Hi!
The whole excersise here is to set the padding bits (if any) of the x
elements and keep the non-padding bits as is.
B here is unsigned _BitInt(7) type.
On targets with bitint_ext_undef this works well, z = ~0 aka z = -1
sets the 0x7f bits, __builtin_clear_padding clears the remaining bit
(which is 0 anyway but pedantically undefined)
and ~*(unsigned char *) &z is 0x80.
On targets with bitint_ext_full (or bitint_ext_partial), z = ~0 still
sets the 0x7f bits, __builtin_clear_padding doesn't do anything (is
optimized away as there are no padding bits), but we still get 0x80,
which is not what I wanted, I wanted 0 on such targets.
The problem is that I chose an unsigned _BitInt(7) type for z.
One possibility would be to use signed _BitInt(7) instead, in that case
z = ~0 aka z = -1 on bitint_ext_undef targets sets 0x7f bits and probably
the upper bit too and __builtin_clear_padding then clears that bit, while
on other targets it sets all 0xff bits (because it is sign extended) and
__builtin_clear_padding does nothing.
The following patch instead uses memset to set all bits and then clear
padding.
gcc.dg/vect/vect-bitint-1.c test doesn't need fixing, because it uses
a signed _BitInt(17) type already:
_BitInt(17) z = ~0;
__builtin_clear_padding (&z);
Tested on s390x-linux (where this test failed before) and x86_64-linux
and i686-linux (where it worked and still works).
2026-09-18 Jakub Jelinek <[email protected]>
* gcc.dg/bitint-142.c (main): Don't set z to ~0, instead memset
it to ~0.
--- a/gcc/testsuite/gcc.dg/bitint-142.c 2026-09-17 12:40:27.059845682 -0400
+++ b/gcc/testsuite/gcc.dg/bitint-142.c 2026-09-18 04:32:23.932682040 -0400
@@ -26,7 +26,7 @@ main ()
B f[16] = { 8, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0 };
for (unsigned i = 0; i < 16; ++i)
x[i] = i;
- z = ~0;
+ __builtin_memset (&z, ~0, sizeof (z));
__builtin_clear_padding (&z);
for (unsigned i = 0; i < 16; ++i)
((unsigned char *) &x[0])[i] |= ~*(unsigned char *) &z;
Jakub
