[PATCH 00/16] tcg: Better handling of constants

Richard Henderson posted 16 patches 4 years ago
Failed in applying to current master (apply log)
There is a newer version of this series
include/exec/gen-icount.h    |  25 ++--
include/tcg/tcg-op.h         |  17 +--
include/tcg/tcg-opc.h        |  11 +-
include/tcg/tcg.h            |  30 +++++
accel/tcg/plugin-gen.c       |  49 ++++---
tcg/aarch64/tcg-target.inc.c |  12 +-
tcg/arm/tcg-target.inc.c     |   1 -
tcg/i386/tcg-target.inc.c    | 110 +++++++++-------
tcg/mips/tcg-target.inc.c    |   2 -
tcg/optimize.c               | 204 ++++++++++++++---------------
tcg/ppc/tcg-target.inc.c     |  24 ++--
tcg/riscv/tcg-target.inc.c   |   2 -
tcg/s390/tcg-target.inc.c    |   2 -
tcg/sparc/tcg-target.inc.c   |   2 -
tcg/tcg-op-vec.c             |  83 ++++++------
tcg/tcg-op.c                 | 216 ++++++++++++++----------------
tcg/tcg.c                    | 246 +++++++++++++++++++++++++++++------
tcg/tci.c                    |   4 +-
tcg/tci/tcg-target.inc.c     |   6 +-
19 files changed, 602 insertions(+), 444 deletions(-)
[PATCH 00/16] tcg: Better handling of constants
Posted by Richard Henderson 4 years ago
This promotes constants to full-fledged temporaries, which are then
hashed so we have only a single copy across the TB.  If an opcode
requires forcing one into a register, then we will only do this
once -- at least until the register is killed, e.g. by a function call.

While this is probably an modest improvement for integer code, it is
significant for SVE vector code.  In particular, where a generator function
loads a constant for the operation, and then the generator function is
called N times for the N x 128-bit vector.  Previously we'd be loading
up the same constant N times and now we do so only once.

The existing tcg_const_{type}() functions are unchanged, allocating
a new temporary and initializing it.  The constant propagation pass
of the optimizer will generally remove the temporary when it turns
out not to be modified further.

This adds new tcg_constant_{type}() functions which produce a read-only
temporary containing the constant which need not be freed.  I have
updated the generic expanders to take advantage of this, but have not
touched the target front ends.

This also, in the end, allows the complete removal of the tcg opcodes
that create a constant: INDEX_op_movi_{i32,i64} and INDEX_op_dupi_vec.
Loading of constants into hard registers is completely controlled by
the register allocator.


r~


Richard Henderson (16):
  tcg: Add temp_readonly
  tcg: Introduce TYPE_CONST temporaries
  tcg: Use tcg_constant_i32 with icount expander
  tcg: Use tcg_constant_{i32,i64} with tcg int expanders
  tcg: Use tcg_constant_{i32,vec} with tcg vec expanders
  tcg: Use tcg_constant_{i32,i64} with tcg plugins
  tcg: Rename struct tcg_temp_info to TempOptInfo
  tcg/optimize: Adjust TempOptInfo allocation
  tcg/optimize: Use tcg_constant_internal with constant folding
  tcg/tci: Add special tci_movi_{i32,i64} opcodes
  tcg: Remove movi and dupi opcodes
  tcg: Use tcg_out_dupi_vec from temp_load
  tcg: Increase tcg_out_dupi_vec immediate to int64_t
  tcg: Add tcg_reg_alloc_dup2
  tcg/i386: Use tcg_constant_vec with tcg vec expanders
  tcg: Remove tcg_gen_dup{8,16,32,64}i_vec

 include/exec/gen-icount.h    |  25 ++--
 include/tcg/tcg-op.h         |  17 +--
 include/tcg/tcg-opc.h        |  11 +-
 include/tcg/tcg.h            |  30 +++++
 accel/tcg/plugin-gen.c       |  49 ++++---
 tcg/aarch64/tcg-target.inc.c |  12 +-
 tcg/arm/tcg-target.inc.c     |   1 -
 tcg/i386/tcg-target.inc.c    | 110 +++++++++-------
 tcg/mips/tcg-target.inc.c    |   2 -
 tcg/optimize.c               | 204 ++++++++++++++---------------
 tcg/ppc/tcg-target.inc.c     |  24 ++--
 tcg/riscv/tcg-target.inc.c   |   2 -
 tcg/s390/tcg-target.inc.c    |   2 -
 tcg/sparc/tcg-target.inc.c   |   2 -
 tcg/tcg-op-vec.c             |  83 ++++++------
 tcg/tcg-op.c                 | 216 ++++++++++++++----------------
 tcg/tcg.c                    | 246 +++++++++++++++++++++++++++++------
 tcg/tci.c                    |   4 +-
 tcg/tci/tcg-target.inc.c     |   6 +-
 19 files changed, 602 insertions(+), 444 deletions(-)

-- 
2.20.1


Re: [PATCH 00/16] tcg: Better handling of constants
Posted by Alex Bennée 4 years ago
Richard Henderson <richard.henderson@linaro.org> writes:

> This promotes constants to full-fledged temporaries, which are then
> hashed so we have only a single copy across the TB.  If an opcode
> requires forcing one into a register, then we will only do this
> once -- at least until the register is killed, e.g. by a function call.
>
> While this is probably an modest improvement for integer code, it is
> significant for SVE vector code.  In particular, where a generator function
> loads a constant for the operation, and then the generator function is
> called N times for the N x 128-bit vector.  Previously we'd be loading
> up the same constant N times and now we do so only once.
>
> The existing tcg_const_{type}() functions are unchanged, allocating
> a new temporary and initializing it.  The constant propagation pass
> of the optimizer will generally remove the temporary when it turns
> out not to be modified further.
>
> This adds new tcg_constant_{type}() functions which produce a read-only
> temporary containing the constant which need not be freed.  I have
> updated the generic expanders to take advantage of this, but have not
> touched the target front ends.
>
> This also, in the end, allows the complete removal of the tcg opcodes
> that create a constant: INDEX_op_movi_{i32,i64} and INDEX_op_dupi_vec.
> Loading of constants into hard registers is completely controlled by
> the register allocator.
>
>
> r~
>
>
> Richard Henderson (16):
>   tcg: Add temp_readonly

The series failed to apply at the first patch. Should it be based on any
other patches?

-- 
Alex Bennée

Re: [PATCH 00/16] tcg: Better handling of constants
Posted by Richard Henderson 4 years ago
On 4/20/20 8:39 AM, Alex Bennée wrote:
> The series failed to apply at the first patch. Should it be based on any
> other patches?

While I had split it out of the middle of a branch, I had not realized there
are dependencies.

Best guess is the "tcg: Clean up tcg_gen_gvec_dupi interface​" series is what
you need.


r~