|
|
Richard Henderson wrote:
I expect there will be *minor* code quality regressions on many
platforms, similar to the issue with i386 string patterns. This I
will think is acceptable. Only if there are *major* code quality
issues am I willing to go back and look at this further.
I've run EEMBC using libraries generated with a gcc 3.4.3 derived compiler,
but the benchmark code compiled with mainline from Thursday with and without
the patch. Differences of the geometric means are below 0.12%, but
individual
tests in the automotive suite show more differences. aifftr01 is down
by 1.2%.
Size also shows that the text size of the aifftr01 benchmark object file
has grown
from 12788 to 12916 bytes, a 1.1% increase.
The total text size of all the bmark.o files rose from 155088 to 155248
bytes.
|
|