[PATCH: PR/43129] Simplify global variable's address loading with option -fpic

[Carrot Wei <carrot at google dot com>]

Hi

The bug http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43129 describes how arm
backend generates ineffecient code to load global variable's address in terms
of both code size and performance when option -fpic is specified. It also
describes how it can be improved by relocation R_ARM_GOT_PREL.

This patch contains a new pass simplify-got to implement the suggested
optimization. It is placed between gcse and ira.

This new pass can also be applied to different targets when they satisfy

1. There should be at least 2 methods to load a global variable's address from
GOT. Usually it means using different relocation types.

2. By default all global variable's address loading use the same method: first
load the base address of GOT into a register pic_reg, then the real global
variable's address is loaded as
           (SET offset_reg <the offset from GOT base to the GOT entry>)
           (SET address_reg (MEM (PLUS pic_reg offset_reg)))

3. If the number of global address loading is less than a target specific
threshold the alternative method may be better.

4. The alternative method doesn't use the base of GOT (pic_reg).

This patch depends on the arm relocation R_ARM_GOT_PREL which is only
implemented in mainline binutils. So this patch won't be check in until the
next major release of binutils.


Test:
x86_64 bootstrap.
arm regression test on qemu.
build CSiBE for arm with options "-Os -fpic -mthumb -fno-short-enums",
the total size is:
without simplify-got    2830665
with simplify-got       2825737

ChangeLog:

2010-04-13  Wei Guozhi  <carrot@google.com>

        PR/43129
        tree-pass.h (pass_simplify_got): New variable declaration.
        timevar.def (TV_SIMPLIFY_GOT): New timevar id.
        passes.c (init_optimization_passes): Add the new pass.
        simplify-got.c : New file implement simplify-got pass.
        Makefile.in (simplify-got.o): New object.
        hooks.h (hook_rtx_void_null): New function declaration.
        hooks.c (hook_rtx_void_null): New function.
        target.h (simplify_got_access): New structure contains simplify-got
		related hooks.
        target-def.h (TARGET_GET_PIC_REG): Define hook TARGET_GET_PIC_REG.
        (TARGET_CLEAR_PIC_REG): Likewise.
        (TARGET_LOADED_GLOBAL_VAR): Likewise.
        (TARGET_MAX_USAGE_TO_AVOID_PIC_REG): Likewise.
        (TARGET_LOAD_GLOBAL_ADDRESS): Likewise.
        doc/tm.texi (TARGET_GET_PIC_REG): Document this new hook.
        (TARGET_CLEAR_PIC_REG): Likewise.
        (TARGET_LOADED_GLOBAL_VAR): Likewise.
        (TARGET_MAX_USAGE_TO_AVOID_PIC_REG): Likewise.
        (TARGET_LOAD_GLOBAL_ADDRESS): Likewise.
        config/arm/arm.md (UNSPEC_GOT_PREL_SYM): New unspec value.
        config/arm/arm.c (arm_get_pic_reg): New function.
        (arm_clear_pic_reg): New function.
        (arm_max_usage_to_avoid_pic_reg): New function.
        (arm_loaded_global_var): New function.
        (arm_load_global_address): New function.
        (arm_output_addr_const_extra): Output const expression for
        new relocation R_ARM_GOT_PREL.

2010-04-13  Wei Guozhi  <carrot@google.com>

        PR/43129
        gcc.target/arm/got1.c: New testcase.
        gcc.target/arm/got2.c: New testcase.


thanks
Wei Guozhi

Attachment: got.diff
Description: Binary data