ? 524288 : 0;
401 beq cr0, LBB1_2 ;entry
414 This sort of thing occurs a lot due to globalopt.
416 ===-------------------------------------------------------------------------===
420 define i32 @bar(i32 %x) nounwind readnone ssp {
422 %0 = icmp eq i32 %x, 0 ; <i1> [#uses=1]
423 %neg = sext i1 %0 to i32 ; <i32> [#uses=1]
435 it would be better to produce:
442 ===-------------------------------------------------------------------------===
444 We generate horrible ppc code for this:
456 addi r5, r5, 1 ;; Extra IV for the exit value compare.
460 xoris r6, r5, 30 ;; This is due to a large immediate.
461 cmplwi cr0, r6, 33920
464 //===---------------------------------------------------------------------===//
468 inline std::pair<unsigned, bool> full_add(unsigned a, unsigned b)
469 { return std::make_pair(a + b, a + b < a); }
470 bool no_overflow(unsigned a, unsigned b)
471 { return !full_add(a, b).second; }
488 rlwinm r2, r2, 29, 31, 31
492 //===---------------------------------------------------------------------===//
494 We compile some FP comparisons into an mfcr with two rlwinms and an or. For
497 int test(double x, double y) { return islessequal(x, y);}
498 int test2(double x, double y) { return islessgreater(x, y);}
499 int test3(double x, double y) { return !islessequal(x, y);}
501 Compiles into (all three are similar, but the bits differ):
506 rlwinm r3, r2, 29, 31, 31
507 rlwinm r2, r2, 31, 31, 31
511 GCC compiles this into:
520 which is more efficient and can use mfocr. See PR642 for some more context.
522 //===---------------------------------------------------------------------===//
524 void foo(float *data, float d) {
526 for (i = 0; i < 8000; i++)
529 void foo2(float *data, float d) {
532 for (i = 0; i < 8000; i++) {
545 cmplwi cr0, r4, 32000
554 cmplwi cr0, r4, 32000
559 The 'mr' could be eliminated to folding the add into the cmp better.
561 //===---------------------------------------------------------------------===//
562 Codegen for the following (low-probability) case deteriorated considerably
563 when the correctness fixes for unordered comparisons went in (PR 642, 58871).
564 It should be possible to recover the code quality described in the comments.
566 ; RUN: llvm-as < %s | llc -march=ppc32 | grep or | count 3
567 ; This should produce one 'or' or 'cror' instruction per function.
569 ; RUN: llvm-as < %s | llc -march=ppc32 | grep mfcr | count 3
572 define i32 @test(double %x, double %y) nounwind {
574 %tmp3 = fcmp ole double %x, %y ; <i1> [#uses=1]
575 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
579 define i32 @test2(double %x, double %y) nounwind {
581 %tmp3 = fcmp one double %x, %y ; <i1> [#uses=1]
582 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
586 define i32 @test3(double %x, double %y) nounwind {
588 %tmp3 = fcmp ugt double %x, %y ; <i1> [#uses=1]
589 %tmp34 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
593 //===---------------------------------------------------------------------===//
594 for the following code:
596 void foo (float *__restrict__ a, int *__restrict__ b, int n) {
600 we load b[n] to GPR, then move it VSX register and convert it float. We should
601 use vsx scalar integer load instructions to avoid direct moves
603 //===----------------------------------------------------------------------===//
604 ; RUN: llvm-as < %s | llc -march=ppc32 | not grep fneg
606 ; This could generate FSEL with appropriate flags (FSEL is not IEEE-safe, and
607 ; should not be generated except with -enable-finite-only-fp-math or the like).
608 ; With the correctness fixes for PR642 (58871) LowerSELECT_CC would need to
609 ; recognize a more elaborate tree than a simple SETxx.
611 define double @test_FNEG_sel(double %A, double %B, double %C) {
612 %D = fsub double -0.000000e+00, %A ; <double> [#uses=1]
613 %Cond = fcmp ugt double %D, -0.000000e+00 ; <i1> [#uses=1]
614 %E = select i1 %Cond, double %B, double %C ; <double> [#uses=1]
618 //===----------------------------------------------------------------------===//
619 The save/restore sequence for CR in prolog/epilog is terrible:
620 - Each CR subreg is saved individually, rather than doing one save as a unit.
621 - On Darwin, the save is done after the decrement of SP, which means the offset
622 from SP of the save slot can be too big for a store instruction, which means we
623 need an additional register (currently hacked in 96015+96020; the solution there
624 is correct, but poor).
625 - On SVR4 the same thing can happen, and I don't think saving before the SP
626 decrement is safe on that target, as there is no red zone. This is currently
627 broken AFAIK, although it's not a target I can exercise.
628 The following demonstrates the problem:
629 extern void bar(char *p);
633 __asm__("" ::: "cr2");
636 //===-------------------------------------------------------------------------===
637 Naming convention for instruction formats is very haphazard.
638 We have agreed on a naming scheme as follows:
640 <INST_form>{_<OP_type><OP_len>}+
643 INST_form is the instruction format (X-form, etc.)
644 OP_type is the operand type - one of OPC (opcode), RD (register destination),
645 RS (register source),
646 RDp (destination register pair),
647 RSp (source register pair), IM (immediate),
649 OP_len is the length of the operand in bits
651 VSX register operands would be of length 6 (split across two fields),
652 condition register fields of length 3.
653 We would not need denote reserved fields in names of instruction formats.
655 //===----------------------------------------------------------------------===//
657 Instruction fusion was introduced in ISA 2.06 and more opportunities added in
658 ISA 2.07. LLVM needs to add infrastructure to recognize fusion opportunities
659 and force instruction pairs to be scheduled together.
661 -----------------------------------------------------------------------------
663 More general handling of any_extend and zero_extend:
665 See https://reviews.llvm.org/D24924#555306