[Bug rtl-optimization/70278] New: LRA ICE on trunk for ARM Thumb1 with Os

["andre.simoesdiasvieira at arm dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70278

            Bug ID: 70278
           Summary: LRA ICE on trunk for ARM Thumb1 with Os
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: andre.simoesdiasvieira at arm dot com
  Target Milestone: ---

Created attachment 37999
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=37999&action=edit
reduced e_hypot.c

Hello,

We are running into an ICE in lra on trunk when compiling newlib for an ARM
thumb1 target with -Os.

This happens when compiling newlib/libm/math/e_hypot.c 

See the attached reduced source and the error when executing the command below:
$arm-none-eabi-gcc -march=armv4t -mthumb -Os -S besttry.c
The error:
besttry.c: In function '__ieee754_hypot':
besttry.c:27:1: internal compiler error: in
lra_create_new_reg_with_unique_value, at lra.c:188
 }
 ^
0x94d30b lra_create_new_reg_with_unique_value(machine_mode, rtx_def*,
reg_class, char const*)
        ../src/gcc/lra.c:188
0x94d358 lra_create_new_reg(machine_mode, rtx_def*, reg_class, char const*)
        ../src/gcc/lra.c:228
0x9590e4 split_reg
        ../src/gcc/lra-constraints.c:5034
0x95a2da split_if_necessary
        ../src/gcc/lra-constraints.c:5142
0x95e08c inherit_in_ebb
        ../src/gcc/lra-constraints.c:5527
0x95e08c lra_inheritance()
        ../src/gcc/lra-constraints.c:5813
0x950900 lra(_IO_FILE*)
        ../src/gcc/lra.c:2312
0x9082d1 do_reload
        ../src/gcc/ira.c:5408
0x9082d1 execute
        ../src/gcc/ira.c:5592


My initial investigation into lra shows that this is due to split_reg calling
lra_create_new_reg with mode = VOIDmode. Others with more lra experience might
be able to spot the origin of the issue quicker though.

A bisect shows that this seems to be introduced with revision r234184 where reg
split now calls lra_create_new_reg with 'mode' that seems to be set by
'lra_reg_info[hard_regno].biggest_mode' which in our case will be SImode. This
mode is passed down to lra_create_new_reg_with_unique_value which asserts if
its not VOIDmode.

Cheers,
Andre

PS: Unrelated to this, the code in lra_create_new_reg_with_unique_value looks a
bit odd, I think mode should have been initialized with VOIDmode there.