Two defects, one shape: something the decoder must decide was never stated, so
it fell out of where an entry landed in an unstably sorted table.
CANONICITY. Several mnemonics name one encoding -- SHL and SAL are both /4, JE
and JZ are both 0x74. Decoding a representative of each and diffing against
llvm-mc: 74 patterns, 22 agreed, 52 did not (0F 84 -> JZ where llvm-mc says JE,
A4 -> MOVS vs MOVSB, DB E2 -> FCLEX vs FNCLEX). aliases.odin now declares 57
{alias, canonical} rows, each canonical name MEASURED from llvm-mc rather than
picked, and gen.odin drops an aliased entry as it collects decode entries -- so
the name never reaches the tables. The drop is conditional on the canonical name
covering the byte-identical encoding, which is what lets MOV be the alias at the
A0-A3/B8 moffs forms while staying the only name for 88/89. Aliases stay fully
encodable; only decoding narrows. 119 of 1350 legacy entries dropped, 74 of 74
now agree.
This also fixes the four SAL/SHL failures left by the previous commit, at their
root rather than by extending a hand-written list.
ADDRESS SIZE. 0xE3's mnemonic is chosen by which counter register it tests --
the 67h axis -- and REX.W does not affect address size at all. Modelling JRCXZ
as force_rex_w made its encoding 48 E3 cb where llvm-mc emits a bare E3 cb, gave
JECXZ the bare encoding that is really JRCXZ in long mode, and left three
indistinguishable entries. Encoding_Flags gains addr_size (2 of 6 spare bits);
the matcher refuses a form the mode cannot express, the encoder emits 67h when
it differs from the default, the decoder selects on it before the operand-size
pass. All ten cases match llvm-mc, refusals included.
The tests keep no second copy of the alias table: mnemonics_eq compares
canonical_mnemonic, and run_alias_table_test asserts the stronger property the
generator guarantees -- zero encodings spelled by two surviving mnemonics. It
caught the E3 ambiguity on its own.
Also fixed: gen.odin's write_flags enumerates Encoding_Flags by hand and
silently dropped addr_size from the generated tables -- the flag read back as
its zero value and the instruction was quietly mis-modelled.
271 passed. Bites: dropping {.SAL, .SHL} returns the prime_sieve failure and the
guard names every shift-group opcode; restoring force_rex_w decodes E3 00 as
JCXZ and encodes JRCXZ to three bytes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Riok9vMpkLmo78wsVKJHhz
