On Thu, 3 Mar 2022, David Laight wrote:
> > And indeed it happens with the MIPS target:
> >
> > 803ae47c: 82050000 lb a1,0(s0)
> > 803ae480: 26100001 addiu s0,s0,1
> > 803ae484: 02402025 move a0,s2
> > 803ae488: 0220f809 jalr s1
> > 803ae48c: 30a500ff andi a1,a1,0xff
> >
> > vs current code:
> >
> > 803ae47c: 82050000 lb a1,0(s0)
> > 803ae480: 26100001 addiu s0,s0,1
> > 803ae484: 0220f809 jalr s1
> > 803ae488: 02402025 move a0,s2
> >
> > (NB the last instruction shown after the call instruction, JALR, is in the
> > delay slot that is executed before the PC gets updated). Now arguably the
> > compiler might notice that and use an unsigned LBU load instruction rather
> > than the signed LB load instruction, which would make the ANDI instruction
> > redundant, but still I think we ought to avoid gratuitous type signedness
> > changes.
> >
> > So I'd recommend changing `s' here to `const unsigned char *' or, as I
> > previously suggested, maybe to `const u8 *' even.
>
> Or just not worry that the 'char' value (either [128..127] or [0..255])
> is held in a 'signed int' variable.
> That basically happens every time it is loaded into a register anyway.
That might be true with a hypothetical 8-bit ABI on top of a higher-width
machine architecture. It does happen with the 32-bit MIPS ABI (o32) and a
64-bit architecture, which is why LW (load word signed) is a universal
32-bit and 64-bit instruction while the LWU (load word unsigned) operation
is restricted to 64-bit code.
In this case however a signed `char' value ([-128..127]) is sign-extended
while an unsigned `char' value ([0..255]) is zero-extended, even though
both are carried in a 'signed int' variable from the architecture's point
of view.
Anyway I have looked into it some more and the immediate cause for LBU
not to be used here is the:
if (*s == '\n')
putchar(port, '\r');
conditional. If this part is removed, then LBU does get used for the:
putchar(port, *s);
part and no ANDI is produced.
The reason for that is that the compiler decides to reuse the load used
to evaluate (*s == '\n') (which is done using the plain `char' data type)
for the following `putchar(port, *s)' call if the expression used as the
condition turns out to be false and therefore the value of `*s' has to be
subsequently zero-extended:
b4: 00e08825 move s1,a3
b8: 2413000a li s3,10
bc: 82050000 lb a1,0(s0)
c0: 00000000 nop
c4: 14b30005 bne a1,s3,dc <uart_console_write+0x54>
c8: 00000000 nop
cc: 2405000d li a1,13
d0: 0220f809 jalr s1
d4: 02402025 move a0,s2
d8: 82050000 lb a1,0(s0)
dc: 26100001 addiu s0,s0,1
e0: 02402025 move a0,s2
e4: 0220f809 jalr s1
e8: 30a500ff andi a1,a1,0xff
(the load at bc is reused for the `putchar' call at e4 unless it's `\n',
or otherwise the character is reloaded at d8).
By using a temporary `unsigned char' variable and massaging the source
code suitably GCC can be persuaded to use LBU instead, but the obfuscation
of the source code and the resulting machine code produced seem not worth
the effort IMO, so let's keep it simple.
JFTR,
Maciej
