On Fri, Mar 21, 2025 at 07:12:23PM +0100, Richard Biener wrote:
> So this is the following hunk where I totally misunderstood
> real_to_target when converting this from native_encode_expr and
> a tree to using REAL_VALUE_TYPE:
>
> @@ -15702,13 +15714,16 @@ initial_from_float128(cbl_field_t *field,
> _Float128 va
> lue)
> switch( field->data.capacity )
> {
> case 4:
> - *(_Float32 *)retval = (_Float32) value;
> + value = real_value_truncate (TYPE_MODE (FLOAT), value);
> + real_to_target ((long *)retval, &value, TYPE_MODE (FLOAT));
> break;
> case 8:
> - *(_Float64 *)retval = (_Float64) value;
> + value = real_value_truncate (TYPE_MODE (DOUBLE), value);
> + real_to_target ((long *)retval, &value, TYPE_MODE (DOUBLE));
> break;
> case 16:
> - *(_Float128 *)retval = (_Float128) value;
> + value = real_value_truncate (TYPE_MODE (FLOAT128), value);
> + real_to_target ((long *)retval, &value, TYPE_MODE (FLOAT128));
> break;
>
> an incremental fix is the following, exposing a native_encode_real
> with REAL_VALUE_TYPE input would be nicer, but going back to
> native_encode_expr works. So - can you apply the following and
> re-try? A testcase for the 0.01 vs. 0.0 thing would be nice to have
> as well.
I think you could avoid the jump through build_real + native_encode_expr
by moving parts of native_encode_real into a helper function and export
that helper function from fold-const.h (unless you want to duplicate it).
Basically add a helper where you pass REAL_VALUE_TYPE * + scalar_float_mode
instead of tree.
real_to_target writes up to 6 longs where each long contains 32 bits
and then one needs to store bytes from that according to target endianity.
> --- a/gcc/cobol/genapi.cc
> +++ b/gcc/cobol/genapi.cc
> @@ -15710,20 +15710,24 @@ initial_from_float128(cbl_field_t *field)
>
> case FldFloat:
> {
> + tree tem;
> retval = (char *)xmalloc(field->data.capacity);
> switch( field->data.capacity )
> {
> case 4:
> value = real_value_truncate (TYPE_MODE (FLOAT), value);
> - real_to_target ((long *)retval, &value, TYPE_MODE (FLOAT));
> + tem = build_real (FLOAT, value);
> + native_encode_expr (tem, (unsigned char *)retval, 4, 0);
> break;
Jakub
