[llvm-commits] [llvm] r61424 - in /llvm/trunk/lib/Target/X86: X86Instr64bit.td X86InstrInfo.td

[Eli Friedman]

On Wed, Dec 24, 2008 at 9:00 PM, Chris Lattner <clattner at apple.com> wrote:
>
> On Dec 24, 2008, at 7:28 PM, Eli Friedman wrote:
>
>> On Wed, Dec 24, 2008 at 5:27 PM, Chris Lattner <sabre at nondot.org>
>> wrote:
>>> Author: lattner
>>> Date: Wed Dec 24 19:27:10 2008
>>> New Revision: 61424
>>>
>>> URL: http://llvm.org/viewvc/llvm-project?rev=61424&view=rev
>>> Log:
>>> BT memory operands load from their address operand.
>>
>> Do we really want to map this to the same intrinsic as the register
>> form?  "mov (%esp), %eax; bt %ebx, %eax;" has different semantics from
>> "bt %ebx, (%esp)", and the latter form is significantly slower.
>
> It does?  How so?  The later form reduces register pressure, so absent
> a difference in semantics it is preferable.

Suppose ebx contains the number 32.  The register form modifies the
bottom bit of eax; the memory form modifies bottom bit of the
*following* integer in memory.

Also, even ignoring that, performance is hugely different: on a Core
2, "bt %ebx, %eax" is one uop, but "bt %ebx, (%esp)" is 10 uops.  The
difference isn't quite as severe on other processors, but the reg-reg
form is still significantly faster even if a load from memory is
necessary.

-Eli