[PATCH v2 00/14] target/i386: miscellaneous flags improvements

Paolo Bonzini posted 14 patches 3 weeks, 5 days ago
include/qemu/host-utils.h                |   9 ++
target/i386/cpu.h                        |  33 ++++--
target/i386/helper.h                     |   1 +
target/i386/tcg/helper-tcg.h             |   6 +-
target/i386/tcg/cc_helper_template.h.inc | 127 +++++++++++++++--------
target/i386/cpu-dump.c                   |  18 ++--
target/i386/tcg/cc_helper.c              |  51 +++------
target/i386/tcg/int_helper.c             |   4 +-
target/i386/tcg/translate.c              | 103 ++++++++++++------
target/i386/tcg/decode-new.c.inc         |   2 +-
target/i386/tcg/emit.c.inc               |  24 ++---
11 files changed, 226 insertions(+), 152 deletions(-)
[PATCH v2 00/14] target/i386: miscellaneous flags improvements
Posted by Paolo Bonzini 3 weeks, 5 days ago
This includes improved translation of checks, microoptimization of the
helpers, and improvements to the cc_op_* functions from Richard.

Unlike his original patches[1] I didn't convert cc_op_live() to a switch
statement, instead keeping the array but making sure that all of its
entries are nonzero.  The only zero entry was CC_OP_CLR, which is now
changed to spill the constant value of EFLAGS to cc_op_src.  While this
has a 0.2% cost in number of TCG ops, getting rid of the special
case for CC_OP_CLR makes it even easier to optimize computation of ZF
from CC_OP_DYNAMIC; this is quite common, for example in switch
statements that have CMP/JG/JE sequences (JE followed JL/JG/JA/JB seems
less common than the opposite, though that's not universal).

On a quick-and-dirty run of "ls -lR", the changes add ~750 spills of
0x44 to cc_op_src; but it also reduces to one half the calls to
cc_compute_all (most of them are completely eliminated), and that
is a lot more expensive.

One thing I noticed is that those spills are really huge (11 bytes).
It might help to move cc_* at the very beginning of CPUX86State, because
the number of accesses to cc_* is comparable to the number of accesses
to registers (despite cc_* being mostly written, while registers are
both read and written).

Thanks,

Paolo

[1] https://patchew.org/QEMU/20240701025115.1265117-1-richard.henderson@linaro.org/

v1->v2:
- use MAKE_64BIT_MASK in helper_cc_compute_nz
- get rid completely of parity_table

Paolo Bonzini (10):
  target/i386: use tcg_gen_ext_tl when applicable
  target/i386: remove CC_OP_CLR
  target/i386: optimize computation of ZF from CC_OP_DYNAMIC
  target/i386: optimize TEST+Jxx sequences
  target/i386: add a few more trivial CCPrepare cases
  target/i386: add a note about gen_jcc1
  target/i386: make flag variables unsigned
  target/i386: use compiler builtin to compute PF
  target/i386: use higher-precision arithmetic to compute CF
  target/i386: use + to put flags together

Richard Henderson (4):
  target/i386: Tidy cc_op_str usage
  target/i386: Rearrange CCOp
  target/i386: Introduce cc_op_size
  target/i386: Wrap cc_op_live with a validity check

 include/qemu/host-utils.h                |   9 ++
 target/i386/cpu.h                        |  33 ++++--
 target/i386/helper.h                     |   1 +
 target/i386/tcg/helper-tcg.h             |   6 +-
 target/i386/tcg/cc_helper_template.h.inc | 127 +++++++++++++++--------
 target/i386/cpu-dump.c                   |  18 ++--
 target/i386/tcg/cc_helper.c              |  51 +++------
 target/i386/tcg/int_helper.c             |   4 +-
 target/i386/tcg/translate.c              | 103 ++++++++++++------
 target/i386/tcg/decode-new.c.inc         |   2 +-
 target/i386/tcg/emit.c.inc               |  24 ++---
 11 files changed, 226 insertions(+), 152 deletions(-)

-- 
2.47.0