Re: [PATCH] x86: clean up unnecessarily wide TEST insns

[Denys Vlasenko]

On Fri, Mar 6, 2015 at 11:24 PM, Andi Kleen <andi@xxxxxxxxxxxxxx> wrote:
> Denys Vlasenko <dvlasenk@xxxxxxxxxx> writes:
>
>> By the nature of TEST operation, it is often possible
>> to test a narrower part of the operand:
>>     "testl $3, mem"  -> "testb $3, mem",
>>     "testq $3, %rcx" -> "testb $3, %cl"
>> This results in shorter insns, because TEST insn has no
>> sign-entending byte-immediate forms unlike other ALU ops.
>
> It also results in expensive LCP stalls. Please don't do it.
> If you feel the need to change instructions around like this read
> the optimization manuals first.

Length-changing prefix (LCP) stalls result from 0x66 prefix.

(See https://software.intel.com/en-us/forums/topic/328256).

Basically, LCP happens because adding 66 byte before this instruction:

[test_opcode] [modrm] [imm32]

changes it to

[66] [test_opcode] [modrm] [imm16]

where  [imm16] has *different length* now: 2 bytes instead of 4.
This confuses decoder.


REX prefixes were carefully designed to almost never hit this case:
adding REX prefix does not change instruction length except MOVABS
and MOV [addr],RAX insn.

My patch does not add optimizations which would use 0x66 prefix.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/