Re: [Tinycc-devel] .fill .rept broken by "Clean up 64bit immediate support"

[Michael Matz]

Hi,

On Mon, 16 May 2016, Sergey Korshunoff wrote:

> > With the prev fix I still have a tccboot kernel trap on PCI bus configuration.
>
> What the purpose of the
>
> -ALT(DEF_ASM_OP2(addb, 0x80, 0, OPC_ARITH | OPC_MODRM | OPC_BWLX,
> OPT_IM, OPT_EA | OPT_REG))
> ALT(DEF_ASM_OP2(addw, 0x83, 0, OPC_ARITH | OPC_MODRM | OPC_WLX,
> OPT_IM8S, OPT_EA | OPT_REG))
> +ALT(DEF_ASM_OP2(addb, 0x80, 0, OPC_ARITH | OPC_MODRM | OPC_BWLX,
> OPT_IM, OPT_EA | OPT_REG))
>
> in the "Fix signed constants and opcode order" ?

The purpose is to prefer the short immediate forms of those instructions
if the immediate value fits into 8 bit (signed).  E.g. for
  add $-1,%ebx
the encoding should be "83 c3 ff", not "81 c3 ff ff ff ff".  I've tested
multiple combinations of signed and unsigned constants (including the
borders -129,-128,127,128,255,256) and the assembler itself does the right
thing.  Is it possible that the mishandled instructions doesn't have the
above simple form, but perhaps that the immediate is calculated with asm
expression?

Can you somehow find out which file and instruction is the one that's
miscompiled?  If you can, please attach the pre-processed source file
here.

> With thise change reverted there is no problem with tccboot kernel.

I hope this means it's relatively simple to find out which file contains 
the problem.


Ciao,
Michael.