llvm dev - Jul 2013 - [LLVMdev] [PATCH] x86: disambiguate unqualified btr, bts

On Jul 10, 2013, at 6:54 PM, Stephen Checkoway <s at pahtak.org> wrote:
> On Jul 10, 2013, at 17:44, Jim Grosbach <grosbach at apple.com>
wrote:
>> The length specifier is, as I understand it, required when the
instruction references memory but is optional (and inferred from the registers)
for the register variants.
>> 
>> The best reference I know of for the AT&T syntax is:
http://docs.oracle.com/cd/E19253-01/817-5477/817-5477.pdf
> 
> I'm not sure I'd use the documentation for the Solaris assembler as
authoritative for AT&T syntax, but page 17 does say that if the suffix is
omitted it defaults to long.
Yeah, me either. That’s part of why I’m asking for references. I’d love to know
which ones other people use. Good docs for AT&T syntax are annoyingly hard
to find.
> However, that isn't my experience with gas which uses register operands
to disambiguate, if possible (although I'm on a phone and can't check
right now).
Yep, it’s the memory reference versions which IIUC require a suffix.

I want to make sure we do the right thing here and know why we’re doing it
rather than just adding some aliases because it matches gas’ behavior.

-Jim
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130710/cee90d28/attachment.html>

On Wed, Jul 10, 2013 at 7:15 PM, Jim Grosbach <grosbach at apple.com>
wrote:
>
> On Jul 10, 2013, at 6:54 PM, Stephen Checkoway <s at pahtak.org>
wrote:
>
> On Jul 10, 2013, at 17:44, Jim Grosbach <grosbach at apple.com>
wrote:
>
> The length specifier is, as I understand it, required when the instruction
> references memory but is optional (and inferred from the registers) for the
> register variants.
>
> The best reference I know of for the AT&T syntax is:
> http://docs.oracle.com/cd/E19253-01/817-5477/817-5477.pdf
>
>
> I'm not sure I'd use the documentation for the Solaris assembler as
> authoritative for AT&T syntax, but page 17 does say that if the suffix
is
> omitted it defaults to long.
>
>
> Yeah, me either. That’s part of why I’m asking for references. I’d love to
> know which ones other people use. Good docs for AT&T syntax are
annoyingly
> hard to find.
>
> However, that isn't my experience with gas which uses register operands
to
> disambiguate, if possible (although I'm on a phone and can't check
right
> now).
>
>
> Yep, it’s the memory reference versions which IIUC require a suffix.
>
> I want to make sure we do the right thing here and know why we’re doing it
> rather than just adding some aliases because it matches gas’ behavior.
The reason it's the right thing to do is that the mem/imm forms of
btsw and btsl have exactly the same semantics.

-Eli

Eli Friedman wrote:
> The reason it's the right thing to do is that the mem/imm forms of
> btsw and btsl have exactly the same semantics.
The Intel documentation implies that this is the case:
> If the bit base operand specifies a memory location, it represents the
address of the byte in memory that contains the bit base (bit 0 of the specified
byte) of the bit string (see Figure 3-2). The offset operand then selects a bit
position within the range −231 to 231 − 1 for a register offset and 0 to 31 for
an immediate offset.
However, this doesn't seem to be true if the immediate value is greater than
15. See the attached imm.c which prints the results of bts m16,imm8 (btsw) and
bts m32,imm8 (btsl) with the immediate in [0,63]. For btsw, only the least
significant 4 bits of the immediate seem to be used whereas for btsl, only the
least significant 5 bits seem to be used.

In contrast, bts m16,r16 (btsw) and bts m32,r32 (btsl) are identical for bit
offset operands in [0,63] and are likely identical for [-2^{15}, 2^{15}-1],
although I didn't actually check the others. See the attached reg.c which
changes the immediate constraints to register constraints.

btr behaves similarly.

For the memory, immediate form without the suffix, it seems like the options are
1. If the immediate value is in [0,15], use btsl/btrl since it saves a byte,
otherwise error;
2. Follow both gas's behavior and the Solaris assembler manual Jim Grosbach
linked to which stated that unsuffixed instructions are assumed to be long and
alias bts to btsl and btr to btrl; or
3. Always error even when there is no ambiguity.

I have no opinion on which option LLVM should follow.

-- 
Stephen Checkoway



-------------- next part --------------
A non-text attachment was scrubbed...
Name: imm.c
Type: application/octet-stream
Size: 576 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130713/1669e7a7/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: reg.c
Type: application/octet-stream
Size: 578 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130713/1669e7a7/attachment-0001.obj>