This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [wide-int] int_traits <tree>


On 10/19/2013 10:18 AM, Richard Sandiford wrote:
Kenneth Zadeck <zadeck@naturalbridge.com> writes:
On 10/19/2013 05:01 AM, Richard Sandiford wrote:
Mike Stump <mikestump@comcast.net> writes:
+  // We optimize x < y, where y is 64 or fewer bits.
+  // We have to be careful to not allow comparison to a large positive
+  // unsigned value like 0x8000000000000000, those would be encoded
+  // with a y.len == 2.
+  if (y.precision <= HOST_BITS_PER_WIDE_INT
+      && y.len == 1)
I don't get this.  If y.precision <= HOST_BITS_PER_WIDE_INT then
y.len must be 1.  I realise that tree constants can be stored with
TREE_INT_CST_NUNITS > TYPE_PRECISION / HOST_BITS_PER_WIDE_INT
(so that extensions beyond TYPE_PRECISION are free).  But the
wide-int code is shielded from that by the ::decompose routine.
A wide int never has len > precision / HOST_BITS_PER_WIDE_INT.

Thanks,
Richard
I think that part of this is that neither mike or i really understand
how this stuff works anymore.

in the old version where we had precision 0, we would wait to
canonicalize a constant or a simple integer until we saw what the
precision of the other operand was.   That was what precison 0 meant.
it was kind of like what you are now proposing with this new trait, but
for the reason that we actually did not know what to do than some
concern about escapement.

What i do not understand is what happens what do you get when you bring
in an integer variable that is an unsigned HOST_WIDE_INT with the top
bit set.   In the precision 0 days, if the prec of the other side was 64
or less, the incoming integer took 1 hwi and if the precision was
larger, it took two hwis.  The canonicalization happened late enough so
that there was never a question.
Ah, I think I know what you mean.  The original implementation was:

   template <typename T>
   static inline const HOST_WIDE_INT*
   to_shwi1 (HOST_WIDE_INT *s, unsigned int *l, unsigned int *p, const T& x)
   {
     s[0] = x;
     if (signedp(x)
         || sizeof (T) < sizeof (HOST_WIDE_INT)
         || ! top_bit_set (x))
       {
         *l = 1;
       }
     else
       {
         s[1] = 0;
         *l = 2;
       }
     *p = 0;
     return s;
   }

   void
   wide_int_ro::check_precision (unsigned int *p1, unsigned int *p2,
                                 bool check_equal ATTRIBUTE_UNUSED,
                                 bool check_zero ATTRIBUTE_UNUSED)
   {
     gcc_checking_assert ((!check_zero) || *p1 != 0 || *p2 != 0);

     if (*p1 == 0)
       *p1 = *p2;

     if (*p2 == 0)
       *p2 = *p1;

     gcc_checking_assert ((!check_equal) || *p1 == *p2);
   }

   /* Return true if C1 < C2 using signed comparisons.  */
   template <typename T1, typename T2>
   static inline bool
   lts_p (const T1 &c1, const T2 &c2)
   {
     bool result;
     HOST_WIDE_INT ws1[WIDE_INT_MAX_ELTS];
     HOST_WIDE_INT ws2[WIDE_INT_MAX_ELTS];
     const HOST_WIDE_INT *s1, *s2;  /* Returned data */
     unsigned int cl1, cl2;         /* array lengths  */
     unsigned int p1, p2;           /* precisions */
s1 = to_shwi1 (ws1, &cl1, &p1, c1);
     s2 = to_shwi1 (ws2, &cl2, &p2, c2);
     check_precision (&p1, &p2, false, true);
if (p1 <= HOST_BITS_PER_WIDE_INT
         && p2 <= HOST_BITS_PER_WIDE_INT)
       {
         HOST_WIDE_INT x0 = sext_hwi (s1[0], p1);
         HOST_WIDE_INT x1 = sext_hwi (s2[0], p2);
         result = x0 < x1;
       }
     else
       result = lts_p_large (s1, cl1, p1, s2, cl2, p2);
#ifdef DEBUG_WIDE_INT
     debug_vaa ("wide_int_ro:: %d = (%s lts_p %s\n", result, s1, cl1, p1, s2, cl2, p2);
#endif
     return result;
   }
you need to be careful about asserting too much from the old code. the time line was:

1) we developed the stuff on x86-64
2) you did your patch
3) we ported everything to ppc and our private port.

i really only became very sensitive to this issue during step 3 because the x86 does not exhibit these bugs.


So if you had a 128-bit wide_int and T == unsigned HOST_WIDE_INT,
this lts_p would zero-extend the unsigned HOST_WIDE_INT to 128 bits and
then do a signed comparison.

The thing here is that the "check_equal" argument is false.
So if instead you were comparing a 128-bit wide_int with a 64-bit tree
constant, lts_p would treat that tree constant as a signed 64-bit number,
even if it was TYPE_UNSIGNED.  Similarly if you were comparing a 128-bit
tree constant and a 64-bit tree constant.  You also allowed a comparison
of a 128-bit wide_int with a 64-bit rtx, again treating the 64-bit rtx
as signed.
I do not think that this is what check_equal meant because the 0 precision was a wild card. The 0 precision allowed the values to come in from simple vars and constants and be converted on the fly. However, as i said above, i am not sure how well this worked since the testing for wide stuff was so thin.
So when doing the wi:: conversion, I'd interpreted the desired semantics
for lts_p as being "treat both inputs as signed without extending them",
since that's what the function did in most cases.  It seemed inconsistent
to treat a 64-bit unsigned primitive integer differently from a
64-bit unsigned tree constant.  So at the moment, it doesn't matter
i do not see this as inconsistently as you do. if i have a 6 in the gcc source, i really mean that i want to compare that 6 with a 6 of any type that happens to appear in the user's source program. My 6 has to be generic enough to match anything that the user might throw at it. This was richi's big argument against me having to write foo.lts_p (wide_int (6, foo.get_precision()). The gcc source code writer needs his 6 to be special. Richi was right!!!!
whether any HOST_WIDE_INT input to lts_p is signed or unsigned, just like
it didn't and doesn't matter whether any tree input is signed or unsigned.

If instead we want lts_p to operate to a single unified precision,
like eq_p did:

     s1 = to_shwi1 (ws1, &cl1, &p1, c1);
     s2 = to_shwi1 (ws2, &cl2, &p2, c2);
     check_precision (&p1, &p2, true, false);

and still does, then that's easy enough to change.  Then all extendable
inputs will be extended according to their natural signedness and then
compared signed.  Mixed-precision rtx comparisons would be forbidden.

But that's tangential to the point I was trying to make above,
which is that the rules about valid values for "len" and post-
check_precision "precision" are still the same as in your original
version.  So I think Mike's original patch was right and that this extra
"y.len == 1" check is redundant.  That would still be true if we changed
lts_p as above.
the relative comparisons and the equality comparisons are different. The equality comparisons allowed the precision mismatch because there were places in the front end that hashed tree-csts and so it did comparisons on things whose types were not even similar. We likely could have fixed this by changing the code around to do the type comparsion first, but we chose to make the equality case more general - hense the false for the parameter for check_equal. We never made the similar changes for the relative comparisons.

For this reason, the code really did depend on the precision 0 stuff working, because when you passed in an interger variable or constant, it would get a precision 0 and then get expanded to the proper width before the comparison happened.

My point is that i do not see how that works now because there is no tying of the precisions from the two operands of binary operations. I agree that we do not need the length test for the short circuit code, but we do need to be at a point where (unsigned HWI) 0xfffffffffffffff is canonicalized as two hwis if it is being compared with a number with a 128 precision.

Kenny


Thanks,
Richard


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]