1 //===- README.txt - Notes for improving PowerPC-specific code gen ---------===//
5 * implement do-loop -> bdnz transform
6 * Implement __builtin_trap (ISD::TRAP) as 'tw 31, 0, 0' aka 'trap'.
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 ===-------------------------------------------------------------------------===
158 int %f1(int %a, int %b) {
159 %tmp.1 = and int %a, 15 ; <int> [#uses=1]
160 %tmp.3 = and int %b, 240 ; <int> [#uses=1]
161 %tmp.4 = or int %tmp.3, %tmp.1 ; <int> [#uses=1]
165 without a copy. We make this currently:
168 rlwinm r2, r4, 0, 24, 27
169 rlwimi r2, r3, 0, 28, 31
173 The two-addr pass or RA needs to learn when it is profitable to commute an
174 instruction to avoid a copy AFTER the 2-addr instruction. The 2-addr pass
175 currently only commutes to avoid inserting a copy BEFORE the two addr instr.
177 ===-------------------------------------------------------------------------===
179 Compile offsets from allocas:
182 %X = alloca { int, int }
183 %Y = getelementptr {int,int}* %X, int 0, uint 1
187 into a single add, not two:
194 --> important for C++.
196 ===-------------------------------------------------------------------------===
198 No loads or stores of the constants should be needed:
200 struct foo { double X, Y; };
201 void xxx(struct foo F);
202 void bar() { struct foo R = { 1.0, 2.0 }; xxx(R); }
204 ===-------------------------------------------------------------------------===
206 Darwin Stub LICM optimization:
212 Have to go through an indirect stub if bar is external or linkonce. It would
213 be better to compile it as:
218 which only computes the address of bar once (instead of each time through the
219 stub). This is Darwin specific and would have to be done in the code generator.
220 Probably not a win on x86.
222 ===-------------------------------------------------------------------------===
224 Simple IPO for argument passing, change:
225 void foo(int X, double Y, int Z) -> void foo(int X, int Z, double Y)
227 the Darwin ABI specifies that any integer arguments in the first 32 bytes worth
228 of arguments get assigned to r3 through r10. That is, if you have a function
229 foo(int, double, int) you get r3, f1, r6, since the 64 bit double ate up the
230 argument bytes for r4 and r5. The trick then would be to shuffle the argument
231 order for functions we can internalize so that the maximum number of
232 integers/pointers get passed in regs before you see any of the fp arguments.
234 Instead of implementing this, it would actually probably be easier to just
235 implement a PPC fastcc, where we could do whatever we wanted to the CC,
236 including having this work sanely.
238 ===-------------------------------------------------------------------------===
240 Fix Darwin FP-In-Integer Registers ABI
242 Darwin passes doubles in structures in integer registers, which is very very
243 bad. Add something like a BIT_CONVERT to LLVM, then do an i-p transformation
244 that percolates these things out of functions.
246 Check out how horrible this is:
247 http://gcc.gnu.org/ml/gcc/2005-10/msg01036.html
249 This is an extension of "interprocedural CC unmunging" that can't be done with
252 ===-------------------------------------------------------------------------===
259 return b * 3; // ignore the fact that this is always 3.
265 into something not this:
270 rlwinm r2, r2, 29, 31, 31
272 bgt cr0, LBB1_2 ; UnifiedReturnBlock
274 rlwinm r2, r2, 0, 31, 31
277 LBB1_2: ; UnifiedReturnBlock
281 In particular, the two compares (marked 1) could be shared by reversing one.
282 This could be done in the dag combiner, by swapping a BR_CC when a SETCC of the
283 same operands (but backwards) exists. In this case, this wouldn't save us
284 anything though, because the compares still wouldn't be shared.
286 ===-------------------------------------------------------------------------===
288 We should custom expand setcc instead of pretending that we have it. That
289 would allow us to expose the access of the crbit after the mfcr, allowing
290 that access to be trivially folded into other ops. A simple example:
292 int foo(int a, int b) { return (a < b) << 4; }
299 rlwinm r2, r2, 29, 31, 31
303 ===-------------------------------------------------------------------------===
305 Fold add and sub with constant into non-extern, non-weak addresses so this:
308 void bar(int b) { a = b; }
309 void foo(unsigned char *c) {
326 lbz r2, lo16(_a+3)(r2)
330 ===-------------------------------------------------------------------------===
332 We generate really bad code for this:
334 int f(signed char *a, _Bool b, _Bool c) {
340 ===-------------------------------------------------------------------------===
343 int test(unsigned *P) { return *P >> 24; }
358 ===-------------------------------------------------------------------------===
360 On the G5, logical CR operations are more expensive in their three
361 address form: ops that read/write the same register are half as expensive as
362 those that read from two registers that are different from their destination.
364 We should model this with two separate instructions. The isel should generate
365 the "two address" form of the instructions. When the register allocator
366 detects that it needs to insert a copy due to the two-addresness of the CR
367 logical op, it will invoke PPCInstrInfo::convertToThreeAddress. At this point
368 we can convert to the "three address" instruction, to save code space.
370 This only matters when we start generating cr logical ops.
372 ===-------------------------------------------------------------------------===
374 We should compile these two functions to the same thing:
377 void f(int a, int b, int *P) {
378 *P = (a-b)>=0?(a-b):(b-a);
380 void g(int a, int b, int *P) {
384 Further, they should compile to something better than:
390 bgt cr0, LBB2_2 ; entry
407 ... which is much nicer.
409 This theoretically may help improve twolf slightly (used in dimbox.c:142?).
411 ===-------------------------------------------------------------------------===
413 int foo(int N, int ***W, int **TK, int X) {
416 for (t = 0; t < N; ++t)
417 for (i = 0; i < 4; ++i)
418 W[t / X][i][t % X] = TK[i][t];
423 We generate relatively atrocious code for this loop compared to gcc.
425 We could also strength reduce the rem and the div:
426 http://www.lcs.mit.edu/pubs/pdf/MIT-LCS-TM-600.pdf
428 ===-------------------------------------------------------------------------===
430 float foo(float X) { return (int)(X); }
445 We could use a target dag combine to turn the lwz/extsw into an lwa when the
446 lwz has a single use. Since LWA is cracked anyway, this would be a codesize
449 ===-------------------------------------------------------------------------===
451 We generate ugly code for this:
453 void func(unsigned int *ret, float dx, float dy, float dz, float dw) {
455 if(dx < -dw) code |= 1;
456 if(dx > dw) code |= 2;
457 if(dy < -dw) code |= 4;
458 if(dy > dw) code |= 8;
459 if(dz < -dw) code |= 16;
460 if(dz > dw) code |= 32;
464 ===-------------------------------------------------------------------------===
466 Complete the signed i32 to FP conversion code using 64-bit registers
467 transformation, good for PI. See PPCISelLowering.cpp, this comment:
469 // FIXME: disable this lowered code. This generates 64-bit register values,
470 // and we don't model the fact that the top part is clobbered by calls. We
471 // need to flag these together so that the value isn't live across a call.
472 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
474 Also, if the registers are spilled to the stack, we have to ensure that all
475 64-bits of them are save/restored, otherwise we will miscompile the code. It
476 sounds like we need to get the 64-bit register classes going.
478 ===-------------------------------------------------------------------------===
480 %struct.B = type { i8, [3 x i8] }
482 define void @bar(%struct.B* %b) {
484 %tmp = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
485 %tmp = load i32* %tmp ; <uint> [#uses=1]
486 %tmp3 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=1]
487 %tmp4 = load i32* %tmp3 ; <uint> [#uses=1]
488 %tmp8 = bitcast %struct.B* %b to i32* ; <uint*> [#uses=2]
489 %tmp9 = load i32* %tmp8 ; <uint> [#uses=1]
490 %tmp4.mask17 = shl i32 %tmp4, i8 1 ; <uint> [#uses=1]
491 %tmp1415 = and i32 %tmp4.mask17, 2147483648 ; <uint> [#uses=1]
492 %tmp.masked = and i32 %tmp, 2147483648 ; <uint> [#uses=1]
493 %tmp11 = or i32 %tmp1415, %tmp.masked ; <uint> [#uses=1]
494 %tmp12 = and i32 %tmp9, 2147483647 ; <uint> [#uses=1]
495 %tmp13 = or i32 %tmp12, %tmp11 ; <uint> [#uses=1]
496 store i32 %tmp13, i32* %tmp8
506 rlwimi r2, r4, 0, 0, 0
510 We could collapse a bunch of those ORs and ANDs and generate the following
515 rlwinm r4, r2, 1, 0, 0
520 ===-------------------------------------------------------------------------===
524 unsigned test6(unsigned x) {
525 return ((x & 0x00FF0000) >> 16) | ((x & 0x000000FF) << 16);
532 rlwinm r3, r3, 16, 0, 31
541 rlwinm r3,r3,16,24,31
546 ===-------------------------------------------------------------------------===
548 Consider a function like this:
550 float foo(float X) { return X + 1234.4123f; }
552 The FP constant ends up in the constant pool, so we need to get the LR register.
553 This ends up producing code like this:
562 addis r2, r2, ha16(.CPI_foo_0-"L00000$pb")
563 lfs f0, lo16(.CPI_foo_0-"L00000$pb")(r2)
569 This is functional, but there is no reason to spill the LR register all the way
570 to the stack (the two marked instrs): spilling it to a GPR is quite enough.
572 Implementing this will require some codegen improvements. Nate writes:
574 "So basically what we need to support the "no stack frame save and restore" is a
575 generalization of the LR optimization to "callee-save regs".
577 Currently, we have LR marked as a callee-save reg. The register allocator sees
578 that it's callee save, and spills it directly to the stack.
580 Ideally, something like this would happen:
582 LR would be in a separate register class from the GPRs. The class of LR would be
583 marked "unspillable". When the register allocator came across an unspillable
584 reg, it would ask "what is the best class to copy this into that I *can* spill"
585 If it gets a class back, which it will in this case (the gprs), it grabs a free
586 register of that class. If it is then later necessary to spill that reg, so be
589 ===-------------------------------------------------------------------------===
593 return X ? 524288 : 0;
601 beq cr0, LBB1_2 ;entry
614 This sort of thing occurs a lot due to globalopt.
616 ===-------------------------------------------------------------------------===
618 We currently compile 32-bit bswap:
620 declare i32 @llvm.bswap.i32(i32 %A)
621 define i32 @test(i32 %A) {
622 %B = call i32 @llvm.bswap.i32(i32 %A)
629 rlwinm r2, r3, 24, 16, 23
631 rlwimi r2, r3, 8, 24, 31
632 rlwimi r4, r3, 8, 8, 15
633 rlwimi r4, r2, 0, 16, 31
637 it would be more efficient to produce:
640 rlwinm r3,r3,8,0xffffffff
642 rlwimi r3,r0,24,16,23
645 ===-------------------------------------------------------------------------===
647 test/CodeGen/PowerPC/2007-03-24-cntlzd.ll compiles to:
649 __ZNK4llvm5APInt17countLeadingZerosEv:
652 or r2, r2, r2 <<-- silly.
656 The dead or is a 'truncate' from 64- to 32-bits.
658 ===-------------------------------------------------------------------------===
660 We generate horrible ppc code for this:
672 addi r5, r5, 1 ;; Extra IV for the exit value compare.
676 xoris r6, r5, 30 ;; This is due to a large immediate.
677 cmplwi cr0, r6, 33920
680 //===---------------------------------------------------------------------===//
684 inline std::pair<unsigned, bool> full_add(unsigned a, unsigned b)
685 { return std::make_pair(a + b, a + b < a); }
686 bool no_overflow(unsigned a, unsigned b)
687 { return !full_add(a, b).second; }
704 rlwinm r2, r2, 29, 31, 31
708 //===---------------------------------------------------------------------===//
710 We compile some FP comparisons into an mfcr with two rlwinms and an or. For
713 int test(double x, double y) { return islessequal(x, y);}
714 int test2(double x, double y) { return islessgreater(x, y);}
715 int test3(double x, double y) { return !islessequal(x, y);}
717 Compiles into (all three are similar, but the bits differ):
722 rlwinm r3, r2, 29, 31, 31
723 rlwinm r2, r2, 31, 31, 31
727 GCC compiles this into:
736 which is more efficient and can use mfocr. See PR642 for some more context.
738 //===---------------------------------------------------------------------===//