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 ===-------------------------------------------------------------------------===
10 Support 'update' load/store instructions. These are cracked on the G5, but are
13 With preinc enabled, this:
15 long *%test4(long *%X, long *%dest) {
16 %Y = getelementptr long* %X, int 4
18 store long %A, long* %dest
33 with -sched=list-burr, I get:
42 ===-------------------------------------------------------------------------===
44 We compile the hottest inner loop of viterbi to:
55 bne cr0, LBB1_83 ;bb420.i
57 The CBE manages to produce:
68 This could be much better (bdnz instead of bdz) but it still beats us. If we
69 produced this with bdnz, the loop would be a single dispatch group.
71 ===-------------------------------------------------------------------------===
88 This is effectively a simple form of predication.
90 ===-------------------------------------------------------------------------===
92 Lump the constant pool for each function into ONE pic object, and reference
93 pieces of it as offsets from the start. For functions like this (contrived
94 to have lots of constants obviously):
96 double X(double Y) { return (Y*1.23 + 4.512)*2.34 + 14.38; }
101 lis r2, ha16(.CPI_X_0)
102 lfd f0, lo16(.CPI_X_0)(r2)
103 lis r2, ha16(.CPI_X_1)
104 lfd f2, lo16(.CPI_X_1)(r2)
106 lis r2, ha16(.CPI_X_2)
107 lfd f1, lo16(.CPI_X_2)(r2)
108 lis r2, ha16(.CPI_X_3)
109 lfd f2, lo16(.CPI_X_3)(r2)
113 It would be better to materialize .CPI_X into a register, then use immediates
114 off of the register to avoid the lis's. This is even more important in PIC
117 Note that this (and the static variable version) is discussed here for GCC:
118 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
120 Here's another example (the sgn function):
121 double testf(double a) {
122 return a == 0.0 ? 0.0 : (a > 0.0 ? 1.0 : -1.0);
125 it produces a BB like this:
127 lis r2, ha16(LCPI1_0)
128 lfs f0, lo16(LCPI1_0)(r2)
129 lis r2, ha16(LCPI1_1)
130 lis r3, ha16(LCPI1_2)
131 lfs f2, lo16(LCPI1_2)(r3)
132 lfs f3, lo16(LCPI1_1)(r2)
137 ===-------------------------------------------------------------------------===
139 PIC Code Gen IPO optimization:
141 Squish small scalar globals together into a single global struct, allowing the
142 address of the struct to be CSE'd, avoiding PIC accesses (also reduces the size
143 of the GOT on targets with one).
145 Note that this is discussed here for GCC:
146 http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html
148 ===-------------------------------------------------------------------------===
150 Implement Newton-Rhapson method for improving estimate instructions to the
151 correct accuracy, and implementing divide as multiply by reciprocal when it has
152 more than one use. Itanium will want this too.
154 ===-------------------------------------------------------------------------===
156 Compile offsets from allocas:
159 %X = alloca { int, int }
160 %Y = getelementptr {int,int}* %X, int 0, uint 1
164 into a single add, not two:
171 --> important for C++.
173 ===-------------------------------------------------------------------------===
175 No loads or stores of the constants should be needed:
177 struct foo { double X, Y; };
178 void xxx(struct foo F);
179 void bar() { struct foo R = { 1.0, 2.0 }; xxx(R); }
181 ===-------------------------------------------------------------------------===
183 Darwin Stub LICM optimization:
189 Have to go through an indirect stub if bar is external or linkonce. It would
190 be better to compile it as:
195 which only computes the address of bar once (instead of each time through the
196 stub). This is Darwin specific and would have to be done in the code generator.
197 Probably not a win on x86.
199 ===-------------------------------------------------------------------------===
201 Simple IPO for argument passing, change:
202 void foo(int X, double Y, int Z) -> void foo(int X, int Z, double Y)
204 the Darwin ABI specifies that any integer arguments in the first 32 bytes worth
205 of arguments get assigned to r3 through r10. That is, if you have a function
206 foo(int, double, int) you get r3, f1, r6, since the 64 bit double ate up the
207 argument bytes for r4 and r5. The trick then would be to shuffle the argument
208 order for functions we can internalize so that the maximum number of
209 integers/pointers get passed in regs before you see any of the fp arguments.
211 Instead of implementing this, it would actually probably be easier to just
212 implement a PPC fastcc, where we could do whatever we wanted to the CC,
213 including having this work sanely.
215 ===-------------------------------------------------------------------------===
217 Fix Darwin FP-In-Integer Registers ABI
219 Darwin passes doubles in structures in integer registers, which is very very
220 bad. Add something like a BIT_CONVERT to LLVM, then do an i-p transformation
221 that percolates these things out of functions.
223 Check out how horrible this is:
224 http://gcc.gnu.org/ml/gcc/2005-10/msg01036.html
226 This is an extension of "interprocedural CC unmunging" that can't be done with
229 ===-------------------------------------------------------------------------===
236 return b * 3; // ignore the fact that this is always 3.
242 into something not this:
247 rlwinm r2, r2, 29, 31, 31
249 bgt cr0, LBB1_2 ; UnifiedReturnBlock
251 rlwinm r2, r2, 0, 31, 31
254 LBB1_2: ; UnifiedReturnBlock
258 In particular, the two compares (marked 1) could be shared by reversing one.
259 This could be done in the dag combiner, by swapping a BR_CC when a SETCC of the
260 same operands (but backwards) exists. In this case, this wouldn't save us
261 anything though, because the compares still wouldn't be shared.
263 ===-------------------------------------------------------------------------===
265 We should custom expand setcc instead of pretending that we have it. That
266 would allow us to expose the access of the crbit after the mfcr, allowing
267 that access to be trivially folded into other ops. A simple example:
269 int foo(int a, int b) { return (a < b) << 4; }
276 rlwinm r2, r2, 29, 31, 31
280 ===-------------------------------------------------------------------------===
282 Fold add and sub with constant into non-extern, non-weak addresses so this:
285 void bar(int b) { a = b; }
286 void foo(unsigned char *c) {
303 lbz r2, lo16(_a+3)(r2)
307 ===-------------------------------------------------------------------------===
309 We generate really bad code for this:
311 int f(signed char *a, _Bool b, _Bool c) {
317 ===-------------------------------------------------------------------------===
320 int test(unsigned *P) { return *P >> 24; }
335 ===-------------------------------------------------------------------------===
337 On the G5, logical CR operations are more expensive in their three
338 address form: ops that read/write the same register are half as expensive as
339 those that read from two registers that are different from their destination.
341 We should model this with two separate instructions. The isel should generate
342 the "two address" form of the instructions. When the register allocator
343 detects that it needs to insert a copy due to the two-addresness of the CR
344 logical op, it will invoke PPCInstrInfo::convertToThreeAddress. At this point
345 we can convert to the "three address" instruction, to save code space.
347 This only matters when we start generating cr logical ops.
349 ===-------------------------------------------------------------------------===
351 We should compile these two functions to the same thing:
354 void f(int a, int b, int *P) {
355 *P = (a-b)>=0?(a-b):(b-a);
357 void g(int a, int b, int *P) {
361 Further, they should compile to something better than:
367 bgt cr0, LBB2_2 ; entry
384 ... which is much nicer.
386 This theoretically may help improve twolf slightly (used in dimbox.c:142?).
388 ===-------------------------------------------------------------------------===
390 int foo(int N, int ***W, int **TK, int X) {
393 for (t = 0; t < N; ++t)
394 for (i = 0; i < 4; ++i)
395 W[t / X][i][t % X] = TK[i][t];
400 We generate relatively atrocious code for this loop compared to gcc.
402 We could also strength reduce the rem and the div:
403 http://www.lcs.mit.edu/pubs/pdf/MIT-LCS-TM-600.pdf
405 ===-------------------------------------------------------------------------===
407 float foo(float X) { return (int)(X); }
422 We could use a target dag combine to turn the lwz/extsw into an lwa when the
423 lwz has a single use. Since LWA is cracked anyway, this would be a codesize
426 ===-------------------------------------------------------------------------===
428 We generate ugly code for this:
430 void func(unsigned int *ret, float dx, float dy, float dz, float dw) {
432 if(dx < -dw) code |= 1;
433 if(dx > dw) code |= 2;
434 if(dy < -dw) code |= 4;
435 if(dy > dw) code |= 8;
436 if(dz < -dw) code |= 16;
437 if(dz > dw) code |= 32;
441 ===-------------------------------------------------------------------------===
443 Complete the signed i32 to FP conversion code using 64-bit registers
444 transformation, good for PI. See PPCISelLowering.cpp, this comment:
446 // FIXME: disable this lowered code. This generates 64-bit register values,
447 // and we don't model the fact that the top part is clobbered by calls. We
448 // need to flag these together so that the value isn't live across a call.
449 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
451 Also, if the registers are spilled to the stack, we have to ensure that all
452 64-bits of them are save/restored, otherwise we will miscompile the code. It
453 sounds like we need to get the 64-bit register classes going.
455 ===-------------------------------------------------------------------------===
457 %struct.B = type { i8, [3 x i8] }
459 define void @bar(%struct.B* %b) {
461 %tmp = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
462 %tmp = load i32* %tmp ; <uint> [#uses=1]
463 %tmp3 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
464 %tmp4 = load i32* %tmp3 ; <uint> [#uses=1]
465 %tmp8 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=2]
466 %tmp9 = load i32* %tmp8 ; <uint> [#uses=1]
467 %tmp4.mask17 = shl i32 %tmp4, i8 1 ; <uint> [#uses=1]
468 %tmp1415 = and i32 %tmp4.mask17, 2147483648 ; <uint> [#uses=1]
469 %tmp.masked = and i32 %tmp, 2147483648 ; <uint> [#uses=1]
470 %tmp11 = or i32 %tmp1415, %tmp.masked ; <uint> [#uses=1]
471 %tmp12 = and i32 %tmp9, 2147483647 ; <uint> [#uses=1]
472 %tmp13 = or i32 %tmp12, %tmp11 ; <uint> [#uses=1]
473 store i32 %tmp13, i32* %tmp8
483 rlwimi r2, r4, 0, 0, 0
487 We could collapse a bunch of those ORs and ANDs and generate the following
492 rlwinm r4, r2, 1, 0, 0
497 ===-------------------------------------------------------------------------===
501 unsigned test6(unsigned x) {
502 return ((x & 0x00FF0000) >> 16) | ((x & 0x000000FF) << 16);
509 rlwinm r3, r3, 16, 0, 31
518 rlwinm r3,r3,16,24,31
523 ===-------------------------------------------------------------------------===
525 Consider a function like this:
527 float foo(float X) { return X + 1234.4123f; }
529 The FP constant ends up in the constant pool, so we need to get the LR register.
530 This ends up producing code like this:
539 addis r2, r2, ha16(.CPI_foo_0-"L00000$pb")
540 lfs f0, lo16(.CPI_foo_0-"L00000$pb")(r2)
546 This is functional, but there is no reason to spill the LR register all the way
547 to the stack (the two marked instrs): spilling it to a GPR is quite enough.
549 Implementing this will require some codegen improvements. Nate writes:
551 "So basically what we need to support the "no stack frame save and restore" is a
552 generalization of the LR optimization to "callee-save regs".
554 Currently, we have LR marked as a callee-save reg. The register allocator sees
555 that it's callee save, and spills it directly to the stack.
557 Ideally, something like this would happen:
559 LR would be in a separate register class from the GPRs. The class of LR would be
560 marked "unspillable". When the register allocator came across an unspillable
561 reg, it would ask "what is the best class to copy this into that I *can* spill"
562 If it gets a class back, which it will in this case (the gprs), it grabs a free
563 register of that class. If it is then later necessary to spill that reg, so be
566 ===-------------------------------------------------------------------------===
570 return X ? 524288 : 0;
578 beq cr0, LBB1_2 ;entry
591 This sort of thing occurs a lot due to globalopt.
593 ===-------------------------------------------------------------------------===
595 We currently compile 32-bit bswap:
597 declare i32 @llvm.bswap.i32(i32 %A)
598 define i32 @test(i32 %A) {
599 %B = call i32 @llvm.bswap.i32(i32 %A)
606 rlwinm r2, r3, 24, 16, 23
608 rlwimi r2, r3, 8, 24, 31
609 rlwimi r4, r3, 8, 8, 15
610 rlwimi r4, r2, 0, 16, 31
614 it would be more efficient to produce:
617 rlwinm r3,r3,8,0xffffffff
619 rlwimi r3,r0,24,16,23
622 ===-------------------------------------------------------------------------===
624 test/CodeGen/PowerPC/2007-03-24-cntlzd.ll compiles to:
626 __ZNK4llvm5APInt17countLeadingZerosEv:
629 or r2, r2, r2 <<-- silly.
633 The dead or is a 'truncate' from 64- to 32-bits.
635 ===-------------------------------------------------------------------------===
637 We generate horrible ppc code for this:
649 addi r5, r5, 1 ;; Extra IV for the exit value compare.
653 xoris r6, r5, 30 ;; This is due to a large immediate.
654 cmplwi cr0, r6, 33920
657 //===---------------------------------------------------------------------===//
661 inline std::pair<unsigned, bool> full_add(unsigned a, unsigned b)
662 { return std::make_pair(a + b, a + b < a); }
663 bool no_overflow(unsigned a, unsigned b)
664 { return !full_add(a, b).second; }
681 rlwinm r2, r2, 29, 31, 31
685 //===---------------------------------------------------------------------===//
687 We compile some FP comparisons into an mfcr with two rlwinms and an or. For
690 int test(double x, double y) { return islessequal(x, y);}
691 int test2(double x, double y) { return islessgreater(x, y);}
692 int test3(double x, double y) { return !islessequal(x, y);}
694 Compiles into (all three are similar, but the bits differ):
699 rlwinm r3, r2, 29, 31, 31
700 rlwinm r2, r2, 31, 31, 31
704 GCC compiles this into:
713 which is more efficient and can use mfocr. See PR642 for some more context.
715 //===---------------------------------------------------------------------===//
717 void foo(float *data, float d) {
719 for (i = 0; i < 8000; i++)
722 void foo2(float *data, float d) {
725 for (i = 0; i < 8000; i++) {
738 cmplwi cr0, r4, 32000
747 cmplwi cr0, r4, 32000
752 The 'mr' could be eliminated to folding the add into the cmp better.
754 //===---------------------------------------------------------------------===//
755 Codegen for the following (low-probability) case deteriorated considerably
756 when the correctness fixes for unordered comparisons went in (PR 642, 58871).
757 It should be possible to recover the code quality described in the comments.
759 ; RUN: llvm-as < %s | llc -march=ppc32 | grep or | count 3
760 ; This should produce one 'or' or 'cror' instruction per function.
762 ; RUN: llvm-as < %s | llc -march=ppc32 | grep mfcr | count 3
765 define i32 @test(double %x, double %y) nounwind {
767 %tmp3 = fcmp ole double %x, %y ; <i1> [#uses=1]
768 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
772 define i32 @test2(double %x, double %y) nounwind {
774 %tmp3 = fcmp one double %x, %y ; <i1> [#uses=1]
775 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
779 define i32 @test3(double %x, double %y) nounwind {
781 %tmp3 = fcmp ugt double %x, %y ; <i1> [#uses=1]
782 %tmp34 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
785 //===----------------------------------------------------------------------===//
786 ; RUN: llvm-as < %s | llc -march=ppc32 | not grep fneg
788 ; This could generate FSEL with appropriate flags (FSEL is not IEEE-safe, and
789 ; should not be generated except with -enable-finite-only-fp-math or the like).
790 ; With the correctness fixes for PR642 (58871) LowerSELECT_CC would need to
791 ; recognize a more elaborate tree than a simple SETxx.
793 define double @test_FNEG_sel(double %A, double %B, double %C) {
794 %D = sub double -0.000000e+00, %A ; <double> [#uses=1]
795 %Cond = fcmp ugt double %D, -0.000000e+00 ; <i1> [#uses=1]
796 %E = select i1 %Cond, double %B, double %C ; <double> [#uses=1]