summaryrefslogtreecommitdiffstats
path: root/llvm/lib/Target/ARM/ARMBaseRegisterInfo.cpp
diff options
context:
space:
mode:
authorChris Bieneman <beanz@apple.com>2014-07-16 20:13:31 +0000
committerChris Bieneman <beanz@apple.com>2014-07-16 20:13:31 +0000
commitdf4b763be5897a315751cd071953e11c7d28a175 (patch)
treefbe63381a6bc5424465127332733ba75f4929f36 /llvm/lib/Target/ARM/ARMBaseRegisterInfo.cpp
parent2458bec7e740643bdfa3c0e89d60704e225b78cb (diff)
downloadbcm5719-llvm-df4b763be5897a315751cd071953e11c7d28a175.tar.gz
bcm5719-llvm-df4b763be5897a315751cd071953e11c7d28a175.zip
[RegisterCoalescer] Moving the RegisterCoalescer subtarget hook onto the TargetRegisterInfo instead of the TargetSubtargetInfo.
llvm-svn: 213188
Diffstat (limited to 'llvm/lib/Target/ARM/ARMBaseRegisterInfo.cpp')
-rw-r--r--llvm/lib/Target/ARM/ARMBaseRegisterInfo.cpp59
1 files changed, 59 insertions, 0 deletions
diff --git a/llvm/lib/Target/ARM/ARMBaseRegisterInfo.cpp b/llvm/lib/Target/ARM/ARMBaseRegisterInfo.cpp
index cdd91c7a703..32b5f4aa294 100644
--- a/llvm/lib/Target/ARM/ARMBaseRegisterInfo.cpp
+++ b/llvm/lib/Target/ARM/ARMBaseRegisterInfo.cpp
@@ -38,6 +38,8 @@
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
+#define DEBUG_TYPE "arm-register-info"
+
#define GET_REGINFO_TARGET_DESC
#include "ARMGenRegisterInfo.inc"
@@ -775,3 +777,60 @@ ARMBaseRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
MI.getOperand(FIOperandNum).ChangeToRegister(ScratchReg, false, false,true);
}
}
+
+bool ARMBaseRegisterInfo::shouldCoalesce(MachineInstr *MI,
+ const TargetRegisterClass *SrcRC,
+ unsigned SubReg,
+ const TargetRegisterClass *DstRC,
+ unsigned DstSubReg,
+ const TargetRegisterClass *NewRC) const {
+ auto MBB = MI->getParent();
+ auto MF = MBB->getParent();
+ const MachineRegisterInfo &MRI = MF->getRegInfo();
+ // If not copying into a sub-register this should be ok because we shouldn't
+ // need to split the reg.
+ if (!DstSubReg)
+ return true;
+ // Small registers don't frequently cause a problem, so we can coalesce them.
+ if (NewRC->getSize() < 32 && DstRC->getSize() < 32 && SrcRC->getSize() < 32)
+ return true;
+
+ auto NewRCWeight =
+ MRI.getTargetRegisterInfo()->getRegClassWeight(NewRC);
+ auto SrcRCWeight =
+ MRI.getTargetRegisterInfo()->getRegClassWeight(SrcRC);
+ auto DstRCWeight =
+ MRI.getTargetRegisterInfo()->getRegClassWeight(DstRC);
+ // If the source register class is more expensive than the destination, the
+ // coalescing is probably profitable.
+ if (SrcRCWeight.RegWeight > NewRCWeight.RegWeight)
+ return true;
+ if (DstRCWeight.RegWeight > NewRCWeight.RegWeight)
+ return true;
+
+ // If the register allocator isn't constrained, we can always allow coalescing
+ // unfortunately we don't know yet if we will be constrained.
+ // The goal of this heuristic is to restrict how many expensive registers
+ // we allow to coalesce in a given basic block.
+ auto AFI = MF->getInfo<ARMFunctionInfo>();
+ auto It = AFI->getCoalescedWeight(MBB);
+
+ DEBUG(dbgs() << "\tARM::shouldCoalesce - Coalesced Weight: "
+ << It->second << "\n");
+ DEBUG(dbgs() << "\tARM::shouldCoalesce - Reg Weight: "
+ << NewRCWeight.RegWeight << "\n");
+
+ // This number is the largest round number that which meets the criteria:
+ // (1) addresses PR18825
+ // (2) generates better code in some test cases (like vldm-shed-a9.ll)
+ // (3) Doesn't regress any test cases (in-tree, test-suite, and SPEC)
+ // In practice the SizeMultiplier will only factor in for straight line code
+ // that uses a lot of NEON vectors, which isn't terribly common.
+ unsigned SizeMultiplier = MBB->size()/100;
+ SizeMultiplier = SizeMultiplier ? SizeMultiplier : 1;
+ if (It->second < NewRCWeight.WeightLimit * SizeMultiplier) {
+ It->second += NewRCWeight.RegWeight;
+ return true;
+ }
+ return false;
+}
OpenPOWER on IntegriCloud