Re: [Qemu-devel] [PATCH v3 00/26] tcg: rework liveness analysis and register allocator

[Richard Henderson]

On 2012-10-20 07:38, Aurelien Jarno wrote:
> This patch series rework the liveness analysis and register allocator
> in order to generate more optimized code, by avoiding a lot of move
> instructions. I have measured a 9% performance improvement in user mode
> and 4% in system mode.
> 
> The idea behind this patch series is to free registers as soon as the
> temps are not used anymore instead of waiting for a basic block end or
> an op with side effects. In addition temps are copied to memory as soon
> as they are not going to be written anymore, this way even globals can
> be marked as "dead", avoiding moves to a new register when inputs and
> outputs are aliased. Finally qemu_ld/st operations do not save back
> globals to memory, but only copy them there. In case of an exception
> the globals have the correct values, and otherwise they do not have
> to be reloaded.
> 
> Overall this greatly reduces the number of moves emitted, and spread
> them all over the TBs, increasing the performances on in-order CPUs.
> This also reduces register spilling, especially on CPUs with few
> registers.
> 
> In practice it means the liveness analysis is providing more
> information to the register allocator, and especially when to the memory
> version of a temp with the content of the associated register. This
> means that the two are now quite linked, and that for some functions the
> code exist in two versions, one used when the liveness analysis is
> enabled which only does some checks with assert(), the other when it is
> disabled. It might be possible to keep only one version, but it implies
> de-optimizing the liveness analysis disabled case. In any case the
> checks with assert() should be kept, as they are quite useful to make
> sure nothing subtly breaks.
> 
> 
> Changes v2 -> v3:
>  - rebased against master
>  - patch 7: fixed indentation
>  - patch 14 and later: renamed TCG_CALL_NO_RG into TCG_CALL_NO_RWG
>                        renamed TCG_CALL_NO_WGSE into TCG_CALL_NO_WG_SE
>                        renamed TCG_CALL_NO_RGSE TCG_CALL_NO_RWG_SE

Reviewed-by: Richard Henderson <address@hidden>

It all looks good.


r~


> 
> Aurelien Jarno (26):
>   tcg: add temp_dead()
>   tcg: add tcg_reg_sync()
>   tcg: add temp_sync()
>   tcg: sync output arguments on liveness request
>   tcg: rework liveness analysis
>   tcg: improve tcg_reg_alloc_movi()
>   tcg: rewrite tcg_reg_alloc_mov()
>   tcg: always mark dead input arguments as dead
>   tcg: start with local temps in TEMP_VAL_MEM state
>   tcg: don't explicitly save globals and temps
>   tcg: fix some op flags
>   tcg: forbid ld/st function to modify globals
>   tcg: synchronize globals for ops with side effects
>   tcg: rework TCG helper flags
>   target-alpha: rename helper flags
>   target-arm: rename helper flags
>   target-cris: rename helper flags
>   target-i386: rename helper flags
>   target-microblaze: rename helper flags
>   target-mips: rename helper flags
>   target-ppc: rename helper flags
>   target-s390x: rename helper flags
>   target-sh4: rename helper flags
>   target-sparc: rename helper flags
>   target-xtensa: rename helper flags
>   tcg: remove compatiblity call flags
> 
>  target-alpha/helper.h      |  176 ++++++++---------
>  target-arm/helper.h        |   18 +-
>  target-cris/helper.h       |   18 +-
>  target-i386/helper.h       |    4 +-
>  target-microblaze/helper.h |    6 +-
>  target-mips/helper.h       |  106 +++++------
>  target-ppc/helper.h        |   38 ++--
>  target-s390x/helper.h      |   76 ++++----
>  target-sh4/helper.h        |    6 +-
>  target-sparc/helper.h      |   50 ++---
>  target-xtensa/helper.h     |   16 +-
>  tcg/README                 |   22 ++-
>  tcg/optimize.c             |    3 +-
>  tcg/tcg-op.h               |   18 +-
>  tcg/tcg-opc.h              |   29 ++-
>  tcg/tcg.c                  |  449 
> +++++++++++++++++++++++++++-----------------
>  tcg/tcg.h                  |   29 ++-
>  17 files changed, 595 insertions(+), 469 deletions(-)
>