-
Logged In: YES
user_id=254187
The measured perf impact is indeed lower than expected, but
this does not automatically make it unimportant. Not being
able to delete a long sequence of dead code is embarrassing,
and could have perf impact elsewhere too as Chengyong
indicated.
I looked at the troubled spot in
GRA_LIVE_Complute_Local_Info() as Liu Yang mentioned. I
think the...
2002-05-30 23:26:56 UTC in Open Research Compiler - Aurora
-
Logged In: YES
user_id=254187
The measured perf impact is indeed lower than expected, but
this does not automatically make it unimportant. Being able
to delete a long sequence of dead code is embarrassing, and
could have perf impact elsewhere too as Chengyong indicated.
I looked at the troubled spot in
GRA_LIVE_Complute_Local_Info() as Liu Yang mentioned. I
think the original...
2002-05-30 23:23:44 UTC in Open Research Compiler - Aurora
-
Logged In: YES
user_id=254187
I suspect the first mov ar.lc is placed by a phase like the
counted loop (cloop) optimization. The second one is done
by the loop opt (swp/unrolling). We should first check
whether EBO is invoked, and if so, why didn't it get rid of
the dead mov ar.lc=30.
If EBO has a real difficulty in removing this dead mov (say
due to the nature of an application...
2002-04-26 22:09:55 UTC in Open Research Compiler - Aurora
-
Logged In: YES
user_id=254187
Because the cmp did not have a completer (the default is
cond), PRDB cannot assert that p6 and p7 are disjoint. This
is because with a cond completer, p6 and p7 could be both
true if they were both true before this inst and the
condition is false.
However, this is obviously that should not be the case by
looking at the source code. In fact, this could...
2002-04-26 21:59:07 UTC in Open Research Compiler - Aurora
-
Logged In: YES
user_id=254187
I am concerned to insert an instruction before the branch
to compute the complementary predicate, since that may cost
an extra inst. I believe in most cases it is sufficient to
add a new in the complementary position of the compare. For
example,
cmp.eq.unc p, p0 = a, b
(p) br ..
If we need a complementary predicate to change the layout,
put q in the...
2002-03-20 20:42:31 UTC in Open Research Compiler - Aurora
-
It helps a lot if each instruction in asm has its
associated source line number printed in the comment.
This shall be easy to do with small changes in code
emission.
addl r2=@ltoff(mtfFreq#),gp ;; //
[1] mtfFreq
ld8 r20=[r20] // [2]
<source line number here>
nop.i 0 //.
2002-01-29 23:13:23 UTC in Open Research Compiler - Aurora
-
Logged In: YES
user_id=254187
This one is certainly not as straightforward as the
elimination of save/restore 'br'. But we could check after
reg alloc is done to see whether any callee-save pr has
been modified. If not, we shall avoid the save and restore
of 'pr'.
2001-07-30 23:10:01 UTC in Open Research Compiler - Aurora
