On Thu, Oct 06, 2022 at 06:57:40PM -0400, Michael Collison wrote:
> I am trying to improve code generation for coremark to match a recent
> improvement that was made in LLVM.
>
> I added the following transformation to match.pd which attempts to replace a
> branch with straight line code:
>
> /* (cond (and (x , 0x1) == 0), y, (z ^ y) ) -> (-(and (x , 0x1)) & z ) ^ y
> */
> (simplify
> (cond (eq (bit_and @0 integer_onep@1)
> integer_zerop)
> @2
> (bit_xor:c @3 @2))
> (bit_xor (bit_and (negate (bit_and @0 @1)) @3) @2))
>
> I get a internal error, but in stepping through the debugger I can see the
> pattern matches, but fails when when it tries to further simplify and match
> another pattern in match.pd:
>
> /* x & C -> x if we know that x & ~C == 0. */
> #if GIMPLE
> (simplify
> (bit_and SSA_NAME@0 INTEGER_CST@1)
> (if (INTEGRAL_TYPE_P (TREE_TYPE (@0))
> && wi::bit_and_not (get_nonzero_bits (@0), wi::to_wide (@1)) == 0)
> @0))
> #endif
>
> The crash occurs in wi::bit_and_not. Before digging further I want to ask if
> there is a problem with the way I wrote the transformation?
Yes. The way you wrote it, @0 and @1 (and the zero) will have the same type
(or compatible) and @2 and @3 too, but the replacement expression relies
on all of @0, @1, @2 and @3 to have compatible types.
If you have
int x;
long long y, z;
..
(x & 1) == 0 ? y : z ^ y
then
(-(x & 1) & z) ^ y
is invalid in GIMPLE.
It can be even more incompatible, e.g. y and z could be integral vectors
while x could be scalar integer, etc.
If both TREE_TYPE (@0) and type (aka TREE_TYPE (@2)/TREE_TYPE (@3)) are
scalar, then perhaps you could just add (convert? ...) around bit_and,
otherwise I think you'd better require TREE_TYPE (@0) and type to be
compatible types.
On the other side, you probably should handle also
(x & 1) != 0 ? z ^ y : y
Jakub
