fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-05 20:28:04 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

mesa: Don't pass paramter to __builtin_clz which would hve undefined result.

Browse source 
__builtin_clz with parameter 0 has undefined value. When
using -O3 optimizing this would result to too large next
power of two value.

Fix is to check if passed value is 1 and modify formula
for that case.
This commit is contained in:
Pauli Nieminen 2010-02-14 14:16:20 +02:00
parent 7344e20ce4
commit 6e958832af
1 changed files with 11 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

16
src/mesa/main/imports.h
View file

@ -412,13 +412,19 @@ _mesa_is_pow_two(int x)
* Source for the fallback implementation is
* Sean Eron Anderson's webpage "Bit Twiddling Hacks"
* http://graphics.stanford.edu/~seander/bithacks.html
*
* When using builtin function have to do some work
* for case when passed values 1 to prevent hiting
* undefined result from __builtin_clz. Undefined
* results would be different depending on optimization
* level used for build.
*/
static INLINE int32_t
_mesa_next_pow_two_32(uint32_t x)
{
#ifdef __GNUC__
x--;
return 1 << ((__builtin_clz(x) ^ 31) + 1);
uint32_t y = (x != 1);
return (1 + y) << ((__builtin_clz(x - y) ^ 31) );
#else
x--;
x |= x >> 1;
@ -435,11 +441,11 @@ static INLINE int64_t
_mesa_next_pow_two_64(uint64_t x)
{
#ifdef __GNUC__
x--;
uint64_t y = (x != 1);
if (sizeof(x) == sizeof(long))
return 1 << ((__builtin_clzl(x) ^ 63) + 1);
return (1 + y) << ((__builtin_clzl(x - y) ^ 63));
else
return 1 << ((__builtin_clzll(x) ^ 63) + 1);
return (1 + y) << ((__builtin_clzll(x - y) ^ 63));
#else
x--;
x |= x >> 1;