[PATCH for-9.1 00/19] target/i386: convert 1-byte opcodes to new decoder

Paolo Bonzini posted 19 patches 3 weeks, 2 days ago
Patches applied successfully (tree, apply log)
git fetch https://github.com/patchew-project/qemu tags/patchew/20240409164323.776660-1-pbonzini@redhat.com
Maintainers: Richard Henderson <richard.henderson@linaro.org>, Paolo Bonzini <pbonzini@redhat.com>, Eduardo Habkost <eduardo@habkost.net>
include/tcg/tcg.h                           |    6 +
target/i386/helper.h                        |   11 -
target/i386/tcg/decode-new.h                |   23 +-
target/i386/tcg/shift_helper_template.h.inc |  108 -
target/i386/tcg/int_helper.c                |   34 -
target/i386/tcg/translate.c                 | 3717 ++++---------------
target/i386/tcg/decode-new.c.inc            |  602 ++-
target/i386/tcg/emit.c.inc                  | 1560 +++++++-
8 files changed, 2914 insertions(+), 3147 deletions(-)
delete mode 100644 target/i386/tcg/shift_helper_template.h.inc
[PATCH for-9.1 00/19] target/i386: convert 1-byte opcodes to new decoder
Posted by Paolo Bonzini 3 weeks, 2 days ago
This series includes changes to the x86 TCG decoder that switch the
1-byte opcodes to the table-driven decoder (except for x87).  A few
easy 2-byte opcodes are also converted (BSWAP, SETcc, CMOVcc,
MOVZX/MOVSX and those that are extensions of 1-byte opcodes like PUSH/POP
FS/GS, LFS/LGS/LSS).

After optimization, the generated code is generally similar to what
is produced by the old decoder, with some differences for 32-bit
multiplications and rotate operations (RCL/RCR, and ROL/ROR less so).

This reaches a point where prefix decoding is done entirely in the new
decoder; when the opcode is loaded, if needed it will defer to
translate.c for the actual translation of the instruction.

Quite surprisingly, even without removing this duplicate code the
patch remove more lines than it adds, even though the table-driven
translator is theoretically more verbose (1 line per entry in the tables
plus all the function declarations for group decoders and emitters).
This shows how operand decoding is spread all over the place in
translate.c.

These have been ready for a few months; now that it seems clearer that
issue 2092 is a generic problem with vhost-user, it is time to get
this upstream.

Paolo

Based-on: <20240406223248.502699-1-richard.henderson@linaro.org>


Paolo Bonzini (19):
  target/i386: use TSTEQ/TSTNE to test low bits
  target/i386: use TSTEQ/TSTNE to check flags
  target/i386: remove mask from CCPrepare
  target/i386: do not use s->tmp0 and s->tmp4 to compute flags
  target/i386: reintroduce debugging mechanism
  target/i386: move 00-5F opcodes to new decoder
  target/i386: extract gen_far_call/jmp, reordering temporaries
  target/i386: allow instructions with more than one immediate
  target/i386: move 60-BF opcodes to new decoder
  target/i386: generalize gen_movl_seg_T0
  target/i386: move C0-FF opcodes to new decoder (except for x87)
  target/i386: merge and enlarge a few ranges for call to disas_insn_new
  target/i386: move remaining conditional operations to new decoder
  target/i386: move BSWAP to new decoder
  target/i386: port extensions of one-byte opcodes to new decoder
  target/i386: remove now-converted opcodes from old decoder
  target/i386: decode x87 instructions in a separate function
  target/i386: split legacy decoder into a separate function
  target/i386: remove duplicate prefix decoding

 include/tcg/tcg.h                           |    6 +
 target/i386/helper.h                        |   11 -
 target/i386/tcg/decode-new.h                |   23 +-
 target/i386/tcg/shift_helper_template.h.inc |  108 -
 target/i386/tcg/int_helper.c                |   34 -
 target/i386/tcg/translate.c                 | 3717 ++++---------------
 target/i386/tcg/decode-new.c.inc            |  602 ++-
 target/i386/tcg/emit.c.inc                  | 1560 +++++++-
 8 files changed, 2914 insertions(+), 3147 deletions(-)
 delete mode 100644 target/i386/tcg/shift_helper_template.h.inc

-- 
2.44.0