1
0
mirror of https://github.com/KarolS/millfork.git synced 2024-12-25 06:29:17 +00:00
millfork/docs/lang/x86disclaimer.md
2019-07-15 14:21:50 +02:00

3.8 KiB

< back to index

8086 support disclaimer

Millfork does not support Intel 8086 directly. Instead, it generates Intel 8080 code and translates it automatically to 8086 machine code. For convenience, most undocumented 8085 instructions and Z80 instructions using IX are also translated.

This means that:

  • code is going to be large and slow;

  • there is no support for writing 8086 assembly;

  • Millfork currently translates majority of Intel 8085 assembly instructions to 8086 machine code, so you can write 8080/Z80 assembly instead.
    Instructions RST (8080), RIM, SIM (8085), RSTV, ARHL, RLDE (8085 undocumented) are not supported.

For example, code like

asm {
  LD HL, x
  LD BC, i
  ADD HL, BC
  JP C, skip
  LD A, IX(5)
  LD (HL), A
  skip:
}

is compiled to

  MOV BX, x
  MOV CX, i
  ADD BX, CX
  JNC .next
  JMP skip
.next:
  MOV AL, BYTE PTR [BP+5]
  MOV BYTE PTR [BX], AL    
skip:

Generated assembly output uses Intel 8086 syntax.

Configuring code generation

There are three options that influence the 8086 code generation:

  • ix_stack (command line equivalent -fuse-ix-for-stack for enabling, -fno-use-index-for-stack for disabling)

  • emit_8085 (command line equivalent -f8085-ops for enabling, -fno-8085-ops for disabling)

  • emit_illegals (command line equivalent -fillegals for enabling, -fno-illegals for disabling)

emit_8085 and emit_illegals have effect only together.

Major deficiencies of generated code

  • hardware multiplication is not used

  • many 16-bit operations are not used

  • many operations are restricted to the AL register

  • the overflow flag is not used

  • signed comparisons are suboptimal and as buggy as on 8080 (8085 has the undocumented K flag that could be used here, but Millfork does not use it)

  • DAS is not used

  • conditional jumps are never optimized to short 2-byte jumps and always use 5 bytes

  • the SI register is reloaded before every use

  • the converter translates the DAD instruction (16-bit ADD on Z80) to ADD, which may change behaviour of assembly code, as 8080's DAD changes only the carry flag, and 8086's ADD changes many flags. Luckily, this is not an issue with Millfork code, as the optimizer does not assume anything about flags after that instruction. The proper sequence is LAHF/ADD r1,r2/RCR SI,1/SAHF/RCL SI,1, but it is obviously too slow.

  • the converter translates the INX instruction (16-bit INC on Z80) to INC, and similarly, the DCX instruction (16-bit DEC on Z80) to DEC , which may change behaviour of assembly code, as 8080's INX and DCX don't change any flags, and 8086's INC and DEC change many flags. Luckily, this is not an issue with Millfork code, as the optimizer does not assume anything about flags after that instruction. The proper sequence is LAHF/INC r (or DEC r)/SAHF, but it is obviously too slow.

Register mapping

The registers are translated as following:

          A → AL  
 B → CH   C → CL   BC → CX  
 D → DH   E → DL   DE → DX  
 H → BH   L → BL   HL → BX  
                   SP → SP
                   IX → BP

The AH register is used as a temporary register for holding the flags.

The SI register is used as a temporary register for holding the address in 8080's LDAX/STAX (LD (DE),A/LD(BC),A/LD A,(DE)/LD A,(BC) on Z80) and 8085's undocumented LDSI/SHLX/LHLX (LD DE,SP+n/LD (DE),HL/LD HL,(DE) in Z80 syntax).

The DI register is currently not used.

Future development

There won't be any major future development related to 8086 support, unless a full 8086 backend that is independent from the 8080 backend is created.

The current solution was developed only as a proof of concept.