* TCWG-830 (4/10)
- Observing tree dumps
- Peeling for alignment happens at -O3 but not at -O2 -ftree-vectorize
Reason: in vect_enhance_data_refs_alignment() for:
a) -O2 -ftree-vectorize: max_allowed_peel == 0
b) -O3: max_allowed_peel == (unsigned) -1;
which equals UINT_MAX and therefore peeling gets allowed.
- Workaround: Pass -param vect-max-peeling-for-alignment=0
- Peeling for alignment with O2 can be enabled by passing
-fvect-cost-model (we don't want this!)
Reason:
opts.c:
/* Tune vectorization related parametees according to cost model. */
if (opts->x_flag_vect_cost_model == VECT_COST_MODEL_CHEAP)
{
maybe_set_param_value (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS,
6, opts->x_param_values, opts_set->x_param_values);
maybe_set_param_value (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS,
0, opts->x_param_values, opts_set->x_param_values);
maybe_set_param_value (PARAM_VECT_MAX_PEELING_FOR_ALIGNMENT,
0, opts->x_param_values, opts_set->x_param_values);
}
The above if condition becomes false when -fvect-cost-model is passed.
- Proposed patch (untested): http://pastebin.com/ftp0mrwH
Patch follows the workaround and passes --param vect-max-peeling-for-alignment=0
if unaligned access is supported.
* TCWG-777 (4/10)
- Observing tree and rtl dumps
- Workaround: for -O1 pass -fno-tree-fre -fno-tree-dominator-opts
Test-case: http://pastebin.com/cjBcSpiT
Generated assembly at -O1 without workaround: http://pastebin.com/jmQGZhN9
Generated assembly at -O1 with workaround: http://pastebin.com/JGj05z66
Is that the expected output for no unnecessary temps in assembly with
workaround ?
Is it profitable over the assembly generated without workaround ?
- Approach currently taken:
a) New pass "remove-temps" (for lack of better name), after nrv (added
as last gimple pass).
b) Transforms:
if (ssa_var != 0)
to
new_ssa_var = SSA_NAME_DEF_STMT (ssa_var)
if (new_ssa_var != 0)
This "unfolds" cse on expressions within if, which was done by fre
(and if fre was disabled then by dom pass).
c) However this approach results in dead stores.
eg:
_8 = flags_7(D) & 1;
if (_8 != 0)
...
is transformed to:
_8 = flags_7(D) & 1;
_32 = flags_7(D) & 1;
if (_32 != 0)
...
so store to _8 is dead store.
I tried to run dse after remove-temps but that didn't work.
RTL 194r.jump eliminates the above dead store as "trivially dead insn".
However I don't think it's a good idea to have dead stores like these
in gimple and rely
on RTL to eliminate them. I could try to make the pass bit smarter to
not generate redundant stores like _32 != 0 in above case.
d) Patch (no intent to commit as-is): http://pastebin.com/AGXnSkrZ
Generated assembly at -O1 with the patch: http://pastebin.com/VmHCVpGC
Patch eliminates temporaries at -O1 but not at -O2.
I have not yet figured out the reason for that.
For if (flags & 1),
In dfinish pass for -O1, the generated RTL is from
zeroextractsi_compare0_scratch
while for -O2, the generated RTL is from andsi3_compare0
e) Is this a problem also on x86 ?
x86 generated assembly with -O1: http://pastebin.com/XMeTXXwK
* Misc (2/10)
- Getting familiar with vectorizer and NEON gcc intrinsics
- Reviewed git tutorials and starting preparation of git doc
- Conference calls
== Next Week ==
- Continue working on TCWG-830 and TCWG-777
- Header file flattening
- Travel to Mumbai on 2nd July (Thursday) for US Visa OFC appointment.