Richard Biener <rguent...@suse.de> writes:
> On Thu, 23 May 2019, Jiufu Guo wrote:
>
>> Hi,
>>
>> Richard Biener <rguent...@suse.de> writes:
>>
>> > On Tue, 21 May 2019, Jiufu Guo wrote:
>> >
>> >> Hi,
>> >>
>> >> This patch implements a new opportunity of jump threading for PR77820.
>> >> In this optimization, conditional jumps are merged with unconditional
>> >> jump.
>> >> And then moving CMP result to GPR is eliminated.
>> >>
>> >> It looks like below:
>> >>
>> >> <P0>
>> >> p0 = a CMP b
>> >> goto <X>;
>> >>
>> >> <P1>
>> >> p1 = c CMP d
>> >> goto <X>;
>> >>
>> >> <X>
>> >> # phi = PHI <p0 (P0), p1 (P1)>
>> >> if (phi != 0) goto <Y>; else goto <Z>;
>> >>
>> >> Could be transformed to:
>> >>
>> >> <P0>
>> >> p0 = a CMP b
>> >> if (p0 != 0) goto <Y>; else goto <Z>;
>> >>
>> >> <P1>
>> >> p1 = c CMP d
>> >> if (p1 != 0) goto <Y>; else goto <Z>;
>> >>
>> >>
>> >> This optimization eliminates:
>> >> 1. saving CMP result: p0 = a CMP b.
>> >> 2. additional CMP on branch: if (phi != 0).
>> >> 3. converting CMP result if there is phi = (INT_CONV) p0 if there is.
>> >>
>> >> Bootstrapped and tested on powerpc64le with no regressions(one case is
>> >> improved)
>> >> and new testcases are added. Is this ok for trunk?
>> >>
>> >> Thanks!
>> >> Jiufu Guo
>> >>
>> ...
>> >> diff --git a/gcc/tree-ssa-threadedge.c b/gcc/tree-ssa-threadedge.c
>> >> index c3ea2d6..23000f6 100644
>> >> --- a/gcc/tree-ssa-threadedge.c
>> >> +++ b/gcc/tree-ssa-threadedge.c
>> >> @@ -1157,6 +1157,90 @@ thread_through_normal_block (edge e,
>> >> return 0;
>> >> }
>> >>
>> >> +/* Return true if PHI's INDEX-th incoming value is a CMP, and the CMP is
>> >> + defined in the incoming basic block. Otherwise return false. */
>> >> +static bool
>> >> +cmp_from_unconditional_block (gphi *phi, int index)
>> >> +{
>> >> + tree value = gimple_phi_arg_def (phi, index);
>> >> + if (!(TREE_CODE (value) == SSA_NAME && has_single_use (value)))
>> >> + return false;
>> >
>> > Not sure why we should reject a constant here but I guess we
>> > expect it to find a simplified condition anyways ;)
>> >
>> Const could be accepted here, like "# t_9 = PHI <5(3), t_17(4)>". I
>> found this case is already handled by other jump-threading code, like
>> 'ethread' pass.
>>
>> >> +
>> >> + gimple *def = SSA_NAME_DEF_STMT (value);
>> >> +
>> >> + if (!is_gimple_assign (def))
>> >> + return false;
>> >> +
>> >> + if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def)))
>> >> + {
>> >> + value = gimple_assign_rhs1 (def);
>> >> + if (!(TREE_CODE (value) == SSA_NAME && has_single_use (value)))
>> >> + return false;
>> >> +
>> >> + def = SSA_NAME_DEF_STMT (value);
>> >> +
>> >> + if (!is_gimple_assign (def))
>> >> + return false;
>> >
>> > too much vertial space.
>> >
>> Thanks, I will refine it.
>> >> + }
>> >> +
>> >> + if (TREE_CODE_CLASS (gimple_assign_rhs_code (def)) != tcc_comparison)
>> >> + return false;
>> >> +
>> >> + /* Check if phi's incoming value is defined in the incoming
>> >> basic_block. */
>> >> + edge e = gimple_phi_arg_edge (phi, index);
>> >> + if (def->bb != e->src)
>> >> + return false;
>> >
>> > why does this matter?
>> >
>> Through preparing pathes and duplicating block, this transform can also
>> help to combine a cmp in previous block and a gcond in current block.
>> "if (def->bb != e->src)" make sure the cmp is define in the incoming
>> block of the current; and then combining "cmp with gcond" is safe. If
>> the cmp is defined far from the incoming block, it would be hard to
>> achieve the combining, and the transform may not needed.
>
> We're in SSA form so the "combining" doesn't really care where the
> definition comes from.
>
>> >> +
>> >> + if (!single_succ_p (def->bb))
>> >> + return false;
>> >
>> > Or this? The actual threading will ensure this will hold true.
>> >
>> Yes, other thread code check this and ensure it to be true, like
>> function thread_through_normal_block. Since this new function is invoked
>> outside thread_through_normal_block, so, checking single_succ_p is also
>> needed for this case.
>
> I mean threading will isolate the path making this trivially true.
> It's also no requirement for combining, in fact due to the single-use
> check the definition can be sinked across the edge already (if
> the edges dest didn't have multiple predecessors which this threading
> will fix as well).
>
I would relax these check and have a test.
And I refactor the code a little as below. Thanks for any comments!
bool
edge_forwards_cmp_to_conditional_jump_through_empty_bb_p (edge e)
{
basic_block bb = e->dest;
/* See if there is only one stmt which is gcond. */
gimple *gs = last_and_only_stmt (bb);
if (gs == NULL || gimple_code (gs) != GIMPLE_COND)
return false;
/* See if gcond's condition is "(phi !=/== 0/1)". */
tree cond = gimple_cond_lhs (gs);
if (TREE_CODE (cond) != SSA_NAME
|| gimple_code (SSA_NAME_DEF_STMT (cond)) != GIMPLE_PHI
|| gimple_bb (SSA_NAME_DEF_STMT (cond)) != bb)
return false;
enum tree_code code = gimple_cond_code (gs);
tree rhs = gimple_cond_rhs (gs);
if (!(code == NE_EXPR || code == EQ_EXPR || integer_onep (rhs)
|| integer_zerop (rhs)))
return false;
gphi *phi = as_a<gphi *> (SSA_NAME_DEF_STMT (cond));
edge_iterator ei;
edge in_e;
FOR_EACH_EDGE (in_e, ei, bb->preds)
{
/* Check if phi's incoming value is CMP */
gimple *def;
tree value = PHI_ARG_DEF_FROM_EDGE (phi, in_e);
if (TREE_CODE (value) == SSA_NAME && has_single_use (value)
&& is_gimple_assign (SSA_NAME_DEF_STMT (value)))
def = SSA_NAME_DEF_STMT (value);
else
return false;
/* Or if it is (INTCONV) (a CMP b). */
if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def)))
{
value = gimple_assign_rhs1 (def);
if (TREE_CODE (value) == SSA_NAME && has_single_use (value)
&& is_gimple_assign (SSA_NAME_DEF_STMT (value)))
def = SSA_NAME_DEF_STMT (value);
else
return false;
}
if (TREE_CODE_CLASS (gimple_assign_rhs_code (def)) != tcc_comparison)
return false;
}
return true;
}
Thanks,
Jiufu Guo
>> >> + return true;
>> >> +}
>> >> +
>> >> +/* There are basic blocks look like:
>> >> + <P0>
>> >> + p0 = a CMP b ; or p0 = (INT)( a CMP b)
>> >> + goto <X>;
>> >> +
>> >> + <P1>
>> >> + p1 = c CMP d
>> >> + goto <X>;
>> >> +
>> >> + <X>
>> >> + # phi = PHI <p0 (P0), p1 (P1)>
>> >> + if (phi != 0) goto <Y>; else goto <Z>;
>> >> +
>> >> + Then, <X>: a trivial join block.
>> >> +
>> >> + Check if BB is <X> in like above. */
>> >> +
>> >> +bool
>> >> +is_trivial_join_block (basic_block bb)
>> >
>> > I'd make this work on a specific edge.
>> >
>> > edge_forwards_conditional_to_conditional_jump_through_empty_bb_p (edge e)
>> > {
>> > basic_block b = e->dest;
>> >
>> > maybe too elaborate name ;)
>> >
>> Thanks for help to name the function! It is very valuable for me ;)
>> >> +{
>> >> + gimple *gs = last_and_only_stmt (bb);
>> >> + if (gs == NULL)
>> >> + return false;
>> >> +
>> >> + if (gimple_code (gs) != GIMPLE_COND)
>> >> + return false;
>> >> +
>> >> + tree cond = gimple_cond_lhs (gs);
>> >> +
>> >> + if (TREE_CODE (cond) != SSA_NAME)
>> >> + return false;
>> >
>> > space after if( too much vertical space in this function
>> > for my taste btw.
>> Will update this.
>> >
>> > For the forwarding to work we want a NE_EXPR or EQ_EXPR
>> > as gimple_cond_code and integer_one_p or integer_zero_p
>> > gimple_cond_rhs.
>> Right, checking those would be more safe. Since no issue found, during
>> bootstrap and regression tests, so I did not add these checking. I will
>> add this checking.
>> >
>> >> +
>> >> + if (gimple_code (SSA_NAME_DEF_STMT (cond)) != GIMPLE_PHI)
>> >> + return false;
>> >> +
>> >> + gphi *phi = as_a<gphi *> (SSA_NAME_DEF_STMT (cond));
>> >
>> > I think to match your pattern you want to check that
>> > gimple_bb (phi) == bb as well here.
>> Right, it should be checked. I will update.
>> >
>> >> + for (unsigned int i = 0; i < phi->nargs; i++)
>> >> + if (!cmp_from_unconditional_block (phi, i))
>> >
>> > Just process the incoming edge argument and inline the
>> > helper. You can use PHI_ARG_DEF_FROM_EDGE here.
>> I will refine code, and try to use it.
>> >
>> > Thanks for integrating this into jump-threading - it does look
>> > like a good fit.
>> >
>> > How often does this trigger during bootstrap?
>> Thanks for your sugguestion, this could help to evaluate patch. During
>> bootstrap(stage 2 or 3), in gcc source code, 1300-1500 basic blocks are
>> fullfile this tranform.
>
> Thanks,
> Richard.
