rmsiface.txt

RMS NOTATION TUTORIAL

BEHAVIOR

  =: write-only
  +: read/writeable
  &: earlyclobber (impacts register sharing, defensive copying, etc.)
  %: commutative
  ,: groups alternative constraints
  ?: discourages group of constraints

SELECTION

  Autonomous

     a: ax/eax/rax
     b: bx/ebx/rbx
     c: cx/cbx/cbx
     d: dx/edx/rdx
     S: si/esi/rsi
     D: di/edi/rdi
    Yz: %xmm0
    Yp: integer register when TARGET_PARTIAL_REG_STALL (?)
    Ya: ???
    otherwise use "g" and give it a variable

  Algorithmic

    r:  pick one of a,b,c,d,D,S,r8-15 registers, referenced as %0,etc.
    q:  pick one of a,b,c,d,r8-r15 for lo-byte access, e.g. %b0,%w0,%k0,etc.
    Q:  pick one of a,b,c,d for hi-byte access, e.g. %h0,etc.
    U:  pick one of a,c,d,D,S,r8-11 (call-clobbered)
    R:  pick one of a,b,c,d,di,si,bp,sp (all models)
    y:  pick mmx register (or plain floating point?)
    x:  pick sse register
    m:  memory
    o:  memory offsetable by an immediate, referenced as %0,2+%0,etc.
    p:  memory, intended for load/push address and segments (movl %@:%p1, %0)
    g:  probably shorthand for "rmi" combo
    X:  allow anything
    k:  pick evex operand mask register i.e. k0-7 (avx512)
    v:  pick sse register w/ evex encoding i.e. %xmm0-%xmm31 (avx512)
    Yr: pick sse register w/o rex prefix (ignored if -mavx)
    Yk: pick evex operand mask register usable as predicate i.e. k1-7 (avx512)
    Yd: pick sse that's evex encodable for -mavx512dq (otherwise any sse4.1)
    Yv: pick sse that's evex encodable for -mavx512vl (otherwise any sse)
    Yh: pick sse that's evex encodable with number factor of four
    Yb: pick register for GOT base (intended for -fno-plt and __tls_get_addr)

  Combos

     rm: pick register or memory address (converting immediates)
    rmi: pick register or memory address (allowing immediates)
    etc.

  Immediates

    i: integer literal or compiler/assembler constexpr or linker embedding
    n: integer literal or compiler constexpr?
    s: integer constexpr but not literal (or known at link time?)
    C: SSE constant zero operand
    E: const double
    F: const double or const vector
    G: 1=+1.0,2=log₂10?,3=log₂𝑒?,4=π?,3=log₁₀2?,4=logₑ2?,5=+0.0? (fld?)
    e: i∊[-2^31,2^31) for sign-extending immediates
    Z: i∊[0,2^32) for zero-extending immediates
    I: i∊[0,31] (5 bits for 32-bit shifts)
    J: i∊[0,63] (6 bits for 64-bit shifts)
    K: i∊[-128,127]
    L: permit uncasted char/short literal as zero-extended operand to andl?
    M: i∊[0,3] (intended for index scaling, e.g. "mov\t(?,?,1<<%0),?")
    N: i∊[0,255] (for in & out)
    O: i∊[0,32] (wut?)
    M: 2-bit integer constant (shifts for index scaling)
    I: 5-bit integer constant (for 32-bit shifts)
    J: 6-bit integer constant (for 64-bit shifts)
    K: signed 8-bit integer constant
    G: 80387 constant
    C: sse constant zero operand

  Addresses

    Ts: address w/o segment register
    Ti: address w/o index or rip (see also mpx)
    Tb: address w/o base or rip (see also mpx)
    Tv: evex-sib address operand (avx512)

  Transcendentals

     f: any stack slot
     t: top of stack (%st(0)) and possibly converts xmm to ymm
     u: second top of stack (%st(1))
    Ts: address operand without segment register

  Specials

    %[{}|] avx512 syntax escaping (whew)
    %%REG  explicitly-supplied register (used w/ clobbers)
    %=     generates number unique to each instance

AUGMENTATION

  %pN    print raw
  %PN    print w/ @plt
  %aN    print address
  %bN    print lo-byte form, e.g. xchgb %b0,%%al (QImode 8-bit)
  %hN    print hi-byte form, e.g. xchgb %h0,%%ah (QImode 8-bit)
  %wN    print lo-word form, e.g. xchgw %w0,%%ax (HImode 16-bit)
  %kN    print dword form, e.g. xchgl %k0,%%eax (SImode 32-bit)
  %qN    print qword form, e.g. xchgq %q0,%%rax (DImode 64-bit)
  %HN    access high 8 bytes of SSE register, or +8 displacement
  %kN    print 32-bit form of operand, e.g. xchgl %k0,%%eax
  %cN    print constant without punctuation, e.g. lea %c0(%1),%2
  %lN    print label without punctuation, e.g. jumps
  %nN    negated literal, e.g. lea %n0(%1),%2
  %zN    print only opcode suffix for operand type
  %VN    print register name without %, e.g. call foo%V0

INSTRUCTION NOTATION

  Bell Laboratories

    mov $1,r0
    mov r0,r2
    mov r0,var
    mov r2,(r3)
    mov 16(r2,r5),r0
    / fortran

  NASA

    mov eax, 1
    mov ebx, eax
    mov offset var, eax
    mov dword ptr [ecx], ebx
    add eax, dword ptr [fs:ebx+esi*4+42h]
    ; etc.

  Gnubell

    movw $1,%ax
    movl %eax,%ebx
    movl %eax,var
    movl %ebx,(%ecx)
    addl 0x42(%fs,%ebx,%esi,4),%eax
    lock addl $0x12331337,%fs:-0x1337(%ebx,%esi,1)
    / not guaranteed
    # not guaranteed

NOTES

  Lettering is arbitrarily defined by a lisp config file dating back
  to gcc 1.x and each target author is granted freedom to choose his
  or her own lettering which community reserves ability to change as
  these aren't standardized in any way so please use w/ some caution
  and note 4.2.1 holds a certain degree of significance in timelines

Shouldn’t line 18 be c: cx/ecx/rcx? https://gcc.gnu.org/onlinedocs/gcc/Machine-Constraints.html#Machine-Constraints Seems to say so (“the c register”).

Thanks @swolchok!