x86: Emit LAHF/SAHF instead of PUSHF/POPF
authorJF Bastien <jfb@google.com>
Mon, 10 Aug 2015 20:59:36 +0000 (20:59 +0000)
committerJF Bastien <jfb@google.com>
Mon, 10 Aug 2015 20:59:36 +0000 (20:59 +0000)
commit010fa7174d0367d545c0d766dafc95f150ac8ecc
tree69247b72fa66c4741c7ec1ac59ec69f86e055f18
parente03b9f566ce9ee709fc83d840c1ff1d01fca35c8
x86: Emit LAHF/SAHF instead of PUSHF/POPF

NaCl's sandbox doesn't allow PUSHF/POPF out of security concerns (priviledged emulators have forgotten to mask system bits in the past, and EFLAGS's DF bit is a constant source of hilarity). Commit r220529 fixed PR20376 by saving cmpxchg's flags result using EFLAGS, this commit now generated LAHF/SAHF instead, for all of x86 (not just NaCl) because it leads to an overall performance gain over PUSHF/POPF.

As with the previous patch this code generation is pretty bad because it occurs very later, after register allocation, and in many cases it rematerializes flags which were already available (e.g. already in a register through SETE). Fortunately it's somewhat rare that this code needs to fire.

I did [[ https://github.com/jfbastien/benchmark-x86-flags | a bit of benchmarking ]], the results on an Intel Haswell E5-2690 CPU at 2.9GHz are:

| Time per call (ms)  | Runtime (ms) | Benchmark                      |
| 0.000012514         |      6257    | sete.i386                      |
| 0.000012810         |      6405    | sete.i386-fast                 |
| 0.000010456         |      5228    | sete.x86-64                    |
| 0.000010496         |      5248    | sete.x86-64-fast               |
| 0.000012906         |      6453    | lahf-sahf.i386                 |
| 0.000013236         |      6618    | lahf-sahf.i386-fast            |
| 0.000010580         |      5290    | lahf-sahf.x86-64               |
| 0.000010304         |      5152    | lahf-sahf.x86-64-fast          |
| 0.000028056         |     14028    | pushf-popf.i386                |
| 0.000027160         |     13580    | pushf-popf.i386-fast           |
| 0.000023810         |     11905    | pushf-popf.x86-64              |
| 0.000026468         |     13234    | pushf-popf.x86-64-fast         |

Clearly `PUSHF`/`POPF` are suboptimal. It doesn't really seems to be worth teaching LLVM about individual flags, at least not for this purpose.

Reviewers: rnk, jvoung, t.p.northover

Subscribers: llvm-commits

Differential revision: http://reviews.llvm.org/D6629

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@244503 91177308-0d34-0410-b5e6-96231b3b80d8
lib/CodeGen/MachineInstrBundle.cpp
lib/Target/X86/X86InstrInfo.cpp
test/CodeGen/X86/cmpxchg-clobber-flags.ll