Making use of VFP / NEON floating point multiply-accumulate / subtraction is
difficult on current ARM implementations for a few reasons.
1. Even though a single vmla has latency that is one cycle shorter than a pair
of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause
additional pipeline stall. So it's frequently better to single codegen
vmul + vadd.
2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to
stall for 4 cycles. We need to schedule them apart.
3. A vmla followed vmla is a special case. Obvious issuing back to back RAW
vmla + vmla is very bad. But this isn't ideal either:
vmul
vadd
vmla
Instead, we want to expand the second vmla:
vmla
vmul
vadd
Even with the 4 cycle vmul stall, the second sequence is still 2 cycles
faster.
Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough
but it isn't the optimial solution. This patch attempts to make it possible to
use vmla / vmls in cases where it is profitable.
A. Add missing isel predicates which cause vmla to be codegen'ed.
B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to
compute a fmul and a fmla.
C. Add additional isel checks for vmla, avoid cases where vmla is feeding into
fp instructions (except for the #3 exceptional case).
D. Add ARM hazard recognizer to model the vmla / vmls hazards.
E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the
vmla / vmls will trigger one of the special hazards.
Work in progress, only A+B are enabled.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@120960 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 4838b7d..afec5c2 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -15,6 +15,7 @@
#include "ARM.h"
#include "ARMAddressingModes.h"
#include "ARMConstantPoolValue.h"
+#include "ARMHazardRecognizer.h"
#include "ARMMachineFunctionInfo.h"
#include "ARMRegisterInfo.h"
#include "ARMGenInstrInfo.inc"
@@ -40,9 +41,58 @@
EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden,
cl::desc("Enable ARM 2-addr to 3-addr conv"));
+
+/// ARM_MLxEntry - Record information about MLA / MLS instructions.
+struct ARM_MLxEntry {
+ unsigned MLxOpc; // MLA / MLS opcode
+ unsigned MulOpc; // Expanded multiplication opcode
+ unsigned AddSubOpc; // Expanded add / sub opcode
+ bool NegAcc; // True if the acc is negated before the add / sub.
+ bool HasLane; // True if instruction has an extra "lane" operand.
+};
+
+static const ARM_MLxEntry ARM_MLxTable[] = {
+ // MLxOpc, MulOpc, AddSubOpc, NegAcc, HasLane
+ // fp scalar ops
+ { ARM::VMLAS, ARM::VMULS, ARM::VADDS, false, false },
+ { ARM::VMLSS, ARM::VMULS, ARM::VSUBS, false, false },
+ { ARM::VMLAD, ARM::VMULD, ARM::VADDD, false, false },
+ { ARM::VMLSD, ARM::VMULD, ARM::VSUBD, false, false },
+ { ARM::VMLAfd_sfp, ARM::VMULfd_sfp, ARM::VADDfd_sfp, false, false },
+ { ARM::VMLSfd_sfp, ARM::VMULfd_sfp, ARM::VSUBfd_sfp, false, false },
+ { ARM::VNMLAS, ARM::VNMULS, ARM::VSUBS, true, false },
+ { ARM::VNMLSS, ARM::VMULS, ARM::VSUBS, true, false },
+ { ARM::VNMLAD, ARM::VNMULD, ARM::VSUBD, true, false },
+ { ARM::VNMLSD, ARM::VMULD, ARM::VSUBD, true, false },
+
+ // fp SIMD ops
+ { ARM::VMLAfd, ARM::VMULfd, ARM::VADDfd, false, false },
+ { ARM::VMLSfd, ARM::VMULfd, ARM::VSUBfd, false, false },
+ { ARM::VMLAfq, ARM::VMULfq, ARM::VADDfq, false, false },
+ { ARM::VMLSfq, ARM::VMULfq, ARM::VSUBfq, false, false },
+ { ARM::VMLAslfd, ARM::VMULslfd, ARM::VADDfd, false, true },
+ { ARM::VMLSslfd, ARM::VMULslfd, ARM::VSUBfd, false, true },
+ { ARM::VMLAslfq, ARM::VMULslfq, ARM::VADDfq, false, true },
+ { ARM::VMLSslfq, ARM::VMULslfq, ARM::VSUBfq, false, true },
+};
+
ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI)
: TargetInstrInfoImpl(ARMInsts, array_lengthof(ARMInsts)),
Subtarget(STI) {
+ for (unsigned i = 0, e = array_lengthof(ARM_MLxTable); i != e; ++i) {
+ if (!MLxEntryMap.insert(std::make_pair(ARM_MLxTable[i].MLxOpc, i)).second)
+ assert(false && "Duplicated entries?");
+ MLxHazardOpcodes.insert(ARM_MLxTable[i].AddSubOpc);
+ MLxHazardOpcodes.insert(ARM_MLxTable[i].MulOpc);
+ }
+}
+
+ScheduleHazardRecognizer *ARMBaseInstrInfo::
+CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II) const {
+ if (Subtarget.isThumb2() || Subtarget.hasVFP2())
+ return (ScheduleHazardRecognizer *)
+ new ARMHazardRecognizer(II, *this, getRegisterInfo(), Subtarget);
+ return TargetInstrInfoImpl::CreateTargetPostRAHazardRecognizer(II);
}
MachineInstr *
@@ -197,7 +247,6 @@
return NewMIs[0];
}
-
// Branch analysis.
bool
ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB,
@@ -2196,3 +2245,19 @@
}
return false;
}
+
+bool
+ARMBaseInstrInfo::isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
+ unsigned &AddSubOpc,
+ bool &NegAcc, bool &HasLane) const {
+ DenseMap<unsigned, unsigned>::const_iterator I = MLxEntryMap.find(Opcode);
+ if (I == MLxEntryMap.end())
+ return false;
+
+ const ARM_MLxEntry &Entry = ARM_MLxTable[I->second];
+ MulOpc = Entry.MulOpc;
+ AddSubOpc = Entry.AddSubOpc;
+ NegAcc = Entry.NegAcc;
+ HasLane = Entry.HasLane;
+ return true;
+}