Chandler Carruth a6a87b595d [PM] Change the core design of the TTI analysis to use a polymorphic
type erased interface and a single analysis pass rather than an
extremely complex analysis group.

The end result is that the TTI analysis can contain a type erased
implementation that supports the polymorphic TTI interface. We can build
one from a target-specific implementation or from a dummy one in the IR.

I've also factored all of the code into "mix-in"-able base classes,
including CRTP base classes to facilitate calling back up to the most
specialized form when delegating horizontally across the surface. These
aren't as clean as I would like and I'm planning to work on cleaning
some of this up, but I wanted to start by putting into the right form.

There are a number of reasons for this change, and this particular
design. The first and foremost reason is that an analysis group is
complete overkill, and the chaining delegation strategy was so opaque,
confusing, and high overhead that TTI was suffering greatly for it.
Several of the TTI functions had failed to be implemented in all places
because of the chaining-based delegation making there be no checking of
this. A few other functions were implemented with incorrect delegation.
The message to me was very clear working on this -- the delegation and
analysis group structure was too confusing to be useful here.

The other reason of course is that this is *much* more natural fit for
the new pass manager. This will lay the ground work for a type-erased
per-function info object that can look up the correct subtarget and even
cache it.

Yet another benefit is that this will significantly simplify the
interaction of the pass managers and the TargetMachine. See the future
work below.

The downside of this change is that it is very, very verbose. I'm going
to work to improve that, but it is somewhat an implementation necessity
in C++ to do type erasure. =/ I discussed this design really extensively
with Eric and Hal prior to going down this path, and afterward showed
them the result. No one was really thrilled with it, but there doesn't
seem to be a substantially better alternative. Using a base class and
virtual method dispatch would make the code much shorter, but as
discussed in the update to the programmer's manual and elsewhere,
a polymorphic interface feels like the more principled approach even if
this is perhaps the least compelling example of it. ;]

Ultimately, there is still a lot more to be done here, but this was the
huge chunk that I couldn't really split things out of because this was
the interface change to TTI. I've tried to minimize all the other parts
of this. The follow up work should include at least:

1) Improving the TargetMachine interface by having it directly return
   a TTI object. Because we have a non-pass object with value semantics
   and an internal type erasure mechanism, we can narrow the interface
   of the TargetMachine to *just* do what we need: build and return
   a TTI object that we can then insert into the pass pipeline.
2) Make the TTI object be fully specialized for a particular function.
   This will include splitting off a minimal form of it which is
   sufficient for the inliner and the old pass manager.
3) Add a new pass manager analysis which produces TTI objects from the
   target machine for each function. This may actually be done as part
   of #2 in order to use the new analysis to implement #2.
4) Work on narrowing the API between TTI and the targets so that it is
   easier to understand and less verbose to type erase.
5) Work on narrowing the API between TTI and its clients so that it is
   easier to understand and less verbose to forward.
6) Try to improve the CRTP-based delegation. I feel like this code is
   just a bit messy and exacerbating the complexity of implementing
   the TTI in each target.

Many thanks to Eric and Hal for their help here. I ended up blocked on
this somewhat more abruptly than I expected, and so I appreciate getting
it sorted out very quickly.

Differential Revision: http://reviews.llvm.org/D7293

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@227669 91177308-0d34-0410-b5e6-96231b3b80d8
2015-01-31 03:43:40 +00:00
..
2015-01-19 21:11:14 +00:00
2015-01-10 01:57:21 +00:00

//===- README.txt - Notes for improving PowerPC-specific code gen ---------===//

TODO:
* lmw/stmw pass a la arm load store optimizer for prolog/epilog

===-------------------------------------------------------------------------===

This code:

unsigned add32carry(unsigned sum, unsigned x) {
 unsigned z = sum + x;
 if (sum + x < x)
     z++;
 return z;
}

Should compile to something like:

	addc r3,r3,r4
	addze r3,r3

instead we get:

	add r3, r4, r3
	cmplw cr7, r3, r4
	mfcr r4 ; 1
	rlwinm r4, r4, 29, 31, 31
	add r3, r3, r4

Ick.

===-------------------------------------------------------------------------===

We compile the hottest inner loop of viterbi to:

        li r6, 0
        b LBB1_84       ;bb432.i
LBB1_83:        ;bb420.i
        lbzx r8, r5, r7
        addi r6, r7, 1
        stbx r8, r4, r7
LBB1_84:        ;bb432.i
        mr r7, r6
        cmplwi cr0, r7, 143
        bne cr0, LBB1_83        ;bb420.i

The CBE manages to produce:

	li r0, 143
	mtctr r0
loop:
	lbzx r2, r2, r11
	stbx r0, r2, r9
	addi r2, r2, 1
	bdz later
	b loop

This could be much better (bdnz instead of bdz) but it still beats us.  If we
produced this with bdnz, the loop would be a single dispatch group.

===-------------------------------------------------------------------------===

Lump the constant pool for each function into ONE pic object, and reference
pieces of it as offsets from the start.  For functions like this (contrived
to have lots of constants obviously):

double X(double Y) { return (Y*1.23 + 4.512)*2.34 + 14.38; }

We generate:

_X:
        lis r2, ha16(.CPI_X_0)
        lfd f0, lo16(.CPI_X_0)(r2)
        lis r2, ha16(.CPI_X_1)
        lfd f2, lo16(.CPI_X_1)(r2)
        fmadd f0, f1, f0, f2
        lis r2, ha16(.CPI_X_2)
        lfd f1, lo16(.CPI_X_2)(r2)
        lis r2, ha16(.CPI_X_3)
        lfd f2, lo16(.CPI_X_3)(r2)
        fmadd f1, f0, f1, f2
        blr

It would be better to materialize .CPI_X into a register, then use immediates
off of the register to avoid the lis's.  This is even more important in PIC 
mode.

Note that this (and the static variable version) is discussed here for GCC:
http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html

Here's another example (the sgn function):
double testf(double a) {
       return a == 0.0 ? 0.0 : (a > 0.0 ? 1.0 : -1.0);
}

it produces a BB like this:
LBB1_1: ; cond_true
        lis r2, ha16(LCPI1_0)
        lfs f0, lo16(LCPI1_0)(r2)
        lis r2, ha16(LCPI1_1)
        lis r3, ha16(LCPI1_2)
        lfs f2, lo16(LCPI1_2)(r3)
        lfs f3, lo16(LCPI1_1)(r2)
        fsub f0, f0, f1
        fsel f1, f0, f2, f3
        blr 

===-------------------------------------------------------------------------===

PIC Code Gen IPO optimization:

Squish small scalar globals together into a single global struct, allowing the 
address of the struct to be CSE'd, avoiding PIC accesses (also reduces the size
of the GOT on targets with one).

Note that this is discussed here for GCC:
http://gcc.gnu.org/ml/gcc-patches/2006-02/msg00133.html

===-------------------------------------------------------------------------===

Darwin Stub removal:

We still generate calls to foo$stub, and stubs, on Darwin.  This is not
necessary when building with the Leopard (10.5) or later linker, as stubs are
generated by ld when necessary.  Parameterizing this based on the deployment
target (-mmacosx-version-min) is probably enough.  x86-32 does this right, see
its logic.

===-------------------------------------------------------------------------===

Darwin Stub LICM optimization:

Loops like this:
  
  for (...)  bar();

Have to go through an indirect stub if bar is external or linkonce.  It would 
be better to compile it as:

     fp = &bar;
     for (...)  fp();

which only computes the address of bar once (instead of each time through the 
stub).  This is Darwin specific and would have to be done in the code generator.
Probably not a win on x86.

===-------------------------------------------------------------------------===

Simple IPO for argument passing, change:
  void foo(int X, double Y, int Z) -> void foo(int X, int Z, double Y)

the Darwin ABI specifies that any integer arguments in the first 32 bytes worth
of arguments get assigned to r3 through r10. That is, if you have a function
foo(int, double, int) you get r3, f1, r6, since the 64 bit double ate up the
argument bytes for r4 and r5. The trick then would be to shuffle the argument
order for functions we can internalize so that the maximum number of 
integers/pointers get passed in regs before you see any of the fp arguments.

Instead of implementing this, it would actually probably be easier to just 
implement a PPC fastcc, where we could do whatever we wanted to the CC, 
including having this work sanely.

===-------------------------------------------------------------------------===

Fix Darwin FP-In-Integer Registers ABI

Darwin passes doubles in structures in integer registers, which is very very 
bad.  Add something like a BITCAST to LLVM, then do an i-p transformation that
percolates these things out of functions.

Check out how horrible this is:
http://gcc.gnu.org/ml/gcc/2005-10/msg01036.html

This is an extension of "interprocedural CC unmunging" that can't be done with
just fastcc.

===-------------------------------------------------------------------------===

Fold add and sub with constant into non-extern, non-weak addresses so this:

static int a;
void bar(int b) { a = b; }
void foo(unsigned char *c) {
  *c = a;
}

So that 

_foo:
        lis r2, ha16(_a)
        la r2, lo16(_a)(r2)
        lbz r2, 3(r2)
        stb r2, 0(r3)
        blr

Becomes

_foo:
        lis r2, ha16(_a+3)
        lbz r2, lo16(_a+3)(r2)
        stb r2, 0(r3)
        blr

===-------------------------------------------------------------------------===

We should compile these two functions to the same thing:

#include <stdlib.h>
void f(int a, int b, int *P) {
  *P = (a-b)>=0?(a-b):(b-a);
}
void g(int a, int b, int *P) {
  *P = abs(a-b);
}

Further, they should compile to something better than:

_g:
        subf r2, r4, r3
        subfic r3, r2, 0
        cmpwi cr0, r2, -1
        bgt cr0, LBB2_2 ; entry
LBB2_1: ; entry
        mr r2, r3
LBB2_2: ; entry
        stw r2, 0(r5)
        blr

GCC produces:

_g:
        subf r4,r4,r3
        srawi r2,r4,31
        xor r0,r2,r4
        subf r0,r2,r0
        stw r0,0(r5)
        blr

... which is much nicer.

This theoretically may help improve twolf slightly (used in dimbox.c:142?).

===-------------------------------------------------------------------------===

PR5945: This: 
define i32 @clamp0g(i32 %a) {
entry:
        %cmp = icmp slt i32 %a, 0
        %sel = select i1 %cmp, i32 0, i32 %a
        ret i32 %sel
}

Is compile to this with the PowerPC (32-bit) backend:

_clamp0g:
        cmpwi cr0, r3, 0
        li r2, 0
        blt cr0, LBB1_2
; BB#1:                                                     ; %entry
        mr r2, r3
LBB1_2:                                                     ; %entry
        mr r3, r2
        blr

This could be reduced to the much simpler:

_clamp0g:
        srawi r2, r3, 31
        andc r3, r3, r2
        blr

===-------------------------------------------------------------------------===

int foo(int N, int ***W, int **TK, int X) {
  int t, i;
  
  for (t = 0; t < N; ++t)
    for (i = 0; i < 4; ++i)
      W[t / X][i][t % X] = TK[i][t];
      
  return 5;
}

We generate relatively atrocious code for this loop compared to gcc.

We could also strength reduce the rem and the div:
http://www.lcs.mit.edu/pubs/pdf/MIT-LCS-TM-600.pdf

===-------------------------------------------------------------------------===

We generate ugly code for this:

void func(unsigned int *ret, float dx, float dy, float dz, float dw) {
  unsigned code = 0;
  if(dx < -dw) code |= 1;
  if(dx > dw)  code |= 2;
  if(dy < -dw) code |= 4;
  if(dy > dw)  code |= 8;
  if(dz < -dw) code |= 16;
  if(dz > dw)  code |= 32;
  *ret = code;
}

===-------------------------------------------------------------------------===

%struct.B = type { i8, [3 x i8] }

define void @bar(%struct.B* %b) {
entry:
        %tmp = bitcast %struct.B* %b to i32*              ; <uint*> [#uses=1]
        %tmp = load i32* %tmp          ; <uint> [#uses=1]
        %tmp3 = bitcast %struct.B* %b to i32*             ; <uint*> [#uses=1]
        %tmp4 = load i32* %tmp3                ; <uint> [#uses=1]
        %tmp8 = bitcast %struct.B* %b to i32*             ; <uint*> [#uses=2]
        %tmp9 = load i32* %tmp8                ; <uint> [#uses=1]
        %tmp4.mask17 = shl i32 %tmp4, i8 1          ; <uint> [#uses=1]
        %tmp1415 = and i32 %tmp4.mask17, 2147483648            ; <uint> [#uses=1]
        %tmp.masked = and i32 %tmp, 2147483648         ; <uint> [#uses=1]
        %tmp11 = or i32 %tmp1415, %tmp.masked          ; <uint> [#uses=1]
        %tmp12 = and i32 %tmp9, 2147483647             ; <uint> [#uses=1]
        %tmp13 = or i32 %tmp12, %tmp11         ; <uint> [#uses=1]
        store i32 %tmp13, i32* %tmp8
        ret void
}

We emit:

_foo:
        lwz r2, 0(r3)
        slwi r4, r2, 1
        or r4, r4, r2
        rlwimi r2, r4, 0, 0, 0
        stw r2, 0(r3)
        blr

We could collapse a bunch of those ORs and ANDs and generate the following
equivalent code:

_foo:
        lwz r2, 0(r3)
        rlwinm r4, r2, 1, 0, 0
        or r2, r2, r4
        stw r2, 0(r3)
        blr

===-------------------------------------------------------------------------===

Consider a function like this:

float foo(float X) { return X + 1234.4123f; }

The FP constant ends up in the constant pool, so we need to get the LR register.
 This ends up producing code like this:

_foo:
.LBB_foo_0:     ; entry
        mflr r11
***     stw r11, 8(r1)
        bl "L00000$pb"
"L00000$pb":
        mflr r2
        addis r2, r2, ha16(.CPI_foo_0-"L00000$pb")
        lfs f0, lo16(.CPI_foo_0-"L00000$pb")(r2)
        fadds f1, f1, f0
***     lwz r11, 8(r1)
        mtlr r11
        blr

This is functional, but there is no reason to spill the LR register all the way
to the stack (the two marked instrs): spilling it to a GPR is quite enough.

Implementing this will require some codegen improvements.  Nate writes:

"So basically what we need to support the "no stack frame save and restore" is a
generalization of the LR optimization to "callee-save regs".

Currently, we have LR marked as a callee-save reg.  The register allocator sees
that it's callee save, and spills it directly to the stack.

Ideally, something like this would happen:

LR would be in a separate register class from the GPRs. The class of LR would be
marked "unspillable".  When the register allocator came across an unspillable
reg, it would ask "what is the best class to copy this into that I *can* spill"
If it gets a class back, which it will in this case (the gprs), it grabs a free
register of that class.  If it is then later necessary to spill that reg, so be
it.

===-------------------------------------------------------------------------===

We compile this:
int test(_Bool X) {
  return X ? 524288 : 0;
}

to: 
_test:
        cmplwi cr0, r3, 0
        lis r2, 8
        li r3, 0
        beq cr0, LBB1_2 ;entry
LBB1_1: ;entry
        mr r3, r2
LBB1_2: ;entry
        blr 

instead of:
_test:
        addic r2,r3,-1
        subfe r0,r2,r3
        slwi r3,r0,19
        blr

This sort of thing occurs a lot due to globalopt.

===-------------------------------------------------------------------------===

We compile:

define i32 @bar(i32 %x) nounwind readnone ssp {
entry:
  %0 = icmp eq i32 %x, 0                          ; <i1> [#uses=1]
  %neg = sext i1 %0 to i32              ; <i32> [#uses=1]
  ret i32 %neg
}

to:

_bar:
	cntlzw r2, r3
	slwi r2, r2, 26
	srawi r3, r2, 31
	blr 

it would be better to produce:

_bar: 
        addic r3,r3,-1
        subfe r3,r3,r3
        blr

===-------------------------------------------------------------------------===

We generate horrible ppc code for this:

#define N  2000000
double   a[N],c[N];
void simpleloop() {
   int j;
   for (j=0; j<N; j++)
     c[j] = a[j];
}

LBB1_1: ;bb
        lfdx f0, r3, r4
        addi r5, r5, 1                 ;; Extra IV for the exit value compare.
        stfdx f0, r2, r4
        addi r4, r4, 8

        xoris r6, r5, 30               ;; This is due to a large immediate.
        cmplwi cr0, r6, 33920
        bne cr0, LBB1_1

//===---------------------------------------------------------------------===//

This:
        #include <algorithm>
        inline std::pair<unsigned, bool> full_add(unsigned a, unsigned b)
        { return std::make_pair(a + b, a + b < a); }
        bool no_overflow(unsigned a, unsigned b)
        { return !full_add(a, b).second; }

Should compile to:

__Z11no_overflowjj:
        add r4,r3,r4
        subfc r3,r3,r4
        li r3,0
        adde r3,r3,r3
        blr

(or better) not:

__Z11no_overflowjj:
        add r2, r4, r3
        cmplw cr7, r2, r3
        mfcr r2
        rlwinm r2, r2, 29, 31, 31
        xori r3, r2, 1
        blr 

//===---------------------------------------------------------------------===//

We compile some FP comparisons into an mfcr with two rlwinms and an or.  For
example:
#include <math.h>
int test(double x, double y) { return islessequal(x, y);}
int test2(double x, double y) {  return islessgreater(x, y);}
int test3(double x, double y) {  return !islessequal(x, y);}

Compiles into (all three are similar, but the bits differ):

_test:
	fcmpu cr7, f1, f2
	mfcr r2
	rlwinm r3, r2, 29, 31, 31
	rlwinm r2, r2, 31, 31, 31
	or r3, r2, r3
	blr 

GCC compiles this into:

 _test:
	fcmpu cr7,f1,f2
	cror 30,28,30
	mfcr r3
	rlwinm r3,r3,31,1
	blr
        
which is more efficient and can use mfocr.  See PR642 for some more context.

//===---------------------------------------------------------------------===//

void foo(float *data, float d) {
   long i;
   for (i = 0; i < 8000; i++)
      data[i] = d;
}
void foo2(float *data, float d) {
   long i;
   data--;
   for (i = 0; i < 8000; i++) {
      data[1] = d;
      data++;
   }
}

These compile to:

_foo:
	li r2, 0
LBB1_1:	; bb
	addi r4, r2, 4
	stfsx f1, r3, r2
	cmplwi cr0, r4, 32000
	mr r2, r4
	bne cr0, LBB1_1	; bb
	blr 
_foo2:
	li r2, 0
LBB2_1:	; bb
	addi r4, r2, 4
	stfsx f1, r3, r2
	cmplwi cr0, r4, 32000
	mr r2, r4
	bne cr0, LBB2_1	; bb
	blr 

The 'mr' could be eliminated to folding the add into the cmp better.

//===---------------------------------------------------------------------===//
Codegen for the following (low-probability) case deteriorated considerably 
when the correctness fixes for unordered comparisons went in (PR 642, 58871).
It should be possible to recover the code quality described in the comments.

; RUN: llvm-as < %s | llc -march=ppc32  | grep or | count 3
; This should produce one 'or' or 'cror' instruction per function.

; RUN: llvm-as < %s | llc -march=ppc32  | grep mfcr | count 3
; PR2964

define i32 @test(double %x, double %y) nounwind  {
entry:
	%tmp3 = fcmp ole double %x, %y		; <i1> [#uses=1]
	%tmp345 = zext i1 %tmp3 to i32		; <i32> [#uses=1]
	ret i32 %tmp345
}

define i32 @test2(double %x, double %y) nounwind  {
entry:
	%tmp3 = fcmp one double %x, %y		; <i1> [#uses=1]
	%tmp345 = zext i1 %tmp3 to i32		; <i32> [#uses=1]
	ret i32 %tmp345
}

define i32 @test3(double %x, double %y) nounwind  {
entry:
	%tmp3 = fcmp ugt double %x, %y		; <i1> [#uses=1]
	%tmp34 = zext i1 %tmp3 to i32		; <i32> [#uses=1]
	ret i32 %tmp34
}
//===----------------------------------------------------------------------===//
; RUN: llvm-as < %s | llc -march=ppc32 | not grep fneg

; This could generate FSEL with appropriate flags (FSEL is not IEEE-safe, and 
; should not be generated except with -enable-finite-only-fp-math or the like).
; With the correctness fixes for PR642 (58871) LowerSELECT_CC would need to
; recognize a more elaborate tree than a simple SETxx.

define double @test_FNEG_sel(double %A, double %B, double %C) {
        %D = fsub double -0.000000e+00, %A               ; <double> [#uses=1]
        %Cond = fcmp ugt double %D, -0.000000e+00               ; <i1> [#uses=1]
        %E = select i1 %Cond, double %B, double %C              ; <double> [#uses=1]
        ret double %E
}

//===----------------------------------------------------------------------===//
The save/restore sequence for CR in prolog/epilog is terrible:
- Each CR subreg is saved individually, rather than doing one save as a unit.
- On Darwin, the save is done after the decrement of SP, which means the offset
from SP of the save slot can be too big for a store instruction, which means we
need an additional register (currently hacked in 96015+96020; the solution there
is correct, but poor).
- On SVR4 the same thing can happen, and I don't think saving before the SP
decrement is safe on that target, as there is no red zone.  This is currently
broken AFAIK, although it's not a target I can exercise.
The following demonstrates the problem:
extern void bar(char *p);
void foo() {
  char x[100000];
  bar(x);
  __asm__("" ::: "cr2");
}