The Rosetta split covered only VEX scalar compares. #5642 measured the
same lost load after legacy CMPSS and CMPSD, and the decoder already
hands them to the matcher with the right length and RIP offset.
Recognise F3 or F2, segment overrides and an optional REX before
0F C2 /r ib, with the same predicate rule as the VEX forms. A 66, 67 or
LOCK prefix, or both F2 and F3, keeps the instruction inline as before.
DecodeInstruction returned before reading the ModRM of FF /3 and FF /5,
so a far CALL or JMP through a RIP-relative pointer had no RIP-relative
displacement. A stub that moved such a JMP copied the displacement
without relocating it.
The decoder was compared with iced-x86 (AMD decoding) over 3.99 million
encodings: every opcode of the legacy, 0F, 0F38 and 0F3A maps under 34
prefix combinations and the VEX 0F, 0F38 and 0F3A maps, each with 48
ModRM forms. Apart from this, it agrees on length, RIP-relative offset,
control flow kind and branch target for every valid Zen 2 instruction
except near branches with a 66 prefix, recorded as technical debt, and
kernel-only returns, which it treats as sequential.
Add the comparison's metadata as cases: one RIP-relative encoding per
opcode, map and immediate size, every FF /r form, and the branch forms
with their flow kind and target. The existing cases check lengths only;
a loop or jrcxz decoded as an unconditional jump passed the whole suite.
The AMD-only converter moves the instructions next to a short AMD-only
instruction into its stub, and refuses to move one that a branch
enters. It knew only direct jump and call targets, so a switch case
entered through a jump table could land inside the jump written over
the moved instructions.
The converter now also reads relative jump tables: a RIP-relative LEA
whose address lies in a loaded segment starts a table of 32-bit
offsets from that address, read while each target starts an
instruction of the segment being converted. To read them, Convert
takes the program headers and converts their executable load
segments; the relinker passes all headers, and
ElfReader::ReadCodeSegments, now unused, is removed.
The converter test builds a dispatch through such a table whose second
case enters the instruction after a 4-byte 256-bit store, and checks
that the store takes the instruction before it instead. It also checks
that reading stops at an entry that does not start an instruction,
that only a LEA forms a table address, that tables are read only
inside loaded segments, that only executable load segments are
converted, and that a segment past the end of the file fails the
conversion.
PPSA12544: one 256-bit store in Il2cppUserAssemblies moved the
instruction after it, which is case 1 of a switch; it now goes through
an island. The other relinked files are unchanged.
AI-assisted: yes
Rosetta 2 ends the process when an unaligned 256-bit store crosses into
a page the store may not write ("rosetta error: unexpectedly need to
EmulateForward on a synchronous exception") instead of raising the
fault. A store inside such a page, and a 128-bit store across one,
fault normally. The macOS write watch keeps watched pages read-only
between collections, so a guest that stores 32 bytes across the start
of a watched page ends the process.
--to-rosetta now lowers VMOVUPS, VMOVUPD and VMOVDQU with a 256-bit
memory destination to a stub that stores the low half with the same
instruction at VEX.L=0 and the high half with VEXTRACTF128 at the
operand plus 16, so each half faults like any other store. RIP-relative
operands are relocated from the instruction's place in the site. A
store whose upper half has no 32-bit displacement fails the conversion.
Aligned stores (VMOVAPS, VMOVAPD, VMOVDQA and the non-temporal forms)
cannot cross a page and are left alone.
Every store is lowered. A 4-byte store has no room for the jump to its
stub, so its stub also takes one neighbouring instruction. The
following one: a return ends the stub, a call pushes its original
return address and jumps, a jump or conditional branch is re-encoded
with a 32-bit displacement, and another 256-bit store is lowered in the
same stub. Otherwise the preceding one, unless it is a call, a jump, a
return or a trap. Register calls, LOOP, JRCXZ and branches with a
prefix that changes them do not move. When neither neighbour can move,
the store becomes a short jump to an island: a run of movable
instructions within reach is moved to its own stub, and the bytes it
frees hold the jump to the store's stub. A store with neither fails the
conversion.
The converter tests compare the halves with clang's encodings for base,
SIB, disp8 overflowing to disp32, disp32, RSP-relative, FS-absolute,
r13-based and RIP-relative operands, and check that aligned stores,
loads, register forms and masked stores are not split. They check the
stub of a short store followed by a return, a call (direct, with an
address-size prefix, and through a RIP-relative pointer), short, near
and RIP-relative jumps, short and near conditional branches and another
store; that a following register call, LOOP, JRCXZ, prefixed branch or
CLWB makes the store take the instruction before it; islands past
instructions already moved, a store that needs its own stub, a branch
target, a register call and the short-jump reach; and that a store with
no room fails.
PPSA12544: 55295 stores split in the executable, 278 of them in the
stub of the store before them; every module converts with none kept.
AI-assisted: yes
DecodeInstruction did not treat 0F 0F as having a ModRM byte, so a
RIP-relative 3DNow! operand was not reported to the analyses and the
--to-intel stub relocation that read HasRipRelativeDisp.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VZWLdTjf1S57iLgn2rcaFb
0F 0F was decoded as a two-byte instruction. 3DNow! instructions are
0F 0F ModRM [SIB] [displacement] imm8, the trailing byte selecting the
operation (AMD64 Architecture Programmer's Manual Volume 3, Appendix A,
3DNow! opcodes).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VZWLdTjf1S57iLgn2rcaFb
Rosetta runs AVX2, FMA, BMI1/2, F16C and MOVBE, but, unlike an Intel
host, it has neither RDSEED, CLWB nor RDPID. --to-rosetta converts
everything --to-intel does and lowers these three as well:
- RDSEED becomes RDRAND (the /6 form of the same opcode), which also
returns a random value and sets CF.
- CLWB only writes a cache line back, so it becomes a NOP of the same
length.
- RDPID becomes a load of 0 through an optional stub: Rosetta's TSC_AUX
is a constant, not a processor index, so every thread reports
processor 0. A site with no movable instruction after it for the jump
is kept.
The flag implies --to-intel and conflicts with --windows. It is
independent of --macos: it only changes what the conversion produces, so
--macos alone leaves these instructions as they are, as on an Intel Mac,
and a title for Rosetta is relinked with --macos --to-rosetta. Intel
output is unchanged: the target is a parameter of the converter, Intel by
default.
The x86 decoder had no XOP (AMD 8F escape) support and read 8F E8.. as
POP r/m, losing instruction boundaries. Since #2238 roots code addresses
taken by RIP-relative LEA, code analysis reaches CPU-dispatched XOP
paths, such as the vpperm routines in Stremio Plus (PPSA74126), OpenNOW
(PPSA99082) and PPSA99420, and their Windows conversion failed with
"Code analysis: overlapping instruction boundaries". Decode 8F followed
by a map of 8 or more as a three-byte XOP prefix (8F can only be POP r/m
with ModRM.reg 0) with a ModRM operand, an 8-bit immediate in map 8 and
a 32-bit immediate in map 0Ah, in both the length and operand decoders.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ClassifyInstruction assumed a ModRM byte right after every opcode that
was not a near branch. A length-changing rewrite therefore adjusted
immediates such as MOVABS imm64 that looked like a RIP-relative ModRM,
and missed RIP-relative operands of VEX and 0F 38 / 0F 3A instructions
and the rel8 targets of LOOP and JRCXZ. Take the RIP-relative
displacement and branch target from DecodeInstruction instead.
UD1 (0F B9 /r) always raises #UD, like UD2, but DecodeInstruction
reported it as sequential. Reachability analysis then followed the
bytes after it, so an import reachable only through that fallthrough
was kept by the strict unused NID filter. UD1 keeps its ModRM, so a
RIP-relative operand is still reported.
SHA-1 and SHA-256 instructions with a [rip+disp32] operand, and
instructions with one that a short site absorbs, now move into the
stub. The stub records each displacement as a relocation relative to
the site, and the Linux, guest and Windows placements resolve it once
the stub address is known. Indirect branches through [rip+disp32]
still fail the relink.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each in-place substitution went through X64InstructionRewriter::Rewrite,
which copies the whole code segment and decodes every instruction in it
to adjust references. The converter only accepts equal-length
replacements, so the adjustment is always zero and the copy and scan are
wasted: conversion time grew by one segment scan per substitution. Copy
the replacement bytes into the segment instead.
PUSH GS, POP GS and RSM (0F A8, 0F A9, 0F AA) have no ModRM byte, but fell in the 0F A3-AB ModRM range, so the decoder read the next byte as a ModRM and sized its displacement. MOV to and from control and debug registers (0F 20-23) ignore the ModRM mod field and always name a register, but were sized and reported as memory forms, including a RIP-relative disp32 for mod 00 r/m 101.
The decoder adds the TEST immediate of opcodes F6 and F7 when ModRM.reg
is 0 or 1. The check excluded VEX and two-byte opcodes but not EVEX, so
an EVEX.0F F6 (VPSADBW) whose ModRM.reg is 0 or 1 was given an extra
imm8 and decoded one byte too long. VPSADBW has no immediate (Intel SDM
Vol. 2B, VPSADBW).
A0-A3 (MOV AL/eAX/rAX to and from moffs) have no ModRM byte and carry an
8-byte absolute address in 64-bit mode, or 4 bytes with a 0x67 prefix.
The decoder added nothing for the address, so the linear scanner decoded
the address bytes as code.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In 64-bit mode REX.W takes precedence over the 0x66 operand-size prefix,
so the immediate of 81, C7, 69, 68, A9, the 05..3D accumulator forms and
the F7 TEST form stays imm32. The decoder only applied this to B8+r and
returned a length two bytes short for the others.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SHA1RNDS4/NEXTE/MSG1/MSG2 with a memory source made `--to-intel` fail
the whole relink. The stub now spills a scratch xmm, loads the operand
into it with the unaligned load already used for SHA-256 and runs the
register-form lowering. The SHA1RNDS4 immediate is read after the SIB
byte and displacement. RIP-relative operands still fail: the stub runs
at a different address.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SHA256RNDS2/MSG1/MSG2 with a memory source made `--to-intel` fail the
whole relink. The stub now spills a scratch xmm, loads the operand into
it with MOVDQU (AMD does not fault on a misaligned SHA operand, so the
unaligned load keeps that behaviour) and runs the register-form
lowering. RSP-based operands get the spill depth added to their
displacement. RIP-relative operands still fail: the stub runs at a
different address.
Co-Authored-By: Claude <noreply@anthropic.com>
RSQRTPS and RCPPS are approximations whose low bits differ between Intel
and AMD. Intel returns 0x3F7FF000 for 1/sqrt(1.0), and two Newton-Raphson
steps on it settle on 0.99999994 instead of 1.0. PPSA21564 renormalises
its spring quaternions that way: an identity quaternion keeps w at
0.99999994 with a zero axis, its axis-angle conversion then normalises the
zero axis into NaNs, and the NaN bone poses reach keyframed physics bodies.
A NaN body later moves to an invalid broadphase cell and the game crashes
a few seconds into its hub level.
--to-intel now replaces the VEX.128 register forms with an out-of-line stub
that computes 1/sqrt(x) with SQRTPS and DIVPS (1/x with DIVPS), keeping
the upper half of the destination cleared and the scratch register intact.
A site with no room for the jump, including one followed by a branch
target or a RIP-relative instruction, keeps the native instruction and is
reported as kept instead of failing the relink.
The APM leaves bits 127:64 of the destination undefined after EXTRQ and
INSERTQ, but hardware zeroes them, and the register-form INSERTQ stub
already does the same (#185). The other paths kept the old high qword, so
a following register-form INSERTQ that reads its length and index from
that qword computed the wrong low qword too.
Every stub path that left the high qword now ends with movq dst, dst.
The in-place PSRLQ, PMOVZX and PBLENDW forms leave it non-zero and have
no room for the movq, so those sites now take a stub. The two in-place
forms that emitted nothing now emit movq dst, dst.
Co-Authored-By: Claude <noreply@anthropic.com>
Adrià