This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Redesign jump threading profile updates


On 10/01/14 14:04, Teresa Johnson wrote:
The block frequencies are very small in this case leading to rounding
errors when computing the edge frequency. The rounding error was then
propagated into the recomputed probabilities, leading to insanities in
the outgoing edge probabilities on the jump threading path. Eventually
during rtl expansion these insanities somehow caused an ICE.

(Before this patch the probabilities weren't even being updated,
leading to insanities in other cases where they needed to be updated.)

Specifically, in this case we had a block with frequency = 1. It had
two outgoing edges each with probability 50%. Because the
EDGE_FREQUENCY computation uses a rounding divide, the outgoing edge
frequencies were both computed as 1. Later use of these block and edge
frequencies to recompute the probability lead to both outgoing edges
getting 100%.

To address this, in the routine freqs_to_counts_path can simply scale
up the frequencies when recording them in the count field. I added a
scale by REG_BR_PROB_BASE. The largest count possible would therefore
be REG_BR_PROB_BASE * BB_FREQ_MAX which is 10000 * 10000 = 100mil
which safely fits in gcov_type.

Here is the patch I am testing. Confirmed it fixes the reported issue.
Currently bootstrapping and testing on x86_64-unknown-linux-gnu. Ok
for trunk if it passes?

(The whitespace is getting messed up when I copy the patch in here -
the indentations do line up in the patch.)

Thanks,
Teresa

2014-10-01  Teresa Johnson  <tejohnson@google.com>

         * tree-ssa-threadupdate.c (freqs_to_counts_path): Scale frequencies
         up when synthesizing counts to avoid rounding errors.
OK.
jeff


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]