On 05/05/2016 12:01 AM, Shiva Chen wrote:
Hi,
SMS transformation would change the kernel loop iteration count.
To do this, SMS pass will find the register contain loop count
and generate the instructions to adjust loop count.
Currently, SMS will try to find count_reg by recognizing doloop_end
pattern which means if the target didn't define doloop_end pattern
or the loop not suitable for doloop optimization, SMS pass won't active.
The patch use induction variable analysis to find count_reg instead of
find doloop_end pattern.
So these patches need to be bootstrapped and regression tested. Since
SMS is not the default on any major platforms, I would recommend first
hacking SMS to be enabled by default. That isn't a patch you're going
to submit, but instead it allows you to do bootstrap and regression
testing on x86_64, ppc64 or whatever desktop/server machines you have
access to.
I did that with patch #1 just to see what would happen and I get an
assert triggered in generate_prolog_epilog:
gcc_assert (REG_P (sub_reg));
Where sub_reg and count_reg are:
(subreg:SI (reg:DI 146 [ ivtmp.11 ]) 0)
AFAICT (reg:DI 146) is the actual IV, but the test actually occurs in
SImode:
(insn 80 79 82 7 (parallel [
(set (reg:DI 146 [ ivtmp.11 ])
(plus:DI (reg:DI 146 [ ivtmp.11 ])
(const_int -1 [0xffffffffffffffff])))
(clobber (reg:CC 17 flags))
]) 212 {*adddi_1}
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil)))
(insn 82 80 83 7 (set (reg:CCGOC 17 flags)
(compare:CCGOC (subreg:SI (reg:DI 146 [ ivtmp.11 ]) 0)
(const_int 0 [0])))
../../../../gcc/libstdc++-v3/libsupc++/hash_bytes.cc:54 3 {*cmpsi_ccno_1}
(nil))
(jump_insn 83 82 84 7 (set (pc)
(if_then_else (ge (reg:CCGOC 17 flags)
(const_int 0 [0]))
(label_ref:DI 81)
(pc)))
../../../../gcc/libstdc++-v3/libsupc++/hash_bytes.cc:54 635 {*jcc_1}
(expr_list:REG_DEAD (reg:CCGOC 17 flags)
(int_list:REG_BR_PROB 8500 (nil)))
-> 81)
So you either need to be filtering out cases where the IV is in a
different mode than its test or handling SUBREGs better throughout the code.
I haven't looked deeply at any of the patches yet. I won't until
they're in better shape from a correctness standpoint.
Jeff
