[Bug c/83528] New: Nios2: redundant pointers to the record fields

["already5chosen at yahoo dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83528

            Bug ID: 83528
           Summary: Nios2: redundant pointers to the record fields
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: already5chosen at yahoo dot com
  Target Milestone: ---

Created attachment 42942
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=42942&action=edit
eaxmple of bad code generation for Nios2 target

In the loop over array of records Nios2 code generator accesses fields in the
record through multiple pointers instead of single pointer.
As a result of it, in my example each loop iteration updates 5 pointers instead
of 1. In real world code I observed cases with even bigger amount of redundant
pointers, sometimes so many that it causes spills of registers to the stack.

gcc version 5.3.0 (Altera 17.1 Build 590)
Yes, I know, gcc5 is no longer supported, but that's the latest version
supplied
by Altera/Intel which means that it's the latest version accessible by simple
ned user like myself.

system type: nios2-elf-gcc

gcc options: -O2.
-O1 is very similar. -O3 produces catastrophic code, but being a sane person I
don't expect that -O3 on "minor" target would be usable.

complete command line:
nios2-elf-gcc -O2 -Wall -S foo.c

the compiler output: no messages or warnings.

Pay attention that nothing like that happens with MIPS32 target, which has
exactly the same addressing modes as Nios2 and is nearly identical to Nios2 in
most aspects.
So, it seems to me that intermediate levels of gcc are not suspect. The problem
is at the back end.