diff options
author | Hal Finkel <hfinkel@anl.gov> | 2014-03-29 05:29:01 +0000 |
---|---|---|
committer | Hal Finkel <hfinkel@anl.gov> | 2014-03-29 05:29:01 +0000 |
commit | 44b2b9dc1a6192fda90990ec9eec922e3f8d2049 (patch) | |
tree | 799b084ff01548b0c8e4e2a051363a6a4ac11fde /lib/Target/PowerPC/PPCInstrInfo.cpp | |
parent | c06afdcb65acd3f1fb28ce6280fed3a2d0db764c (diff) | |
download | llvm-44b2b9dc1a6192fda90990ec9eec922e3f8d2049.tar.gz llvm-44b2b9dc1a6192fda90990ec9eec922e3f8d2049.tar.bz2 llvm-44b2b9dc1a6192fda90990ec9eec922e3f8d2049.tar.xz |
[PowerPC] Add subregister classes for f64 VSX values
We had stored both f64 values and v2f64, etc. values in the VSX registers. This
worked, but was suboptimal because we would always spill 16-byte values even
through we almost always had scalar 8-byte values. This resulted in an
increase in stack-size use, extra memory bandwidth, etc. To fix this, I've
added 64-bit subregisters of the Altivec registers, and combined those with the
existing scalar floating-point registers to form a class of VSX scalar
floating-point registers. The ABI code has also been enhanced to use this
register class and some other necessary improvements have been made.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/Target/PowerPC/PPCInstrInfo.cpp')
-rw-r--r-- | lib/Target/PowerPC/PPCInstrInfo.cpp | 33 |
1 files changed, 27 insertions, 6 deletions
diff --git a/lib/Target/PowerPC/PPCInstrInfo.cpp b/lib/Target/PowerPC/PPCInstrInfo.cpp index 0f79901791..939bbdc6cc 100644 --- a/lib/Target/PowerPC/PPCInstrInfo.cpp +++ b/lib/Target/PowerPC/PPCInstrInfo.cpp @@ -744,6 +744,8 @@ void PPCInstrInfo::copyPhysReg(MachineBasicBlock &MBB, // copies are generated, they are close enough to some use that the // lower-latency form is preferable. Opc = PPC::XXLOR; + else if (PPC::VSFRCRegClass.contains(DestReg, SrcReg)) + Opc = PPC::XXLORf; else if (PPC::CRBITRCRegClass.contains(DestReg, SrcReg)) Opc = PPC::CROR; else @@ -815,6 +817,12 @@ PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF, getKillRegState(isKill)), FrameIdx)); NonRI = true; + } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { + NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSDX)) + .addReg(SrcReg, + getKillRegState(isKill)), + FrameIdx)); + NonRI = true; } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { assert(TM.getSubtargetImpl()->isDarwin() && "VRSAVE only needs spill/restore on Darwin"); @@ -906,6 +914,10 @@ PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXVD2X), DestReg), FrameIdx)); NonRI = true; + } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { + NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSDX), DestReg), + FrameIdx)); + NonRI = true; } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { assert(TM.getSubtargetImpl()->isDarwin() && "VRSAVE only needs spill/restore on Darwin"); @@ -1638,7 +1650,7 @@ protected: // The addend and this instruction must be in the same block. - if (AddendMI->getParent() != MI->getParent()) + if (!AddendMI || AddendMI->getParent() != MI->getParent()) continue; // The addend must be a full copy within the same register class. @@ -1646,9 +1658,18 @@ protected: if (!AddendMI->isFullCopy()) continue; - if (MRI.getRegClass(AddendMI->getOperand(0).getReg()) != - MRI.getRegClass(AddendMI->getOperand(1).getReg())) - continue; + unsigned AddendSrcReg = AddendMI->getOperand(1).getReg(); + if (TargetRegisterInfo::isVirtualRegister(AddendSrcReg)) { + if (MRI.getRegClass(AddendMI->getOperand(0).getReg()) != + MRI.getRegClass(AddendSrcReg)) + continue; + } else { + // If AddendSrcReg is a physical register, make sure the destination + // register class contains it. + if (!MRI.getRegClass(AddendMI->getOperand(0).getReg()) + ->contains(AddendSrcReg)) + continue; + } // In theory, there could be other uses of the addend copy before this // fma. We could deal with this, but that would require additional @@ -1678,8 +1699,8 @@ protected: OtherProdOp = 2; } - // If there are no killed product operands, then this transformation is - // likely not profitable. + // If there are no killed product operands, then this transformation is + // likely not profitable. if (!KilledProdOp) continue; |