[roland/arm] ARM: Use movw/movt more when available

  Note you'll need an unreleased binutils version (trunk, 2.25 or tip of 2.24
branch) for the new configure check to pass.

I've tested these configurations on arm-linux-gnueabihf:

* configure check fails; arm-linux-gnueabihf-gcc (4.8.2) defaults to -mthumb

  The only changes in disassembly are in the setjmp/longjmp code.  In the
  shared library versions, all that changes is the order of some literal
  pool items.  The rtld versions actually get two instructions shorter
  because it was previously using the GOT unnecessarily (though alignment
  nop-pads it out to the same text length anyway).  The static versions
  actually use movw/movt, since the linker was never broken for the
  absolute flavors of those relocs.

* configure check fails; arm-linux-gnueabihf-gcc -marm

  The only changes in disassembly are in the setjmp/longjmp code.  The
  differences are very much like the Thumb case above.

* configure check passes; arm-linux-gnueabihf-gcc (4.8.2) defaults to -mthumb
* configure check passes; arm-linux-gnueabihf-gcc -marm

  In both of these, the shared library setjmp/longjmp code changes more.
  Also, every rtld syscall stub changes because it's using movw/movt to
  materialize the address of rtld_errno and every cancellable syscall stub
  changes because it's using movw/movt to materialize the address of
  __libc_multiple_threads.

* configure check passes; arm-linux-gnueabihf-gcc (4.8.2) defaults to -mthumb;
  ARM_NO_INDEX_REGISTER temporarily hacked into arm-features.h
* configure check passes; arm-linux-gnueabihf-gcc -marm;
  ARM_NO_INDEX_REGISTER temporarily hacked into arm-features.h

I added the static copy of the basic setjmp test, since with this change
the assembly diverges more between static and shared and there is no
existing test I noticed that exercises static setjmp/longjmp.

In all those configurations, all the setjmp/ tests (still) pass.

We've talked before about the performance issues of movw/movt (two words,
two instructions with a data dependency) vs. literal pools (one load
instruction that in the best case is a one-cycle cache hit, plus a
non-instruction word that in some cases--but none here--might be shared
between multiple uses).  ARM chip experts I've talked to tell me they think
movw/movt overall is probably a wash at worst.  The two cycles and data
dependency trade off against the double cache pollution.  (The I cache sees
literal pool words that will never be executed, because they are in the
same cache line as executed code; the D cache sees instruction words that
will never be fetched as data, because they are in the same cache line as
literal pool words.)

The rtld and static versions of setjmp and longjmp already had low-hanging
fruit (unnecessary GOT use and unnecessary PIC-friendliness, respectively)
left on the the vine, suggesting maximal performance there was not a
priority.  The syscall paths might matter more, though as I said above I'm
skeptical that movw/movt is really a loser.  I'm not inclined to try to get
empirical performance measurements since this level of stuff is so likely
to vary substantially across chip variants and workloads.  Still, if you
really think preserving the status quo for Linux configurations is the
right thing to do, I can add an arm-features.h macro for disallowing
noninstructions in text (i.e. literal pools) and use the new methods only
in that case, which would be set for arm-nacl and not for arm-linux*.

Thanks,
Roland

2014-10-16  Roland McGrath  <roland@hack.frob.com>

	* sysdeps/arm/__longjmp.S [NEED_HWCAP] [IS_IN_rtld]: Use LDST_PCREL
	macro to get at the _rt_local_ro field.
	[NEED_HWCAP] [!IS_IN_rtld]: Use LDR_GLOBAL to get at _rtld_global_ro
	([PIC] case) or _dl_hwcap ([!PIC] case).
	* sysdeps/arm/setjmp.S: Likewise.

	* config.h.in (ARM_PCREL_MOVW_OK): New macro.
	* sysdeps/arm/configure.ac: New check to define it.
	* sysdeps/arm/configure: Regenerated.
	* sysdeps/arm/sysdep.h [__ASSEMBLER__]: Include <arm-features.h>.
	(LDST_INDEXED_NOINDEX, LDST_INDEXED_INDEX): New macros.
	(LDST_INDEXED, LDST_PC_INDEXED): New macros, differing definitions
	depending on [ARM_NO_INDEX_REGISTER] and [__thumb2__].
	(LDST_PCREL) [!__thumb2__ && ARCH_HAS_T2 && ARM_PCREL_MOVW_OK]:
	Use move/movt pair instead of a load.
	(LDST_GLOBAL): Macro removed.
	(LDR_GLOBAL): New macro replaces it.
	(LDR_HIDDEN): New macro.
	(PTR_MANGLE_LOAD): Use LDR_GLOBAL rather than LDST_GLOBAL.
	Use LDR_HIDDEN instead for __pointer_chk_guard_local.

	* setjmp/tst-setjmp-static.c: New file.
	* setjmp/Makefile (tests): Add it.
	(tests-static): New variable.

[roland/arm] ARM: Use movw/movt more when available

Commit Message

Comments

Patch