[linaro regression] alsa-tools FTBFS with error "unable to find a register to spill in class ‘AREG’"

Created attachment 29552
Test case

The attached test case ICEs on trunk (svn revision: 196329) as follows:

$ x86_64-unknown-linux-gnu-gcc -O2 besttry.c
besttry.c: In function ‘output_spdif’:
besttry.c:26:1: internal compiler error: in assign_by_spills, at lra-assigns.c:1268
 }
 ^
0x7f1cf8 assign_by_spills
 /work/sources/gcc-fsf/master/gcc/lra-assigns.c:1268
0x7f29b3 lra_assign()
 /work/sources/gcc-fsf/master/gcc/lra-assigns.c:1425
0x7eebe1 lra(_IO_FILE*)
 /work/sources/gcc-fsf/master/gcc/lra.c:2307
0x7b6e18 do_reload
 /work/sources/gcc-fsf/master/gcc/ira.c:4619
0x7b6e18 rest_of_handle_reload
 /work/sources/gcc-fsf/master/gcc/ira.c:4731
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <http://gcc.gnu.org/bugs.html> for instructions.

Believe this is a target issue as the original test case does not trigger the failure for arm-none-linux-gnueabihf

I can't reproduce this with r196236

Forgot to add gcc -v output:
Using built-in specs.
COLLECT_GCC=/work/builds/gcc-fsf-master/tools/bin/x86_64-unknown-linux-gnu-gcc
COLLECT_LTO_WRAPPER=/work/builds/gcc-fsf-master/tools/libexec/gcc/x86_64-unknown-linux-gnu/4.8.0/lto-wrapper
Target: x86_64-unknown-linux-gnu
Configured with: /work/sources/gcc-fsf/master/configure --target=x86_64-unknown-linux-gnu --host=x86_64-unknown-linux-gnu --build=x86_64-unknown-linux-gnu --prefix=/work/builds/gcc-fsf-master/tools --enable-languages=c,c++,fortran --with-ppl-include=/usr/include/x86_64-linux-gnu
Thread model: posix
gcc version 4.8.0 20130228 (experimental) (GCC)

(In reply to comment #1)
> I can't reproduce this with r196236

I tried with some old revisions I had on a SLES 11 SP1 machine.

I am getting this issue with r193549 at lra-assigns.c.
reload.c spill failure when I use 190150 (does not have LRA). This seems to be a long known issue in trunk.

I am triaging further

Started with http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=188526
but it was merely latent before that, so it isn't LRA bug, because it fails with reload equally.
I think the problem is in combine, where we have:
ax = call ...
flags = r59 != 0
r60 = flags >= 0 ? ax : r59
r65 = buf
r68 = 768
rep stosd [r65 .. r65 + r68 - 4] = 0
r59 = r60

and the combiner combines the r60 = flags >= 0 ? ax : r59 instruction with
r59 = r60 into r59 = flags >= 0 ? ax : r59 instruction in the last spot, thus extending the lifetime of the ax and flags hard registers across various other instructions.

-fno-tree-coalesce-vars for workarround

(In reply to comment #4)

> and the combiner combines the r60 = flags >= 0 ? ax : r59 instruction with
> r59 = r60 into r59 = flags >= 0 ? ax : r59 instruction in the last spot, thus
> extending the lifetime of the ax and flags hard registers across various other
> instructions.

But AX is member of likely spilled class, so combine should take some more care with this insn. Should we just prevent propagations of all insns that mention likely spilled hard regs?

Actually, looking more at this, I'd say combiner is innocent here, the problem is earlier , during ce1 pass, which transforms:
   16: ax:SI=call [`output_play'] argc:0
      REG_DEAD di:DI
      REG_DEAD si:SI
   17: r60:SI=ax:SI
      REG_DEAD ax:SI
   18: flags:CCGOC=cmp(r59:SI,0)
   19: pc={(flags:CCGOC>=0)?L21:pc}
      REG_DEAD flags:CCGOC
      REG_BR_PROB 0x1c84
   20: NOTE_INSN_BASIC_BLOCK 4
    6: r60:SI=r59:SI
      REG_DEAD r59:SI
   21: L21:
   22: NOTE_INSN_BASIC_BLOCK 5
into:
   16: ax:SI=call [`output_play'] argc:0
      REG_DEAD di:DI
      REG_DEAD si:SI
   17: r60:SI=ax:SI
      REG_DEAD ax:SI
   18: flags:CCGOC=cmp(r59:SI,0)
   53: flags:CCGOC=cmp(r59:SI,0)
   54: r60:SI={(flags:CCGOC>=0)?ax:SI:r59:SI}
   22: NOTE_INSN_BASIC_BLOCK 4

i.e. does what combiner tries to avoid with hard registers. I don't understand
why it can't use r60:SI instead of ax:SI.

Created attachment 29583
gcc48-pr56484.patch

Untested fix.

Author: jakub
Date: Tue Mar 5 22:25:43 2013
New Revision: 196478

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=196478
Log:
 PR rtl-optimization/56484
 * ifcvt.c (noce_process_if_block): If else_bb is NULL, avoid extending
 lifetimes of hard registers on small register class machines.

Added:
    trunk/gcc/testsuite/gcc.c-torture/compile/pr56484.c
Modified:
    trunk/gcc/ifcvt.c

Fixed.