1 //===- README.txt - Notes for improving PowerPC-specific code gen ---------===//
5 * implement do-loop -> bdnz transform
6 * lmw/stmw pass a la arm load store optimizer for prolog/epilog
8 ===-------------------------------------------------------------------------===
12 long f2 (long x) { return 0xfffffff000000000UL; }
13 long f3 (long x) { return 0x1ffffffffUL; }
40 ===-------------------------------------------------------------------------===
44 unsigned add32carry(unsigned sum, unsigned x) {
51 Should compile to something like:
61 rlwinm r4, r4, 29, 31, 31
66 ===-------------------------------------------------------------------------===
68 Support 'update' load/store instructions. These are cracked on the G5, but are
71 With preinc enabled, this:
73 long *%test4(long *%X, long *%dest) {
74 %Y = getelementptr long* %X, int 4
76 store long %A, long* %dest
91 with -sched=list-burr, I get:
100 ===-------------------------------------------------------------------------===
102 We compile the hottest inner loop of viterbi to:
113 bne cr0, LBB1_83 ;bb420.i
115 The CBE manages to produce:
126 This could be much better (bdnz instead of bdz) but it still beats us. If we
127 produced this with bdnz, the loop would be a single dispatch group.
129 ===-------------------------------------------------------------------------===
146 This is effectively a simple form of predication.
148 ===-------------------------------------------------------------------------===
150 Lump the constant pool for each function into ONE pic object, and reference
151 pieces of it as offsets from the start. For functions like this (contrived
152 to have lots of constants obviously):
154 double X(double Y) { return (Y*1.23 + 4.512)*2.34 + 14.38; }
159 lis r2, ha16(.CPI_X_0)
160 lfd f0, lo16(.CPI_X_0)(r2)
161 lis r2, ha16(.CPI_X_1)
162 lfd f2, lo16(.CPI_X_1)(r2)
164 lis r2, ha16(.CPI_X_2)
165 lfd f1, lo16(.CPI_X_2)(r2)
166 lis r2, ha16(.CPI_X_3)
167 lfd f2, lo16(.CPI_X_3)(r2)
171 It would be better to materialize .CPI_X into a register, then use immediates
172 off of the register to avoid the lis's. This is even more important in PIC
175 Note that this (and the static variable version) is discussed here for GCC:
176 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
178 Here's another example (the sgn function):
179 double testf(double a) {
180 return a == 0.0 ? 0.0 : (a > 0.0 ? 1.0 : -1.0);
183 it produces a BB like this:
185 lis r2, ha16(LCPI1_0)
186 lfs f0, lo16(LCPI1_0)(r2)
187 lis r2, ha16(LCPI1_1)
188 lis r3, ha16(LCPI1_2)
189 lfs f2, lo16(LCPI1_2)(r3)
190 lfs f3, lo16(LCPI1_1)(r2)
195 ===-------------------------------------------------------------------------===
197 PIC Code Gen IPO optimization:
199 Squish small scalar globals together into a single global struct, allowing the
200 address of the struct to be CSE'd, avoiding PIC accesses (also reduces the size
201 of the GOT on targets with one).
203 Note that this is discussed here for GCC:
204 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
206 ===-------------------------------------------------------------------------===
208 Implement Newton-Rhapson method for improving estimate instructions to the
209 correct accuracy, and implementing divide as multiply by reciprocal when it has
210 more than one use. Itanium would want this too.
212 ===-------------------------------------------------------------------------===
214 Compile offsets from allocas:
217 %X = alloca { int, int }
218 %Y = getelementptr {int,int}* %X, int 0, uint 1
222 into a single add, not two:
229 --> important for C++.
231 ===-------------------------------------------------------------------------===
233 No loads or stores of the constants should be needed:
235 struct foo { double X, Y; };
236 void xxx(struct foo F);
237 void bar() { struct foo R = { 1.0, 2.0 }; xxx(R); }
239 ===-------------------------------------------------------------------------===
243 We still generate calls to foo$stub, and stubs, on Darwin. This is not
244 necessary when building with the Leopard (10.5) or later linker, as stubs are
245 generated by ld when necessary. Parameterizing this based on the deployment
246 target (-mmacosx-version-min) is probably enough. x86-32 does this right, see
249 ===-------------------------------------------------------------------------===
251 Darwin Stub LICM optimization:
257 Have to go through an indirect stub if bar is external or linkonce. It would
258 be better to compile it as:
263 which only computes the address of bar once (instead of each time through the
264 stub). This is Darwin specific and would have to be done in the code generator.
265 Probably not a win on x86.
267 ===-------------------------------------------------------------------------===
269 Simple IPO for argument passing, change:
270 void foo(int X, double Y, int Z) -> void foo(int X, int Z, double Y)
272 the Darwin ABI specifies that any integer arguments in the first 32 bytes worth
273 of arguments get assigned to r3 through r10. That is, if you have a function
274 foo(int, double, int) you get r3, f1, r6, since the 64 bit double ate up the
275 argument bytes for r4 and r5. The trick then would be to shuffle the argument
276 order for functions we can internalize so that the maximum number of
277 integers/pointers get passed in regs before you see any of the fp arguments.
279 Instead of implementing this, it would actually probably be easier to just
280 implement a PPC fastcc, where we could do whatever we wanted to the CC,
281 including having this work sanely.
283 ===-------------------------------------------------------------------------===
285 Fix Darwin FP-In-Integer Registers ABI
287 Darwin passes doubles in structures in integer registers, which is very very
288 bad. Add something like a BITCAST to LLVM, then do an i-p transformation that
289 percolates these things out of functions.
291 Check out how horrible this is:
292 http://gcc.gnu.org/ml/gcc/2005-10/msg01036.html
294 This is an extension of "interprocedural CC unmunging" that can't be done with
297 ===-------------------------------------------------------------------------===
304 return b * 3; // ignore the fact that this is always 3.
310 into something not this:
315 rlwinm r2, r2, 29, 31, 31
317 bgt cr0, LBB1_2 ; UnifiedReturnBlock
319 rlwinm r2, r2, 0, 31, 31
322 LBB1_2: ; UnifiedReturnBlock
326 In particular, the two compares (marked 1) could be shared by reversing one.
327 This could be done in the dag combiner, by swapping a BR_CC when a SETCC of the
328 same operands (but backwards) exists. In this case, this wouldn't save us
329 anything though, because the compares still wouldn't be shared.
331 ===-------------------------------------------------------------------------===
333 We should custom expand setcc instead of pretending that we have it. That
334 would allow us to expose the access of the crbit after the mfcr, allowing
335 that access to be trivially folded into other ops. A simple example:
337 int foo(int a, int b) { return (a < b) << 4; }
344 rlwinm r2, r2, 29, 31, 31
348 ===-------------------------------------------------------------------------===
350 Fold add and sub with constant into non-extern, non-weak addresses so this:
353 void bar(int b) { a = b; }
354 void foo(unsigned char *c) {
371 lbz r2, lo16(_a+3)(r2)
375 ===-------------------------------------------------------------------------===
377 We generate really bad code for this:
379 int f(signed char *a, _Bool b, _Bool c) {
385 ===-------------------------------------------------------------------------===
388 int test(unsigned *P) { return *P >> 24; }
403 ===-------------------------------------------------------------------------===
405 On the G5, logical CR operations are more expensive in their three
406 address form: ops that read/write the same register are half as expensive as
407 those that read from two registers that are different from their destination.
409 We should model this with two separate instructions. The isel should generate
410 the "two address" form of the instructions. When the register allocator
411 detects that it needs to insert a copy due to the two-addresness of the CR
412 logical op, it will invoke PPCInstrInfo::convertToThreeAddress. At this point
413 we can convert to the "three address" instruction, to save code space.
415 This only matters when we start generating cr logical ops.
417 ===-------------------------------------------------------------------------===
419 We should compile these two functions to the same thing:
422 void f(int a, int b, int *P) {
423 *P = (a-b)>=0?(a-b):(b-a);
425 void g(int a, int b, int *P) {
429 Further, they should compile to something better than:
435 bgt cr0, LBB2_2 ; entry
452 ... which is much nicer.
454 This theoretically may help improve twolf slightly (used in dimbox.c:142?).
456 ===-------------------------------------------------------------------------===
459 define i32 @clamp0g(i32 %a) {
461 %cmp = icmp slt i32 %a, 0
462 %sel = select i1 %cmp, i32 0, i32 %a
466 Is compile to this with the PowerPC (32-bit) backend:
478 This could be reduced to the much simpler:
485 ===-------------------------------------------------------------------------===
487 int foo(int N, int ***W, int **TK, int X) {
490 for (t = 0; t < N; ++t)
491 for (i = 0; i < 4; ++i)
492 W[t / X][i][t % X] = TK[i][t];
497 We generate relatively atrocious code for this loop compared to gcc.
499 We could also strength reduce the rem and the div:
500 http://www.lcs.mit.edu/pubs/pdf/MIT-LCS-TM-600.pdf
502 ===-------------------------------------------------------------------------===
504 float foo(float X) { return (int)(X); }
519 We could use a target dag combine to turn the lwz/extsw into an lwa when the
520 lwz has a single use. Since LWA is cracked anyway, this would be a codesize
523 ===-------------------------------------------------------------------------===
525 We generate ugly code for this:
527 void func(unsigned int *ret, float dx, float dy, float dz, float dw) {
529 if(dx < -dw) code |= 1;
530 if(dx > dw) code |= 2;
531 if(dy < -dw) code |= 4;
532 if(dy > dw) code |= 8;
533 if(dz < -dw) code |= 16;
534 if(dz > dw) code |= 32;
538 ===-------------------------------------------------------------------------===
540 Complete the signed i32 to FP conversion code using 64-bit registers
541 transformation, good for PI. See PPCISelLowering.cpp, this comment:
543 // FIXME: disable this lowered code. This generates 64-bit register values,
544 // and we don't model the fact that the top part is clobbered by calls. We
545 // need to flag these together so that the value isn't live across a call.
546 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
548 Also, if the registers are spilled to the stack, we have to ensure that all
549 64-bits of them are save/restored, otherwise we will miscompile the code. It
550 sounds like we need to get the 64-bit register classes going.
552 ===-------------------------------------------------------------------------===
554 %struct.B = type { i8, [3 x i8] }
556 define void @bar(%struct.B* %b) {
558 %tmp = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
559 %tmp = load i32* %tmp ; <uint> [#uses=1]
560 %tmp3 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
561 %tmp4 = load i32* %tmp3 ; <uint> [#uses=1]
562 %tmp8 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=2]
563 %tmp9 = load i32* %tmp8 ; <uint> [#uses=1]
564 %tmp4.mask17 = shl i32 %tmp4, i8 1 ; <uint> [#uses=1]
565 %tmp1415 = and i32 %tmp4.mask17, 2147483648 ; <uint> [#uses=1]
566 %tmp.masked = and i32 %tmp, 2147483648 ; <uint> [#uses=1]
567 %tmp11 = or i32 %tmp1415, %tmp.masked ; <uint> [#uses=1]
568 %tmp12 = and i32 %tmp9, 2147483647 ; <uint> [#uses=1]
569 %tmp13 = or i32 %tmp12, %tmp11 ; <uint> [#uses=1]
570 store i32 %tmp13, i32* %tmp8
580 rlwimi r2, r4, 0, 0, 0
584 We could collapse a bunch of those ORs and ANDs and generate the following
589 rlwinm r4, r2, 1, 0, 0
594 ===-------------------------------------------------------------------------===
598 unsigned test6(unsigned x) {
599 return ((x & 0x00FF0000) >> 16) | ((x & 0x000000FF) << 16);
606 rlwinm r3, r3, 16, 0, 31
615 rlwinm r3,r3,16,24,31
620 ===-------------------------------------------------------------------------===
622 Consider a function like this:
624 float foo(float X) { return X + 1234.4123f; }
626 The FP constant ends up in the constant pool, so we need to get the LR register.
627 This ends up producing code like this:
636 addis r2, r2, ha16(.CPI_foo_0-"L00000$pb")
637 lfs f0, lo16(.CPI_foo_0-"L00000$pb")(r2)
643 This is functional, but there is no reason to spill the LR register all the way
644 to the stack (the two marked instrs): spilling it to a GPR is quite enough.
646 Implementing this will require some codegen improvements. Nate writes:
648 "So basically what we need to support the "no stack frame save and restore" is a
649 generalization of the LR optimization to "callee-save regs".
651 Currently, we have LR marked as a callee-save reg. The register allocator sees
652 that it's callee save, and spills it directly to the stack.
654 Ideally, something like this would happen:
656 LR would be in a separate register class from the GPRs. The class of LR would be
657 marked "unspillable". When the register allocator came across an unspillable
658 reg, it would ask "what is the best class to copy this into that I *can* spill"
659 If it gets a class back, which it will in this case (the gprs), it grabs a free
660 register of that class. If it is then later necessary to spill that reg, so be
663 ===-------------------------------------------------------------------------===
667 return X ? 524288 : 0;
675 beq cr0, LBB1_2 ;entry
688 This sort of thing occurs a lot due to globalopt.
690 ===-------------------------------------------------------------------------===
694 define i32 @bar(i32 %x) nounwind readnone ssp {
696 %0 = icmp eq i32 %x, 0 ; <i1> [#uses=1]
697 %neg = sext i1 %0 to i32 ; <i32> [#uses=1]
709 it would be better to produce:
716 ===-------------------------------------------------------------------------===
718 We currently compile 32-bit bswap:
720 declare i32 @llvm.bswap.i32(i32 %A)
721 define i32 @test(i32 %A) {
722 %B = call i32 @llvm.bswap.i32(i32 %A)
729 rlwinm r2, r3, 24, 16, 23
731 rlwimi r2, r3, 8, 24, 31
732 rlwimi r4, r3, 8, 8, 15
733 rlwimi r4, r2, 0, 16, 31
737 it would be more efficient to produce:
740 rlwinm r3,r3,8,0xffffffff
742 rlwimi r3,r0,24,16,23
745 ===-------------------------------------------------------------------------===
747 test/CodeGen/PowerPC/2007-03-24-cntlzd.ll compiles to:
749 __ZNK4llvm5APInt17countLeadingZerosEv:
752 or r2, r2, r2 <<-- silly.
756 The dead or is a 'truncate' from 64- to 32-bits.
758 ===-------------------------------------------------------------------------===
760 We generate horrible ppc code for this:
772 addi r5, r5, 1 ;; Extra IV for the exit value compare.
776 xoris r6, r5, 30 ;; This is due to a large immediate.
777 cmplwi cr0, r6, 33920
780 //===---------------------------------------------------------------------===//
784 inline std::pair<unsigned, bool> full_add(unsigned a, unsigned b)
785 { return std::make_pair(a + b, a + b < a); }
786 bool no_overflow(unsigned a, unsigned b)
787 { return !full_add(a, b).second; }
804 rlwinm r2, r2, 29, 31, 31
808 //===---------------------------------------------------------------------===//
810 We compile some FP comparisons into an mfcr with two rlwinms and an or. For
813 int test(double x, double y) { return islessequal(x, y);}
814 int test2(double x, double y) { return islessgreater(x, y);}
815 int test3(double x, double y) { return !islessequal(x, y);}
817 Compiles into (all three are similar, but the bits differ):
822 rlwinm r3, r2, 29, 31, 31
823 rlwinm r2, r2, 31, 31, 31
827 GCC compiles this into:
836 which is more efficient and can use mfocr. See PR642 for some more context.
838 //===---------------------------------------------------------------------===//
840 void foo(float *data, float d) {
842 for (i = 0; i < 8000; i++)
845 void foo2(float *data, float d) {
848 for (i = 0; i < 8000; i++) {
861 cmplwi cr0, r4, 32000
870 cmplwi cr0, r4, 32000
875 The 'mr' could be eliminated to folding the add into the cmp better.
877 //===---------------------------------------------------------------------===//
878 Codegen for the following (low-probability) case deteriorated considerably
879 when the correctness fixes for unordered comparisons went in (PR 642, 58871).
880 It should be possible to recover the code quality described in the comments.
882 ; RUN: llvm-as < %s | llc -march=ppc32 | grep or | count 3
883 ; This should produce one 'or' or 'cror' instruction per function.
885 ; RUN: llvm-as < %s | llc -march=ppc32 | grep mfcr | count 3
888 define i32 @test(double %x, double %y) nounwind {
890 %tmp3 = fcmp ole double %x, %y ; <i1> [#uses=1]
891 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
895 define i32 @test2(double %x, double %y) nounwind {
897 %tmp3 = fcmp one double %x, %y ; <i1> [#uses=1]
898 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
902 define i32 @test3(double %x, double %y) nounwind {
904 %tmp3 = fcmp ugt double %x, %y ; <i1> [#uses=1]
905 %tmp34 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
908 //===----------------------------------------------------------------------===//
909 ; RUN: llvm-as < %s | llc -march=ppc32 | not grep fneg
911 ; This could generate FSEL with appropriate flags (FSEL is not IEEE-safe, and
912 ; should not be generated except with -enable-finite-only-fp-math or the like).
913 ; With the correctness fixes for PR642 (58871) LowerSELECT_CC would need to
914 ; recognize a more elaborate tree than a simple SETxx.
916 define double @test_FNEG_sel(double %A, double %B, double %C) {
917 %D = fsub double -0.000000e+00, %A ; <double> [#uses=1]
918 %Cond = fcmp ugt double %D, -0.000000e+00 ; <i1> [#uses=1]
919 %E = select i1 %Cond, double %B, double %C ; <double> [#uses=1]
923 //===----------------------------------------------------------------------===//
924 The save/restore sequence for CR in prolog/epilog is terrible:
925 - Each CR subreg is saved individually, rather than doing one save as a unit.
926 - On Darwin, the save is done after the decrement of SP, which means the offset
927 from SP of the save slot can be too big for a store instruction, which means we
928 need an additional register (currently hacked in 96015+96020; the solution there
929 is correct, but poor).
930 - On SVR4 the same thing can happen, and I don't think saving before the SP
931 decrement is safe on that target, as there is no red zone. This is currently
932 broken AFAIK, although it's not a target I can exercise.
933 The following demonstrates the problem:
934 extern void bar(char *p);
938 __asm__("" ::: "cr2");