On 08.08.2019 01:58, H.J. Lu wrote:
On Wed, Aug 7, 2019 at 1:27 AM Jan Beulich <jbeulich@suse.com> wrote:
On 06.08.2019 23:38, H.J. Lu wrote:
On Tue, Aug 6, 2019 at 1:26 PM H.J. Lu <hjl.tools@gmail.com> wrote:
On Tue, Aug 6, 2019 at 7:27 AM Jan Beulich <jbeulich@suse.com> wrote:
Commit b76bc5d54e ("x86: don't default variable shift count insns to
8-bit operand size") pointed out a very bad case, but the underlying
problem is, as mentioned on various occasions, much larger: Silently
selecting a (nowhere documented afaict) certain default operand size
when there's no "sizing" suffix and no suitable register operand(s) is
simply dangerous (for the programmer to make mistakes).
If there may be an ambiguity, a size suffix can be used. Assembler isn't
responsible for programmer's errors.
While in Intel syntax mode such mistakes already lead to an error (which
is going to remain that way), AT&T syntax mode now gains warnings in
such cases by default, which can be suppressed or promoted to an error
This punishes the perfectly good assembly sources.
if so desired by the programmer. Furthermore at least general purpose
insns now consistently have a default applied (alongside the warning
emission), rather than accepting some and refusing others.
No warnings are (as before) to be generated for "DefaultSize" insns as
well as ones acting on selector and other fixed-width values. The set of
"DefaultSize" one gets slightly widened for the purposes here.
Widen the default size set avoids generate warnings. It sounds to
me these warnings isn't really necessary.
As set forth as a prereq when I first mentioned this intended change a
few years back, Linux as well as gcc have meanwhile been patched to
avoid emitting of ambiguous operands (and hence triggering of the new
warning).
Note that floating point operations with integer operands are an
exception for now: They continue to use short (16-bit) operands by
default even in 32- and 64-bit modes.
Instruction suffix has been an issue with AT&T syntax. But improvements
shouldn't assemble current working assembly sources cleanly.
I guess you mean "should", not "shouldn't"? And "cleanly" to me does
not imply without warnings, just without change to generated code
(unless the generated code was outright wrong). Pointing out possible
issues should not be restricted to cases that didn't assemble without
error so far. Or else this would be one more argument against e.g.
your recent "REP;" handling adjustment.
It is reasonable
to require that programmers should know what they are doing. We should
1. Require suffix if there may be ambiguity.
This would actively break existing code (reading "require" as "emit an
error if it is missing"). After all it's the present _inconsistency_
in behavior that this series tries to address.
2. Generate a warning if needed under a new option, (-mambiguity-check=?).
First of all this contradicts 1 above: There's no point generating a
warning when we also generate an error. As to a separate option - if
you really don't want me to re-use the existing one (which is a pretty
good fit), I can certainly key this to a new one. But the default will
remain to be for the warning to be enabled.
3. Document the default operand size for AT&T syntax if there may be ambiguity.
Will do.
Since AT&T syntax has no size info on memory operand, we can't tell memory
size when mnemonic supports more than one memory size. Currently AT&T
syntax adds suffix to distinguish different memory sizes for these instructions.
For some instructions, there is a default memory size when there is no suffix.
In this sense, there is no ambiguity for assembler. The only issue is that it
isn't easy for programmer to tell what the default memory size is.
To help programmers with these ambiguities when there is no suffix, we can
1. Document these ambiguities.
2. Issue warning when there is an exception in rules. Warning should
be kept as few as possible. For most parts, the default size is 32-bit.
"For most parts" is already telling enough, I think. Recall how
shl %cl, (%eax)
used to be wrongly deriving byte memory operand size from the
shift count register used? If things had been working flawlessly
so far, I'd maybe agree that there's no need for warnings by
default. But seeing how inconsistent _and broken_ things have
been, this is simply not an option.
add $1, (%bx)
shouldn't generate a warning since the destination is 32-bit.
Good example (but of course I don't agree with the "no warning"
statement). Why is it then that
add $0x89, (%eax)
add $0x12345678, (%eax)
fail to assemble without the patch applied? Once again - may I
ask that you look at the additions to noreg*.s that this and
the subsequent patches do? These are all cases not working
currently, but which should work (and according to you even
without warning), and which this series makes work.