mbox series

[00/48] tcg: optimize redundant sign extensions

Message ID 20210830062451.639572-1-richard.henderson@linaro.org
Headers show
Series tcg: optimize redundant sign extensions | expand

Message

Richard Henderson Aug. 30, 2021, 6:24 a.m. UTC 
Currently, we have support for optimizing redundant zero extensions,
which I think was done with x86 and aarch64 in mind, which zero-extend
all 32-bit operations into the 64-bit register.

But targets like Alpha, MIPS, and RISC-V do sign-extensions instead.
The last 5 patches address this.

But before that, split the quite massive tcg_optimize function.


r~


Richard Henderson (48):
  tcg/optimize: Rename "mask" to "z_mask"
  tcg/optimize: Split out OptContext
  tcg/optimize: Remove do_default label
  tcg/optimize: Change tcg_opt_gen_{mov,movi} interface
  tcg/optimize: Move prev_mb into OptContext
  tcg/optimize: Split out init_arguments
  tcg/optimize: Split out copy_propagate
  tcg/optimize: Split out fold_call
  tcg/optimize: Drop nb_oargs, nb_iargs locals
  tcg/optimize: Change fail return for do_constant_folding_cond*
  tcg/optimize: Return true from tcg_opt_gen_{mov,movi}
  tcg/optimize: Split out finish_folding
  tcg/optimize: Use a boolean to avoid a mass of continues
  tcg/optimize: Split out fold_mb, fold_qemu_{ld,st}
  tcg/optimize: Split out fold_const{1,2}
  tcg/optimize: Split out fold_setcond2
  tcg/optimize: Split out fold_brcond2
  tcg/optimize: Split out fold_brcond
  tcg/optimize: Split out fold_setcond
  tcg/optimize: Split out fold_mulu2_i32
  tcg/optimize: Split out fold_addsub2_i32
  tcg/optimize: Split out fold_movcond
  tcg/optimize: Split out fold_extract2
  tcg/optimize: Split out fold_extract, fold_sextract
  tcg/optimize: Split out fold_deposit
  tcg/optimize: Split out fold_count_zeros
  tcg/optimize: Split out fold_bswap
  tcg/optimize: Split out fold_dup, fold_dup2
  tcg/optimize: Split out fold_mov
  tcg/optimize: Split out fold_xx_to_i
  tcg/optimize: Split out fold_xx_to_x
  tcg/optimize: Split out fold_xi_to_i
  tcg/optimize: Add type to OptContext
  tcg/optimize: Split out fold_to_not
  tcg/optimize: Split out fold_sub_to_neg
  tcg/optimize: Split out fold_xi_to_x
  tcg/optimize: Split out fold_ix_to_i
  tcg/optimize: Split out fold_masks
  tcg/optimize: Expand fold_mulu2_i32 to all 4-arg multiplies
  tcg/optimize: Expand fold_addsub2_i32 to 64-bit ops
  tcg/optimize: Sink commutative operand swapping into fold functions
  tcg/optimize: Add more simplifications for orc
  tcg/optimize: Stop forcing z_mask to "garbage" for 32-bit values
  tcg/optimize: Optimize sign extensions
  tcg/optimize: Propagate sign info for logical operations
  tcg/optimize: Propagate sign info for setcond
  tcg/optimize: Propagate sign info for bit counting
  tcg/optimize: Propagate sign info for shifting

 tcg/optimize.c | 2594 +++++++++++++++++++++++++++++-------------------
 1 file changed, 1577 insertions(+), 1017 deletions(-)

-- 
2.25.1

Comments

Philippe Mathieu-Daudé Aug. 30, 2021, 10 p.m. UTC | #1 
Hi Richard,

On 8/30/21 8:24 AM, Richard Henderson wrote:
> Currently, we have support for optimizing redundant zero extensions,

> which I think was done with x86 and aarch64 in mind, which zero-extend

> all 32-bit operations into the 64-bit register.

> 

> But targets like Alpha, MIPS, and RISC-V do sign-extensions instead.

> The last 5 patches address this.

> 

> But before that, split the quite massive tcg_optimize function.


Looks promising after quick review. Do you have any comparison perf numbers?
Richard Henderson Aug. 31, 2021, 3:50 a.m. UTC | #2 
On 8/30/21 3:00 PM, Philippe Mathieu-Daudé wrote:
> Hi Richard,

> 

> On 8/30/21 8:24 AM, Richard Henderson wrote:

>> Currently, we have support for optimizing redundant zero extensions,

>> which I think was done with x86 and aarch64 in mind, which zero-extend

>> all 32-bit operations into the 64-bit register.

>>

>> But targets like Alpha, MIPS, and RISC-V do sign-extensions instead.

>> The last 5 patches address this.

>>

>> But before that, split the quite massive tcg_optimize function.

> 

> Looks promising after quick review. Do you have any comparison perf numbers?


No, not yet.  I expect there to be *lots* of redundant extensions with the patches for 
RV32 on RV64.  I think only MIPS n32 would compare, and I don't have a chroot for that, 
only n64.


r~