maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/lib/Target/ARM/ARMHazardRecognizer.cpp

105 lines
3.4 KiB
C++
Raw Normal View History

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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
//===-- ARMHazardRecognizer.cpp - ARM postra hazard recognizer ------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
#include "ARMHazardRecognizer.h"
#include "ARMBaseInstrInfo.h"
Clean up ARM subtarget code by using Triple ADT. llvm-svn: 123276
2011-01-12 05:46:47 +08:00
#include "ARMBaseRegisterInfo.h"
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
#include "ARMSubtarget.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/Target/TargetRegisterInfo.h"
using namespace llvm;
static bool hasRAWHazard(MachineInstr *DefMI, MachineInstr *MI,
const TargetRegisterInfo &TRI) {
// FIXME: Detect integer instructions properly.
- Rename TargetInstrDesc, TargetOperandInfo to MCInstrDesc and MCOperandInfo and sink them into MC layer. - Added MCInstrInfo, which captures the tablegen generated static data. Chang TargetInstrInfo so it's based off MCInstrInfo. llvm-svn: 134021
2011-06-29 03:10:37 +08:00
const MCInstrDesc &MCID = MI->getDesc();
unsigned Domain = MCID.TSFlags & ARMII::DomainMask;
Add bundle aware API for querying instruction properties and switch the code generator to it. For non-bundle instructions, these behave exactly the same as the MC layer API. For properties like mayLoad / mayStore, look into the bundle and if any of the bundled instructions has the property it would return true. For properties like isPredicable, only return true if *all* of the bundled instructions have the property. For properties like canFoldAsLoad, isCompare, conservatively return false for bundles. llvm-svn: 146026
2011-12-07 15:15:52 +08:00
if (MI->mayStore())
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
return false;
- Rename TargetInstrDesc, TargetOperandInfo to MCInstrDesc and MCOperandInfo and sink them into MC layer. - Added MCInstrInfo, which captures the tablegen generated static data. Chang TargetInstrInfo so it's based off MCInstrInfo. llvm-svn: 134021
2011-06-29 03:10:37 +08:00
unsigned Opcode = MCID.getOpcode();
VFP single precision arith instructions can go down to NEON pipeline, but on Cortex-A8 only. llvm-svn: 126238
2011-02-23 03:53:14 +08:00
if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD)
return false;
if ((Domain & ARMII::DomainVFP) || (Domain & ARMII::DomainNEON))
return MI->readsRegister(DefMI->getOperand(0).getReg(), &TRI);
return false;
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
}
ScheduleHazardRecognizer::HazardType
Various bits of framework needed for precise machine-level selection DAG scheduling during isel. Most new functionality is currently guarded by -enable-sched-cycles and -enable-sched-hazard. Added InstrItineraryData::IssueWidth field, currently derived from ARM itineraries, but could be initialized differently on other targets. Added ScheduleHazardRecognizer::MaxLookAhead to indicate whether it is active, and if so how many cycles of state it holds. Added SchedulingPriorityQueue::HasReadyFilter to allowing gating entry into the scheduler's available queue. ScoreboardHazardRecognizer now accesses the ScheduleDAG in order to get information about it's SUnits, provides RecedeCycle for bottom-up scheduling, correctly computes scoreboard depth, tracks IssueCount, and considers potential stall cycles when checking for hazards. ScheduleDAGRRList now models machine cycles and hazards (under flags). It tracks MinAvailableCycle, drives the hazard recognizer and priority queue's ready filter, manages a new PendingQueue, properly accounts for stall cycles, etc. llvm-svn: 122541
2010-12-24 13:03:26 +08:00
ARMHazardRecognizer::getHazardType(SUnit *SU, int Stalls) {
assert(Stalls == 0 && "ARM hazards don't support scoreboard lookahead");
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
MachineInstr *MI = SU->getInstr();
if (!MI->isDebugValue()) {
// Look for special VMLA / VMLS hazards. A VMUL / VADD / VSUB following
// a VMLA / VMLS will cause 4 cycle stall.
- Rename TargetInstrDesc, TargetOperandInfo to MCInstrDesc and MCOperandInfo and sink them into MC layer. - Added MCInstrInfo, which captures the tablegen generated static data. Chang TargetInstrInfo so it's based off MCInstrInfo. llvm-svn: 134021
2011-06-29 03:10:37 +08:00
const MCInstrDesc &MCID = MI->getDesc();
if (LastMI && (MCID.TSFlags & ARMII::DomainMask) != ARMII::DomainGeneral) {
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
MachineInstr *DefMI = LastMI;
- Rename TargetInstrDesc, TargetOperandInfo to MCInstrDesc and MCOperandInfo and sink them into MC layer. - Added MCInstrInfo, which captures the tablegen generated static data. Chang TargetInstrInfo so it's based off MCInstrInfo. llvm-svn: 134021
2011-06-29 03:10:37 +08:00
const MCInstrDesc &LastMCID = LastMI->getDesc();
Don't cache the instruction and register info from the TargetMachine, because the internals of TargetMachine could change. llvm-svn: 183488
2013-06-07 13:54:19 +08:00
const TargetMachine &TM =
MI->getParent()->getParent()->getTarget();
const ARMBaseInstrInfo &TII =
*static_cast<const ARMBaseInstrInfo*>(TM.getInstrInfo());
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
// Skip over one non-VFP / NEON instruction.
Add bundle aware API for querying instruction properties and switch the code generator to it. For non-bundle instructions, these behave exactly the same as the MC layer API. For properties like mayLoad / mayStore, look into the bundle and if any of the bundled instructions has the property it would return true. For properties like isPredicable, only return true if *all* of the bundled instructions have the property. For properties like canFoldAsLoad, isCompare, conservatively return false for bundles. llvm-svn: 146026
2011-12-07 15:15:52 +08:00
if (!LastMI->isBarrier() &&
This patch combines several changes from Evan Cheng for rdar://8659675. 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. Enable these fp vmlx codegen changes for Cortex-A9. llvm-svn: 129775
2011-04-20 02:11:57 +08:00
// On A9, AGU and NEON/FPU are muxed.
Don't cache the instruction and register info from the TargetMachine, because the internals of TargetMachine could change. llvm-svn: 183488
2013-06-07 13:54:19 +08:00
!(TII.getSubtarget().isLikeA9() &&
(LastMI->mayLoad() || LastMI->mayStore())) &&
- Rename TargetInstrDesc, TargetOperandInfo to MCInstrDesc and MCOperandInfo and sink them into MC layer. - Added MCInstrInfo, which captures the tablegen generated static data. Chang TargetInstrInfo so it's based off MCInstrInfo. llvm-svn: 134021
2011-06-29 03:10:37 +08:00
(LastMCID.TSFlags & ARMII::DomainMask) == ARMII::DomainGeneral) {
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
MachineBasicBlock::iterator I = LastMI;
if (I != LastMI->getParent()->begin()) {
I = llvm::prior(I);
DefMI = &*I;
}
}
if (TII.isFpMLxInstruction(DefMI->getOpcode()) &&
(TII.canCauseFpMLxStall(MI->getOpcode()) ||
Don't cache the instruction and register info from the TargetMachine, because the internals of TargetMachine could change. llvm-svn: 183488
2013-06-07 13:54:19 +08:00
hasRAWHazard(DefMI, MI, TII.getRegisterInfo()))) {
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
// Try to schedule another instruction for the next 4 cycles.
Various bits of framework needed for precise machine-level selection DAG scheduling during isel. Most new functionality is currently guarded by -enable-sched-cycles and -enable-sched-hazard. Added InstrItineraryData::IssueWidth field, currently derived from ARM itineraries, but could be initialized differently on other targets. Added ScheduleHazardRecognizer::MaxLookAhead to indicate whether it is active, and if so how many cycles of state it holds. Added SchedulingPriorityQueue::HasReadyFilter to allowing gating entry into the scheduler's available queue. ScoreboardHazardRecognizer now accesses the ScheduleDAG in order to get information about it's SUnits, provides RecedeCycle for bottom-up scheduling, correctly computes scoreboard depth, tracks IssueCount, and considers potential stall cycles when checking for hazards. ScheduleDAGRRList now models machine cycles and hazards (under flags). It tracks MinAvailableCycle, drives the hazard recognizer and priority queue's ready filter, manages a new PendingQueue, properly accounts for stall cycles, etc. llvm-svn: 122541
2010-12-24 13:03:26 +08:00
if (FpMLxStalls == 0)
FpMLxStalls = 4;
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
return Hazard;
}
}
}
Various bits of framework needed for precise machine-level selection DAG scheduling during isel. Most new functionality is currently guarded by -enable-sched-cycles and -enable-sched-hazard. Added InstrItineraryData::IssueWidth field, currently derived from ARM itineraries, but could be initialized differently on other targets. Added ScheduleHazardRecognizer::MaxLookAhead to indicate whether it is active, and if so how many cycles of state it holds. Added SchedulingPriorityQueue::HasReadyFilter to allowing gating entry into the scheduler's available queue. ScoreboardHazardRecognizer now accesses the ScheduleDAG in order to get information about it's SUnits, provides RecedeCycle for bottom-up scheduling, correctly computes scoreboard depth, tracks IssueCount, and considers potential stall cycles when checking for hazards. ScheduleDAGRRList now models machine cycles and hazards (under flags). It tracks MinAvailableCycle, drives the hazard recognizer and priority queue's ready filter, manages a new PendingQueue, properly accounts for stall cycles, etc. llvm-svn: 122541
2010-12-24 13:03:26 +08:00
return ScoreboardHazardRecognizer::getHazardType(SU, Stalls);
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
}
void ARMHazardRecognizer::Reset() {
LastMI = 0;
Various bits of framework needed for precise machine-level selection DAG scheduling during isel. Most new functionality is currently guarded by -enable-sched-cycles and -enable-sched-hazard. Added InstrItineraryData::IssueWidth field, currently derived from ARM itineraries, but could be initialized differently on other targets. Added ScheduleHazardRecognizer::MaxLookAhead to indicate whether it is active, and if so how many cycles of state it holds. Added SchedulingPriorityQueue::HasReadyFilter to allowing gating entry into the scheduler's available queue. ScoreboardHazardRecognizer now accesses the ScheduleDAG in order to get information about it's SUnits, provides RecedeCycle for bottom-up scheduling, correctly computes scoreboard depth, tracks IssueCount, and considers potential stall cycles when checking for hazards. ScheduleDAGRRList now models machine cycles and hazards (under flags). It tracks MinAvailableCycle, drives the hazard recognizer and priority queue's ready filter, manages a new PendingQueue, properly accounts for stall cycles, etc. llvm-svn: 122541
2010-12-24 13:03:26 +08:00
FpMLxStalls = 0;
Generalize PostRAHazardRecognizer so it can be used in any pass for both forward and backward scheduling. Rename it to ScoreboardHazardRecognizer (Scoreboard is one word). Remove integer division from the scoreboard's critical path. llvm-svn: 121274
2010-12-09 04:04:29 +08:00
ScoreboardHazardRecognizer::Reset();
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
}
void ARMHazardRecognizer::EmitInstruction(SUnit *SU) {
MachineInstr *MI = SU->getInstr();
if (!MI->isDebugValue()) {
LastMI = MI;
Various bits of framework needed for precise machine-level selection DAG scheduling during isel. Most new functionality is currently guarded by -enable-sched-cycles and -enable-sched-hazard. Added InstrItineraryData::IssueWidth field, currently derived from ARM itineraries, but could be initialized differently on other targets. Added ScheduleHazardRecognizer::MaxLookAhead to indicate whether it is active, and if so how many cycles of state it holds. Added SchedulingPriorityQueue::HasReadyFilter to allowing gating entry into the scheduler's available queue. ScoreboardHazardRecognizer now accesses the ScheduleDAG in order to get information about it's SUnits, provides RecedeCycle for bottom-up scheduling, correctly computes scoreboard depth, tracks IssueCount, and considers potential stall cycles when checking for hazards. ScheduleDAGRRList now models machine cycles and hazards (under flags). It tracks MinAvailableCycle, drives the hazard recognizer and priority queue's ready filter, manages a new PendingQueue, properly accounts for stall cycles, etc. llvm-svn: 122541
2010-12-24 13:03:26 +08:00
FpMLxStalls = 0;
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
}
Generalize PostRAHazardRecognizer so it can be used in any pass for both forward and backward scheduling. Rename it to ScoreboardHazardRecognizer (Scoreboard is one word). Remove integer division from the scoreboard's critical path. llvm-svn: 121274
2010-12-09 04:04:29 +08:00
ScoreboardHazardRecognizer::EmitInstruction(SU);
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
}
void ARMHazardRecognizer::AdvanceCycle() {
Various bits of framework needed for precise machine-level selection DAG scheduling during isel. Most new functionality is currently guarded by -enable-sched-cycles and -enable-sched-hazard. Added InstrItineraryData::IssueWidth field, currently derived from ARM itineraries, but could be initialized differently on other targets. Added ScheduleHazardRecognizer::MaxLookAhead to indicate whether it is active, and if so how many cycles of state it holds. Added SchedulingPriorityQueue::HasReadyFilter to allowing gating entry into the scheduler's available queue. ScoreboardHazardRecognizer now accesses the ScheduleDAG in order to get information about it's SUnits, provides RecedeCycle for bottom-up scheduling, correctly computes scoreboard depth, tracks IssueCount, and considers potential stall cycles when checking for hazards. ScheduleDAGRRList now models machine cycles and hazards (under flags). It tracks MinAvailableCycle, drives the hazard recognizer and priority queue's ready filter, manages a new PendingQueue, properly accounts for stall cycles, etc. llvm-svn: 122541
2010-12-24 13:03:26 +08:00
if (FpMLxStalls && --FpMLxStalls == 0)
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
// Stalled for 4 cycles but still can't schedule any other instructions.
LastMI = 0;
Generalize PostRAHazardRecognizer so it can be used in any pass for both forward and backward scheduling. Rename it to ScoreboardHazardRecognizer (Scoreboard is one word). Remove integer division from the scoreboard's critical path. llvm-svn: 121274
2010-12-09 04:04:29 +08:00
ScoreboardHazardRecognizer::AdvanceCycle();
}
void ARMHazardRecognizer::RecedeCycle() {
llvm_unreachable("reverse ARM hazard checking unsupported");
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. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
}