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 would 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 ===-------------------------------------------------------------------------===
185 We still generate calls to foo$stub, and stubs, on Darwin. This is not
186 necessary when building with the Leopard (10.5) or later linker, as stubs are
187 generated by ld when necessary. Parameterizing this based on the deployment
188 target (-mmacosx-version-min) is probably enough. x86-32 does this right, see
191 ===-------------------------------------------------------------------------===
193 Darwin Stub LICM optimization:
199 Have to go through an indirect stub if bar is external or linkonce. It would
200 be better to compile it as:
205 which only computes the address of bar once (instead of each time through the
206 stub). This is Darwin specific and would have to be done in the code generator.
207 Probably not a win on x86.
209 ===-------------------------------------------------------------------------===
211 Simple IPO for argument passing, change:
212 void foo(int X, double Y, int Z) -> void foo(int X, int Z, double Y)
214 the Darwin ABI specifies that any integer arguments in the first 32 bytes worth
215 of arguments get assigned to r3 through r10. That is, if you have a function
216 foo(int, double, int) you get r3, f1, r6, since the 64 bit double ate up the
217 argument bytes for r4 and r5. The trick then would be to shuffle the argument
218 order for functions we can internalize so that the maximum number of
219 integers/pointers get passed in regs before you see any of the fp arguments.
221 Instead of implementing this, it would actually probably be easier to just
222 implement a PPC fastcc, where we could do whatever we wanted to the CC,
223 including having this work sanely.
225 ===-------------------------------------------------------------------------===
227 Fix Darwin FP-In-Integer Registers ABI
229 Darwin passes doubles in structures in integer registers, which is very very
230 bad. Add something like a BIT_CONVERT to LLVM, then do an i-p transformation
231 that percolates these things out of functions.
233 Check out how horrible this is:
234 http://gcc.gnu.org/ml/gcc/2005-10/msg01036.html
236 This is an extension of "interprocedural CC unmunging" that can't be done with
239 ===-------------------------------------------------------------------------===
246 return b * 3; // ignore the fact that this is always 3.
252 into something not this:
257 rlwinm r2, r2, 29, 31, 31
259 bgt cr0, LBB1_2 ; UnifiedReturnBlock
261 rlwinm r2, r2, 0, 31, 31
264 LBB1_2: ; UnifiedReturnBlock
268 In particular, the two compares (marked 1) could be shared by reversing one.
269 This could be done in the dag combiner, by swapping a BR_CC when a SETCC of the
270 same operands (but backwards) exists. In this case, this wouldn't save us
271 anything though, because the compares still wouldn't be shared.
273 ===-------------------------------------------------------------------------===
275 We should custom expand setcc instead of pretending that we have it. That
276 would allow us to expose the access of the crbit after the mfcr, allowing
277 that access to be trivially folded into other ops. A simple example:
279 int foo(int a, int b) { return (a < b) << 4; }
286 rlwinm r2, r2, 29, 31, 31
290 ===-------------------------------------------------------------------------===
292 Fold add and sub with constant into non-extern, non-weak addresses so this:
295 void bar(int b) { a = b; }
296 void foo(unsigned char *c) {
313 lbz r2, lo16(_a+3)(r2)
317 ===-------------------------------------------------------------------------===
319 We generate really bad code for this:
321 int f(signed char *a, _Bool b, _Bool c) {
327 ===-------------------------------------------------------------------------===
330 int test(unsigned *P) { return *P >> 24; }
345 ===-------------------------------------------------------------------------===
347 On the G5, logical CR operations are more expensive in their three
348 address form: ops that read/write the same register are half as expensive as
349 those that read from two registers that are different from their destination.
351 We should model this with two separate instructions. The isel should generate
352 the "two address" form of the instructions. When the register allocator
353 detects that it needs to insert a copy due to the two-addresness of the CR
354 logical op, it will invoke PPCInstrInfo::convertToThreeAddress. At this point
355 we can convert to the "three address" instruction, to save code space.
357 This only matters when we start generating cr logical ops.
359 ===-------------------------------------------------------------------------===
361 We should compile these two functions to the same thing:
364 void f(int a, int b, int *P) {
365 *P = (a-b)>=0?(a-b):(b-a);
367 void g(int a, int b, int *P) {
371 Further, they should compile to something better than:
377 bgt cr0, LBB2_2 ; entry
394 ... which is much nicer.
396 This theoretically may help improve twolf slightly (used in dimbox.c:142?).
398 ===-------------------------------------------------------------------------===
400 int foo(int N, int ***W, int **TK, int X) {
403 for (t = 0; t < N; ++t)
404 for (i = 0; i < 4; ++i)
405 W[t / X][i][t % X] = TK[i][t];
410 We generate relatively atrocious code for this loop compared to gcc.
412 We could also strength reduce the rem and the div:
413 http://www.lcs.mit.edu/pubs/pdf/MIT-LCS-TM-600.pdf
415 ===-------------------------------------------------------------------------===
417 float foo(float X) { return (int)(X); }
432 We could use a target dag combine to turn the lwz/extsw into an lwa when the
433 lwz has a single use. Since LWA is cracked anyway, this would be a codesize
436 ===-------------------------------------------------------------------------===
438 We generate ugly code for this:
440 void func(unsigned int *ret, float dx, float dy, float dz, float dw) {
442 if(dx < -dw) code |= 1;
443 if(dx > dw) code |= 2;
444 if(dy < -dw) code |= 4;
445 if(dy > dw) code |= 8;
446 if(dz < -dw) code |= 16;
447 if(dz > dw) code |= 32;
451 ===-------------------------------------------------------------------------===
453 Complete the signed i32 to FP conversion code using 64-bit registers
454 transformation, good for PI. See PPCISelLowering.cpp, this comment:
456 // FIXME: disable this lowered code. This generates 64-bit register values,
457 // and we don't model the fact that the top part is clobbered by calls. We
458 // need to flag these together so that the value isn't live across a call.
459 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
461 Also, if the registers are spilled to the stack, we have to ensure that all
462 64-bits of them are save/restored, otherwise we will miscompile the code. It
463 sounds like we need to get the 64-bit register classes going.
465 ===-------------------------------------------------------------------------===
467 %struct.B = type { i8, [3 x i8] }
469 define void @bar(%struct.B* %b) {
471 %tmp = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
472 %tmp = load i32* %tmp ; <uint> [#uses=1]
473 %tmp3 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
474 %tmp4 = load i32* %tmp3 ; <uint> [#uses=1]
475 %tmp8 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=2]
476 %tmp9 = load i32* %tmp8 ; <uint> [#uses=1]
477 %tmp4.mask17 = shl i32 %tmp4, i8 1 ; <uint> [#uses=1]
478 %tmp1415 = and i32 %tmp4.mask17, 2147483648 ; <uint> [#uses=1]
479 %tmp.masked = and i32 %tmp, 2147483648 ; <uint> [#uses=1]
480 %tmp11 = or i32 %tmp1415, %tmp.masked ; <uint> [#uses=1]
481 %tmp12 = and i32 %tmp9, 2147483647 ; <uint> [#uses=1]
482 %tmp13 = or i32 %tmp12, %tmp11 ; <uint> [#uses=1]
483 store i32 %tmp13, i32* %tmp8
493 rlwimi r2, r4, 0, 0, 0
497 We could collapse a bunch of those ORs and ANDs and generate the following
502 rlwinm r4, r2, 1, 0, 0
507 ===-------------------------------------------------------------------------===
511 unsigned test6(unsigned x) {
512 return ((x & 0x00FF0000) >> 16) | ((x & 0x000000FF) << 16);
519 rlwinm r3, r3, 16, 0, 31
528 rlwinm r3,r3,16,24,31
533 ===-------------------------------------------------------------------------===
535 Consider a function like this:
537 float foo(float X) { return X + 1234.4123f; }
539 The FP constant ends up in the constant pool, so we need to get the LR register.
540 This ends up producing code like this:
549 addis r2, r2, ha16(.CPI_foo_0-"L00000$pb")
550 lfs f0, lo16(.CPI_foo_0-"L00000$pb")(r2)
556 This is functional, but there is no reason to spill the LR register all the way
557 to the stack (the two marked instrs): spilling it to a GPR is quite enough.
559 Implementing this will require some codegen improvements. Nate writes:
561 "So basically what we need to support the "no stack frame save and restore" is a
562 generalization of the LR optimization to "callee-save regs".
564 Currently, we have LR marked as a callee-save reg. The register allocator sees
565 that it's callee save, and spills it directly to the stack.
567 Ideally, something like this would happen:
569 LR would be in a separate register class from the GPRs. The class of LR would be
570 marked "unspillable". When the register allocator came across an unspillable
571 reg, it would ask "what is the best class to copy this into that I *can* spill"
572 If it gets a class back, which it will in this case (the gprs), it grabs a free
573 register of that class. If it is then later necessary to spill that reg, so be
576 ===-------------------------------------------------------------------------===
580 return X ? 524288 : 0;
588 beq cr0, LBB1_2 ;entry
601 This sort of thing occurs a lot due to globalopt.
603 ===-------------------------------------------------------------------------===
605 We currently compile 32-bit bswap:
607 declare i32 @llvm.bswap.i32(i32 %A)
608 define i32 @test(i32 %A) {
609 %B = call i32 @llvm.bswap.i32(i32 %A)
616 rlwinm r2, r3, 24, 16, 23
618 rlwimi r2, r3, 8, 24, 31
619 rlwimi r4, r3, 8, 8, 15
620 rlwimi r4, r2, 0, 16, 31
624 it would be more efficient to produce:
627 rlwinm r3,r3,8,0xffffffff
629 rlwimi r3,r0,24,16,23
632 ===-------------------------------------------------------------------------===
634 test/CodeGen/PowerPC/2007-03-24-cntlzd.ll compiles to:
636 __ZNK4llvm5APInt17countLeadingZerosEv:
639 or r2, r2, r2 <<-- silly.
643 The dead or is a 'truncate' from 64- to 32-bits.
645 ===-------------------------------------------------------------------------===
647 We generate horrible ppc code for this:
659 addi r5, r5, 1 ;; Extra IV for the exit value compare.
663 xoris r6, r5, 30 ;; This is due to a large immediate.
664 cmplwi cr0, r6, 33920
667 //===---------------------------------------------------------------------===//
671 inline std::pair<unsigned, bool> full_add(unsigned a, unsigned b)
672 { return std::make_pair(a + b, a + b < a); }
673 bool no_overflow(unsigned a, unsigned b)
674 { return !full_add(a, b).second; }
691 rlwinm r2, r2, 29, 31, 31
695 //===---------------------------------------------------------------------===//
697 We compile some FP comparisons into an mfcr with two rlwinms and an or. For
700 int test(double x, double y) { return islessequal(x, y);}
701 int test2(double x, double y) { return islessgreater(x, y);}
702 int test3(double x, double y) { return !islessequal(x, y);}
704 Compiles into (all three are similar, but the bits differ):
709 rlwinm r3, r2, 29, 31, 31
710 rlwinm r2, r2, 31, 31, 31
714 GCC compiles this into:
723 which is more efficient and can use mfocr. See PR642 for some more context.
725 //===---------------------------------------------------------------------===//
727 void foo(float *data, float d) {
729 for (i = 0; i < 8000; i++)
732 void foo2(float *data, float d) {
735 for (i = 0; i < 8000; i++) {
748 cmplwi cr0, r4, 32000
757 cmplwi cr0, r4, 32000
762 The 'mr' could be eliminated to folding the add into the cmp better.
764 //===---------------------------------------------------------------------===//
765 Codegen for the following (low-probability) case deteriorated considerably
766 when the correctness fixes for unordered comparisons went in (PR 642, 58871).
767 It should be possible to recover the code quality described in the comments.
769 ; RUN: llvm-as < %s | llc -march=ppc32 | grep or | count 3
770 ; This should produce one 'or' or 'cror' instruction per function.
772 ; RUN: llvm-as < %s | llc -march=ppc32 | grep mfcr | count 3
775 define i32 @test(double %x, double %y) nounwind {
777 %tmp3 = fcmp ole double %x, %y ; <i1> [#uses=1]
778 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
782 define i32 @test2(double %x, double %y) nounwind {
784 %tmp3 = fcmp one double %x, %y ; <i1> [#uses=1]
785 %tmp345 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
789 define i32 @test3(double %x, double %y) nounwind {
791 %tmp3 = fcmp ugt double %x, %y ; <i1> [#uses=1]
792 %tmp34 = zext i1 %tmp3 to i32 ; <i32> [#uses=1]
795 //===----------------------------------------------------------------------===//
796 ; RUN: llvm-as < %s | llc -march=ppc32 | not grep fneg
798 ; This could generate FSEL with appropriate flags (FSEL is not IEEE-safe, and
799 ; should not be generated except with -enable-finite-only-fp-math or the like).
800 ; With the correctness fixes for PR642 (58871) LowerSELECT_CC would need to
801 ; recognize a more elaborate tree than a simple SETxx.
803 define double @test_FNEG_sel(double %A, double %B, double %C) {
804 %D = sub double -0.000000e+00, %A ; <double> [#uses=1]
805 %Cond = fcmp ugt double %D, -0.000000e+00 ; <i1> [#uses=1]
806 %E = select i1 %Cond, double %B, double %C ; <double> [#uses=1]