diff options
author | Evan Cheng <evan.cheng@apple.com> | 2006-02-16 01:50:02 +0000 |
---|---|---|
committer | Evan Cheng <evan.cheng@apple.com> | 2006-02-16 01:50:02 +0000 |
commit | d51425a82da3d519b85f86e380358aa22e63442d (patch) | |
tree | 29e1e895fd1e74fcc5d952093691ee40f4c3eed7 /lib | |
parent | f17c42d409ce9d981996dea11918774e84400eea (diff) | |
download | external_llvm-d51425a82da3d519b85f86e380358aa22e63442d.zip external_llvm-d51425a82da3d519b85f86e380358aa22e63442d.tar.gz external_llvm-d51425a82da3d519b85f86e380358aa22e63442d.tar.bz2 |
Use movaps / movapd (instead of movss / movsd) to do FR32 / FR64 reg to reg
transfer.
According to the Intel P4 Optimization Manual:
Moves that write a portion of a register can introduce unwanted
dependences. The movsd reg, reg instruction writes only the bottom
64 bits of a register, not to all 128 bits. This introduces a dependence on
the preceding instruction that produces the upper 64 bits (even if those
bits are not longer wanted). The dependence inhibits register renaming,
and thereby reduces parallelism.
Not to mention movaps is shorter than movss.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26226 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib')
-rw-r--r-- | lib/Target/X86/X86RegisterInfo.cpp | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/lib/Target/X86/X86RegisterInfo.cpp b/lib/Target/X86/X86RegisterInfo.cpp index fdbe422..571bef1 100644 --- a/lib/Target/X86/X86RegisterInfo.cpp +++ b/lib/Target/X86/X86RegisterInfo.cpp @@ -106,9 +106,9 @@ void X86RegisterInfo::copyRegToReg(MachineBasicBlock &MBB, } else if (RC == &X86::RFPRegClass || RC == &X86::RSTRegClass) { Opc = X86::FpMOV; } else if (RC == &X86::FR32RegClass || RC == &X86::V4F4RegClass) { - Opc = X86::MOVSSrr; + Opc = X86::MOVAPSrr; } else if (RC == &X86::FR64RegClass || RC == &X86::V2F8RegClass) { - Opc = X86::MOVSDrr; + Opc = X86::MOVAPDrr; } else { assert(0 && "Unknown regclass"); abort(); |