_FP_FRAC_CLZ_4 picks a word by testing X##_f[3], X##_f[2] and X##_f[1] in turn and counts the leading zeros of the first non-zero one. The X##_f[1] branch passes X##_f[2] to __FP_CLZ, a word already known to be zero, so the result is 3 * _FP_W_TYPE_SIZE - 1 whatever X##_f[1] holds. Normalizing a quad then shifts by up to a word too much whenever the leading set bit falls in word 1. sparc traps every quad operation to do_mathemu(), so hex floats parsed by strtod() and scanf() in ordinary user code come back wrong. With this, the affected float-parse tests pass on a SPARCstation 20. The wrong index has been there since 2.1.97. Cc: stable@vger.kernel.org Signed-off-by: Imre Kaloz --- Notes: - _FP_FRAC_CLZ_4 is the generic 4-word helper in include/math-emu/op-4.h, which is why LKML is on Cc. - The checkpatch error "space required after that ','" is the surrounding soft-fp idiom: every neighbouring line in op-4.h is written __FP_CLZ(R,X##_f[N]). - No Fixes: tag: this has been broken since 2.1.97, before the git history. checkpatch asks for one to go with the stable@ Cc, but there is no commit to name. include/math-emu/op-4.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/include/math-emu/op-4.h b/include/math-emu/op-4.h index ba226f8237a6..429c8f5cbf8a 100644 --- a/include/math-emu/op-4.h +++ b/include/math-emu/op-4.h @@ -170,7 +170,7 @@ } \ else if (X##_f[1]) \ { \ - __FP_CLZ(R,X##_f[2]); \ + __FP_CLZ(R,X##_f[1]); \ R += _FP_W_TYPE_SIZE*2; \ } \ else \ base-commit: 62f4c998b297cf233997a2b4cd6fc2d2df0319c9 -- 2.47.3