Richard Biener <rguent...@suse.de> writes:
> On Fri, 18 Oct 2024, Richard Sandiford wrote:
>
>> This patch extends get_nonzero_bits to handle POLY_INT_CSTs,
>> The easiest (but also most useful) case is that the number
>> of trailing zeros in the runtime value is at least the number
>> of trailing zeros in each individual component.
>>
>> In principle, we could do this for coeffs 1 and above only,
>> and then OR in ceoff 0. This would give ~0x11 for [14, 32], say.
>> But that's future work.
>>
>> gcc/
>> * tree-ssanames.cc (get_nonzero_bits): Handle POLY_INT_CSTs.
>> * match.pd (with_possible_nonzero_bits): Likewise.
>>
>> gcc/testsuite/
>> * gcc.target/aarch64/sve/cnt_fold_4.c: New test.
>> ---
>> gcc/match.pd | 2 +
>> .../gcc.target/aarch64/sve/cnt_fold_4.c | 61 +++++++++++++++++++
>> gcc/tree-ssanames.cc | 3 +
>> 3 files changed, 66 insertions(+)
>> create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/cnt_fold_4.c
>>
>> diff --git a/gcc/match.pd b/gcc/match.pd
>> index 540582dc984..41903554478 100644
>> --- a/gcc/match.pd
>> +++ b/gcc/match.pd
>> @@ -2893,6 +2893,8 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>> possibly set. */
>> (match with_possible_nonzero_bits
>> INTEGER_CST@0)
>> +(match with_possible_nonzero_bits
>> + POLY_INT_CST@0)
>> (match with_possible_nonzero_bits
>> SSA_NAME@0
>> (if (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))))
>> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/cnt_fold_4.c
>> b/gcc/testsuite/gcc.target/aarch64/sve/cnt_fold_4.c
>> new file mode 100644
>> index 00000000000..b7a53701993
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.target/aarch64/sve/cnt_fold_4.c
>> @@ -0,0 +1,61 @@
>> +/* { dg-do compile } */
>> +/* { dg-options "-O2" } */
>> +/* { dg-final { check-function-bodies "**" "" } } */
>> +
>> +#include <arm_sve.h>
>> +
>> +/*
>> +** f1:
>> +** cnth x0
>> +** ret
>> +*/
>> +uint64_t
>> +f1 ()
>> +{
>> + uint64_t x = svcntw ();
>> + x >>= 2;
>> + return x << 3;
>> +}
>> +
>> +/*
>> +** f2:
>> +** [^\n]+
>> +** [^\n]+
>> +** ...
>> +** ret
>> +*/
>> +uint64_t
>> +f2 ()
>> +{
>> + uint64_t x = svcntd ();
>> + x >>= 2;
>> + return x << 3;
>> +}
>> +
>> +/*
>> +** f3:
>> +** cntb x0, all, mul #4
>> +** ret
>> +*/
>> +uint64_t
>> +f3 ()
>> +{
>> + uint64_t x = svcntd ();
>> + x >>= 1;
>> + return x << 6;
>> +}
>> +
>> +/*
>> +** f4:
>> +** [^\n]+
>> +** [^\n]+
>> +** ...
>> +** ret
>> +*/
>> +uint64_t
>> +f4 ()
>> +{
>> + uint64_t x = svcntd ();
>> + x >>= 2;
>> + return x << 2;
>> +}
>> diff --git a/gcc/tree-ssanames.cc b/gcc/tree-ssanames.cc
>> index 4f83fcbb517..d2d1ec18797 100644
>> --- a/gcc/tree-ssanames.cc
>> +++ b/gcc/tree-ssanames.cc
>> @@ -505,6 +505,9 @@ get_nonzero_bits (const_tree name)
>> /* Use element_precision instead of TYPE_PRECISION so complex and
>> vector types get a non-zero precision. */
>> unsigned int precision = element_precision (TREE_TYPE (name));
>> + if (POLY_INT_CST_P (name))
>> + return -known_alignment (wi::to_poly_wide (name));
>> +
>
> Since you don't need precision can you move this right after the
> INTEGER_CST handling?
Oops, yes. An earlier cut did use the precision, but I forgot to
move it to a more sensible place when changing it.
Thanks for the reviews. I've pushed parts 1, 2, and 4-9 with the
changes suggested. Part 3 needs more work, so I'll do that separately.
Richard
