RFC [1/2] divmod transform

On 27 May 2016 at 17:31, Richard Biener <rguenther@suse.de> wrote:
> On Fri, 27 May 2016, Prathamesh Kulkarni wrote:
>
>> On 27 May 2016 at 15:45, Richard Biener <rguenther@suse.de> wrote:
>> > On Wed, 25 May 2016, Prathamesh Kulkarni wrote:
>> >
>> >> On 25 May 2016 at 12:52, Richard Biener <rguenther@suse.de> wrote:
>> >> > On Tue, 24 May 2016, Prathamesh Kulkarni wrote:
>> >> >
>> >> >> On 24 May 2016 at 19:39, Richard Biener <rguenther@suse.de> wrote:
>> >> >> > On Tue, 24 May 2016, Prathamesh Kulkarni wrote:
>> >> >> >
>> >> >> >> On 24 May 2016 at 17:42, Richard Biener <rguenther@suse.de> wrote:
>> >> >> >> > On Tue, 24 May 2016, Prathamesh Kulkarni wrote:
>> >> >> >> >
>> >> >> >> >> On 23 May 2016 at 17:35, Richard Biener <richard.guenther@gmail.com> wrote:
>> >> >> >> >> > On Mon, May 23, 2016 at 10:58 AM, Prathamesh Kulkarni
>> >> >> >> >> > <prathamesh.kulkarni@linaro.org> wrote:
>> >> >> >> >> >> Hi,
>> >> >> >> >> >> I have updated my patch for divmod (attached), which was originally
>> >> >> >> >> >> based on Kugan's patch.
>> >> >> >> >> >> The patch transforms stmts with code TRUNC_DIV_EXPR and TRUNC_MOD_EXPR
>> >> >> >> >> >> having same operands to divmod representation, so we can cse computation of mod.
>> >> >> >> >> >>
>> >> >> >> >> >> t1 = a TRUNC_DIV_EXPR b;
>> >> >> >> >> >> t2 = a TRUNC_MOD_EXPR b
>> >> >> >> >> >> is transformed to:
>> >> >> >> >> >> complex_tmp = DIVMOD (a, b);
>> >> >> >> >> >> t1 = REALPART_EXPR (complex_tmp);
>> >> >> >> >> >> t2 = IMAGPART_EXPR (complex_tmp);
>> >> >> >> >> >>
>> >> >> >> >> >> * New hook divmod_expand_libfunc
>> >> >> >> >> >> The rationale for introducing the hook is that different targets have
>> >> >> >> >> >> incompatible calling conventions for divmod libfunc.
>> >> >> >> >> >> Currently three ports define divmod libfunc: c6x, spu and arm.
>> >> >> >> >> >> c6x and spu follow the convention of libgcc2.c:__udivmoddi4:
>> >> >> >> >> >> return quotient and store remainder in argument passed as pointer,
>> >> >> >> >> >> while the arm version takes two arguments and returns both
>> >> >> >> >> >> quotient and remainder having mode double the size of the operand mode.
>> >> >> >> >> >> The port should hence override the hook expand_divmod_libfunc
>> >> >> >> >> >> to generate call to target-specific divmod.
>> >> >> >> >> >> Ports should define this hook if:
>> >> >> >> >> >> a) The port does not have divmod or div insn for the given mode.
>> >> >> >> >> >> b) The port defines divmod libfunc for the given mode.
>> >> >> >> >> >> The default hook default_expand_divmod_libfunc() generates call
>> >> >> >> >> >> to libgcc2.c:__udivmoddi4 provided the operands are unsigned and
>> >> >> >> >> >> are of DImode.
>> >> >> >> >> >>
>> >> >> >> >> >> Patch passes bootstrap+test on x86_64-unknown-linux-gnu and
>> >> >> >> >> >> cross-tested on arm*-*-*.
>> >> >> >> >> >> Bootstrap+test in progress on arm-linux-gnueabihf.
>> >> >> >> >> >> Does this patch look OK ?
>> >> >> >> >> >
>> >> >> >> >> > diff --git a/gcc/targhooks.c b/gcc/targhooks.c
>> >> >> >> >> > index 6b4601b..e4a021a 100644
>> >> >> >> >> > --- a/gcc/targhooks.c
>> >> >> >> >> > +++ b/gcc/targhooks.c
>> >> >> >> >> > @@ -1965,4 +1965,31 @@ default_optab_supported_p (int, machine_mode,
>> >> >> >> >> > machine_mode, optimization_type)
>> >> >> >> >> >    return true;
>> >> >> >> >> >  }
>> >> >> >> >> >
>> >> >> >> >> > +void
>> >> >> >> >> > +default_expand_divmod_libfunc (bool unsignedp, machine_mode mode,
>> >> >> >> >> > +                              rtx op0, rtx op1,
>> >> >> >> >> > +                              rtx *quot_p, rtx *rem_p)
>> >> >> >> >> >
>> >> >> >> >> > functions need a comment.
>> >> >> >> >> >
>> >> >> >> >> > ISTR it was suggested that ARM change to libgcc2.c__udivmoddi4 style?  In that
>> >> >> >> >> > case we could avoid the target hook.
>> >> >> >> >> Well I would prefer adding the hook because that's more easier -;)
>> >> >> >> >> Would it be ok for now to go with the hook ?
>> >> >> >> >> >
>> >> >> >> >> > +      /* If target overrides expand_divmod_libfunc hook
>> >> >> >> >> > +        then perform divmod by generating call to the target-specifc divmod
>> >> >> >> >> > libfunc.  */
>> >> >> >> >> > +      if (targetm.expand_divmod_libfunc != default_expand_divmod_libfunc)
>> >> >> >> >> > +       return true;
>> >> >> >> >> > +
>> >> >> >> >> > +      /* Fall back to using libgcc2.c:__udivmoddi4.  */
>> >> >> >> >> > +      return (mode == DImode && unsignedp);
>> >> >> >> >> >
>> >> >> >> >> > I don't understand this - we know optab_libfunc returns non-NULL for 'mode'
>> >> >> >> >> > but still restrict this to DImode && unsigned?  Also if
>> >> >> >> >> > targetm.expand_divmod_libfunc
>> >> >> >> >> > is not the default we expect the target to handle all modes?
>> >> >> >> >> Ah indeed, the check for DImode is unnecessary.
>> >> >> >> >> However I suppose the check for unsignedp should be there,
>> >> >> >> >> since we want to generate call to __udivmoddi4 only if operand is unsigned ?
>> >> >> >> >
>> >> >> >> > The optab libfunc for sdivmod should be NULL in that case.
>> >> >> >> Ah indeed, thanks.
>> >> >> >> >
>> >> >> >> >> >
>> >> >> >> >> > That said - I expected the above piece to be simply a 'return true;' ;)
>> >> >> >> >> >
>> >> >> >> >> > Usually we use some can_expand_XXX helper in optabs.c to query if the target
>> >> >> >> >> > supports a specific operation (for example SImode divmod would use DImode
>> >> >> >> >> > divmod by means of widening operands - for the unsigned case of course).
>> >> >> >> >> Thanks for pointing out. So if a target does not support divmod
>> >> >> >> >> libfunc for a mode
>> >> >> >> >> but for a wider mode, then we could zero-extend operands to the wider-mode,
>> >> >> >> >> perform divmod on the wider-mode, and then cast result back to the
>> >> >> >> >> original mode.
>> >> >> >> >> I haven't done that in this patch, would it be OK to do that as a follow up ?
>> >> >> >> >
>> >> >> >> > I think that you should conservatively handle the div_optab query, thus if
>> >> >> >> > the target has a HW division in a wider mode don't use the divmod IFN.
>> >> >> >> > You'd simply iterate over GET_MODE_WIDER_MODE and repeat the
>> >> >> >> > if (optab_handler (div_optab, mode) != CODE_FOR_nothing) check, bailing
>> >> >> >> > out if that is available.
>> >> >> >> Done.
>> >> >> >> >
>> >> >> >> >> > +  /* Disable the transform if either is a constant, since
>> >> >> >> >> > division-by-constant
>> >> >> >> >> > +     may have specialized expansion.  */
>> >> >> >> >> > +  if (TREE_CONSTANT (op1) || TREE_CONSTANT (op2))
>> >> >> >> >> > +    return false;
>> >> >> >> >> >
>> >> >> >> >> > please use CONSTANT_CLASS_P (op1) || CONSTANT_CLASS_P (op2)
>> >> >> >> >> >
>> >> >> >> >> > +  if (TYPE_OVERFLOW_TRAPS (type))
>> >> >> >> >> > +    return false;
>> >> >> >> >> >
>> >> >> >> >> > why's that?  Generally please first test cheap things (trapping, constant-ness)
>> >> >> >> >> > before checking expensive stuff (target_supports_divmod_p).
>> >> >> >> >> I added TYPE_OVERFLOW_TRAPS (type) based on your suggestion in:
>> >> >> >> >> https://www.mail-archive.com/gcc@gcc.gnu.org/msg78534.html
>> >> >> >> >> "When looking at TRUNC_DIV_EXPR you should also exclude
>> >> >> >> >> the case where TYPE_OVERFLOW_TRAPS (type) as that should
>> >> >> >> >> expand using the [su]divv optabs (no trapping overflow
>> >> >> >> >> divmod optab exists)."
>> >> >> >> >
>> >> >> >> > Ok, didn't remember that.
>> >> >> >> >
>> >> >> >> >> >
>> >> >> >> >> > +static bool
>> >> >> >> >> > +convert_to_divmod (gassign *stmt)
>> >> >> >> >> > +{
>> >> >> >> >> > +  if (!divmod_candidate_p (stmt))
>> >> >> >> >> > +    return false;
>> >> >> >> >> > +
>> >> >> >> >> > +  tree op1 = gimple_assign_rhs1 (stmt);
>> >> >> >> >> > +  tree op2 = gimple_assign_rhs2 (stmt);
>> >> >> >> >> > +
>> >> >> >> >> > +  vec<gimple *> stmts = vNULL;
>> >> >> >> >> >
>> >> >> >> >> > use an auto_vec <gimple *> - you currently leak it in at least one place.
>> >> >> >> >> >
>> >> >> >> >> > +      if (maybe_clean_or_replace_eh_stmt (use_stmt, use_stmt))
>> >> >> >> >> > +       cfg_changed = true;
>> >> >> >> >> >
>> >> >> >> >> > note that this suggests you should check whether any of the stmts may throw
>> >> >> >> >> > internally as you don't update / transfer EH info correctly.  So for 'stmt' and
>> >> >> >> >> > all 'use_stmt' check stmt_can_throw_internal and if so do not add it to
>> >> >> >> >> > the list of stmts to modify.
>> >> >> >> >> >
>> >> >> >> >> > Btw, I think you should not add 'stmt' immediately but when iterating over
>> >> >> >> >> > all uses also gather uses in TRUNC_MOD_EXPR.
>> >> >> >> >> >
>> >> >> >> >> > Otherwise looks ok.
>> >> >> >> >> Done changes in this version. I am gathering mod uses same time as div uses,
>> >> >> >> >> so this imposes a constraint that mod dominates mod. I am not sure if
>> >> >> >> >> that's desirable.
>> >> >> >> >
>> >> >> >> > I think you also need a mod_seen variable now that you don't necessarily
>> >> >> >> > end up with 'stmt' in the vector of stmts.  I don't see how there is a
>> >> >> >> > constraint that mod dominates mod - it's just that the top_stmt needs
>> >> >> >> > to dominate all other uses that can be replaced with replacing top_stmt
>> >> >> >> > with a divmod.  It's just that the actual stmt set we choose may now
>> >> >> >> > depend on immediate uses order which on a second thought is bad
>> >> >> >> > as immediate uses order could be affected by debug stmts ... hmm.
>> >> >> >> >
>> >> >> >> > To avoid this please re-add the code adding 'stmt' to stmts immediately
>> >> >> >> > and add a use_stmt != stmt check to the immediate use processing loop
>> >> >> >> > so that we don't end up adding it twice.
>> >> >> >> Well I wonder what will happen for the following case:
>> >> >> >> t1 = x / y;
>> >> >> >> if (cond)
>> >> >> >>   t2 = x % y;
>> >> >> >> else
>> >> >> >>   t3 = x % y;
>> >> >> >>
>> >> >> >> Assuming stmt == top_stmt is "t2 = x % y" and use_stmt is "t3 = x % y",
>> >> >> >> use_stmt will not get added to stmts vector, since top_stmt and
>> >> >> >> use_stmt are not in same bb,
>> >> >> >> and bb's containing top_stmt and use_stmt don't dominate each other.
>> >> >> >> Not sure if this is practical case (I assume fre will hoist mod
>> >> >> >> outside if-else?)
>> >> >> >>
>> >> >> >> Now that we immediately add stmt to stmts vector, I suppose mod_seen
>> >> >> >> shall not be required ?
>> >> >> >
>> >> >> > In that case mod_seen is not needed.  But the situation you say will
>> >> >> > still happen so I wonder if we'd need a better way of iterating over
>> >> >> > immediate uses, like first pushing all candidates into a worklist
>> >> >> > vector and then iterating over that until we find no more candidates.
>> >> >> >
>> >> >> > You can then also handle the case of more than one group of stmts
>> >> >> > (the pass currently doesn't iterate in any particular useful order
>> >> >> > over BBs).
>> >> >> IIUC, we want to perform the transform if:
>> >> >> i) there exists top_stmt with code trunc_div_expr/trunc_mod_expr and
>> >> >> having same operands as stmt.
>> >> >> ii) top_stmt dominates all other stmts with code
>> >> >> trunc_div_expr/trunc_mod_expr and having same operands as top_stmt.
>> >> >>
>> >> >> Firstly, we try to get to top_stmt if it exists, by iterating over uses of stmt,
>> >> >> and then iterate over all uses of top_stmt and add them to stmts vector
>> >> >> only if top_stmt dominates all the stmts with same operands as top_stmt
>> >> >> and have code trunc_div_expr/trunc_mod_expr.
>> >> >>
>> >> >> /* Get to top_stmt.  */
>> >> >> top_stmt = stmt;
>> >> >> top_bb = gimple_bb (stmt);
>> >> >>
>> >> >> FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, op1)
>> >> >> {
>> >> >>   if (use_stmt code is TRUNC_DIV_EXPR or TRUNC_MOD_EXPR
>> >> >>       && use_stmt has same operands as stmt)
>> >> >>     {
>> >> >>       if (gimple_bb (use_stmt) dominates top_bb)
>> >> >>         {
>> >> >>           top_bb = gimple_bb (use_stmt);
>> >> >>           top_stmt = use_stmt;
>> >> >>         }
>> >> >>       else if (gimple_bb (use_stmt) == top_stmt
>> >> >>                && gimple_uid (use_stmt) < top_stmt)
>> >> >>         top_stmt = use_stmt;
>> >> >>     }
>> >> >> }
>> >> >>
>> >> >> /* Speculatively consider top_stmt as dominating all other
>> >> >> div_expr/mod_expr stmts with same operands as stmt.  */
>> >> >> stmts.safe_push (top_stmt);
>> >> >>
>> >> >> /* Walk uses of top_stmt to ensure that all stmts are dominated by top_stmt.  */
>> >> >> top_op1 = gimple_assign_rhs1 (top_stmt);
>> >> >> FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, top_op1)
>> >> >> {
>> >> >>   if (use_stmt code is TRUNC_DIV_EXPR or TRUNC_MOD_EXPR
>> >> >>       && use_stmt has same operands as top_stmt)
>> >> >>     {
>> >> >>       if (use_stmt == top_stmt)
>> >> >>         continue;
>> >> >>
>> >> >>       /* No top_stmt exits, do not proceed with transform  */
>> >> >>       if (top_bb does not dominate gimple_bb (use_stmt))
>> >> >>         return false;
>> >> >>
>> >> >>       stmts.safe_push (use_stmt);
>> >> >>     }
>> >> >> }
>> >> >>
>> >> >> For the case:
>> >> >> t1 = x / y;
>> >> >> if (cond)
>> >> >>   t2 = x % y;
>> >> >> else
>> >> >>   t3 = x % y;
>> >> >>
>> >> >> Assuming stmt is "t2 = x % y", it will walk uses of stmt and set
>> >> >> top_stmt to "t1 = x / y"
>> >> >> Then it will walk all uses of top_stmt:
>> >> >> "t2 = x % y" -> dominated by top_stmt
>> >> >> "t3 = x % y" -> dominated by top_stmt
>> >> >> Since all stmts are dominated by top_stmt, we add all three stmts to
>> >> >> vector of stmts and proceed with transform.
>> >> >>
>> >> >> For the case where, top_stmt dominates original stmt but not all stmts:
>> >> >>
>> >> >> if (cond)
>> >> >>   t1 = x / y;
>> >> >> else
>> >> >> {
>> >> >>   t2 = x % y;
>> >> >>   return;
>> >> >> }
>> >> >>
>> >> >> t3 = x % y;
>> >> >>
>> >> >> Assuming stmt is "t3 = x % y",
>> >> >> Walking stmt uses will set top_stmt to "t1 = x / y";
>> >> >>
>> >> >> Walking immediate uses of top_stmt, we find that "t2 = x % y" is not
>> >> >> dominated by top_stmt,
>> >> >> and hence don't do the transform.
>> >> >>
>> >> >> Does this sound reasonable ?
>> >> >
>> >> > Yes, that's reasonable.
>> >> Thanks, I have attached patch that implements above approach.
>> >> Does it look OK ?
>> >
>> > Please start the top-stmt search with
>> >
>> >   top_stmt = stmt;
>> >   top_bb = gimple_bb (stmt);
>> >
>> > this makes sure to process all stmts via the IL walk in case
>> > the uses have multiple independent "dominated" trees.  This also
>> > simplifies the loop body (no need to check for NULL).  This also
>> > makes mod_seen always true and you can compute div_seen in that
>> > loop as well.
>> Done. Um I don't understand why setting top_stmt to NULL won't process
>> all stmts ? AFAIU it will have one extra iteration compared to
>> initializing top_stmt to stmt (since first iteration would initialize
>> top_stmt to stmt assuming stmt does not throw) ?
>
> If you have
>
>   if (cond)
>     {
>       r = x % y;
>       q = x / y;
>     }
>  else
>     {
>       r = x % y;
>       q = x / y;
>    }
>
> then the loop over the function might end up transforming the else
> block when visiting the then block modulo and thus it will never
> transform the then block.  Because you walk immediate uses which
> do not guarantee that you end up with a top_stmt related to the
> IL point you were coming from - the first iteration does _not_
> necessarily have use_stmt == stmt.
Thanks for the explanation,
I overlooked the fact that for first iteration use_stmt may not equal stmt -;)
>
>> >
>> > Otherwise looks ok now.
>> >
>> >> The patch does not still handle the following case:
>> >> int f(int x, int y)
>> >> {
>> >>   extern int cond;
>> >>   int q, r;
>> >>
>> >>   if (cond)
>> >>     q = x % y;
>> >>   else
>> >>     q = x % y;
>> >>
>> >>   r = x % y;
>> >>   return q + r;
>> >> }
>> >>
>> >> In above case although the mod stmt is not dominated by either div
>> >> stmt, I suppose the transform
>> >> is still possible by inserting DIVMOD (x, y) before if-else ?
>> >
>> > Yeah, same for sincos where doing this requires some LCM algorithm.
>> Well I don't have a good approach for this.
>> I was thinking, before doing the divmod transform, we could walk
>> GIMPLE_COND of "diamond" shape
>> (having both arms), and check "then" bb and "else" bb have same div or
>> mod stmts and in that case put an artificial
>> same stmt above GIMPLE_COND.
>>
>> So the above case would be transformed to:
>>
>> int tmp = x / y;  // artificial top_stmt
>> if (cond)
>>   q = x / y;
>> else
>>   q = x / y;
>>
>> r = x % y;
>> return q + r;
>>
>> and then the divmod transform will see "tmp = x / y" as the topmost stmt.
>> Since top_stmt is artificially introduced, we will replace that with DIVMOD ifn
>> rather than inserting DIVMOD ifn above top_stmt as in other cases.
>
> Yeah, but it is really a general missed optimization that should be not
> required for this transform.
The attached patch ICE's during bootstrap for x86_64, and is reproducible with
following case with -m32 -O2:

typedef long long type;

type f(type x, type y)
{
  type q = x / y;
  type r = x % y;
  return q + r;
}

The ICE happens because the test-case hits
gcc_assert (unsignedp);
in default_expand_divmod_libfunc ().

Surprisingly, optab_libfunc (sdivmod_optab, DImode) returns optab_libfunc
with name "__divmoddi4" although __divmoddi4() is nowhere defined in
libgcc for x86.
(I verified that by forcing the patch to generate call to __divmoddi4,
which results in undefined reference to __divmoddi4).

This happens because in optabs.def we have:
OPTAB_NL(sdivmod_optab, "divmod$a4", UNKNOWN, "divmod", '4', gen_int_libfunc)

and gen_int_libfunc generates "__divmoddi4" on first call to optab_libfunc
and sets optab_libfunc (sdivmod_optab, DImode) to "__divmoddi4".
I wonder if we should remove gen_int_libfunc entry in optabs.def for
sdivmod_optab ?

Thanks,
Prathamesh
>
> Richard.
>
>> >
>> >> For the following test-case, I am surprised why CSE didn't take place before
>> >> widening_mul pass ?
>> >>
>> >> int
>> >> f_1 (int x, int y)
>> >> {
>> >>   int q = x / y;
>> >>   int r1 = 0, r2 = 0;
>> >>   if (cond)
>> >>     r1 = x % y;
>> >>   else
>> >>     r2 = x % y;
>> >>   return q + r1 + r2;
>> >> }
>> >
>> > This is not CSE but code hoisting which is not implemented on GIMPLE
>> > (see PR23286)
>> Ah right, thanks for pointing out the PR.
>>
>> Thanks,
>> Prathamesh
>> >
>> >> The input to widening_mul pass is:
>> >> f_1 (int x, int y)
>> >> {
>> >>   int r2;
>> >>   int r1;
>> >>   int q;
>> >>   int cond.0_1;
>> >>   int _2;
>> >>   int _11;
>> >>
>> >>   <bb 2>:
>> >>   q_7 = x_5(D) / y_6(D);
>> >>   cond.0_1 = cond;
>> >>   if (cond.0_1 != 0)
>> >>     goto <bb 3>;
>> >>   else
>> >>     goto <bb 4>;
>> >>
>> >>   <bb 3>:
>> >>   r1_9 = x_5(D) % y_6(D);
>> >>   goto <bb 5>;
>> >>
>> >>   <bb 4>:
>> >>   r2_10 = x_5(D) % y_6(D);
>> >>
>> >>   <bb 5>:
>> >>   # r1_3 = PHI <r1_9(3), 0(4)>
>> >>   # r2_4 = PHI <0(3), r2_10(4)>
>> >>   _2 = r1_3 + q_7;
>> >>   _11 = _2 + r2_4;
>> >>   return _11;
>> >>
>> >> }
>> >>
>> >> Thanks,
>> >> Prathamesh
>> >> >
>> >> > Richard.
>> >> >
>> >> >> Thanks,
>> >> >> Prathamesh
>> >> >> >
>> >> >> > Richard.
>> >> >> >
>> >> >>
>> >> >>
>> >> >
>> >> > --
>> >> > Richard Biener <rguenther@suse.de>
>> >> > SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)
>> >>
>> >
>> > --
>> > Richard Biener <rguenther@suse.de>
>> > SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)
>>
>
> --
> Richard Biener <rguenther@suse.de>
> SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)

RFC [1/2] divmod transform

Commit Message

Comments

Patch