2017-09-13 23:15:20 +02:00
|
|
|
//===- RegAllocBase.cpp - Register Allocator Base Class -------------------===//
|
2012-01-11 23:28:30 +01:00
|
|
|
//
|
2019-01-19 09:50:56 +01:00
|
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
2012-01-11 23:28:30 +01:00
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
2014-02-22 20:31:28 +01:00
|
|
|
// This file defines the RegAllocBase class which provides common functionality
|
2012-01-11 23:28:30 +01:00
|
|
|
// for LiveIntervalUnion-based register allocators.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#include "RegAllocBase.h"
|
2017-09-13 23:15:20 +02:00
|
|
|
#include "llvm/ADT/SmallVector.h"
|
2012-01-11 23:28:30 +01:00
|
|
|
#include "llvm/ADT/Statistic.h"
|
2017-09-13 23:15:20 +02:00
|
|
|
#include "llvm/CodeGen/LiveInterval.h"
|
2017-12-13 03:51:04 +01:00
|
|
|
#include "llvm/CodeGen/LiveIntervals.h"
|
2012-11-28 20:13:06 +01:00
|
|
|
#include "llvm/CodeGen/LiveRegMatrix.h"
|
2012-01-11 23:28:30 +01:00
|
|
|
#include "llvm/CodeGen/MachineInstr.h"
|
RegAlloc: try to fail more gracefully when out of registers
Summary:
The emitError path allows the program to continue, unlike report_fatal_error.
This is friendlier to use cases where LLVM is embedded in a larger program,
because the caller may be able to deal with the error somewhat gracefully.
Change the number of requested NOP bytes in the AArch64 and PowerPC
test cases to avoid triggering an unrelated assertion. The compilation
still fails, as verified by the test.
Change-Id: Iafb9ca341002a597b82e59ddc7a1f13c78758e3d
Reviewers: arsenm, MatzeB
Subscribers: qcolombet, nemanjai, wdng, javed.absar, kristof.beyls, kbarton, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61489
llvm-svn: 360786
2019-05-15 19:29:58 +02:00
|
|
|
#include "llvm/CodeGen/MachineModuleInfo.h"
|
2012-01-11 23:28:30 +01:00
|
|
|
#include "llvm/CodeGen/MachineRegisterInfo.h"
|
2020-03-08 17:36:29 +01:00
|
|
|
#include "llvm/CodeGen/Spiller.h"
|
2017-11-17 02:07:10 +01:00
|
|
|
#include "llvm/CodeGen/TargetRegisterInfo.h"
|
2012-11-28 20:13:06 +01:00
|
|
|
#include "llvm/CodeGen/VirtRegMap.h"
|
2017-09-13 23:15:20 +02:00
|
|
|
#include "llvm/Pass.h"
|
2012-01-11 23:28:30 +01:00
|
|
|
#include "llvm/Support/CommandLine.h"
|
|
|
|
#include "llvm/Support/Debug.h"
|
|
|
|
#include "llvm/Support/ErrorHandling.h"
|
|
|
|
#include "llvm/Support/Timer.h"
|
2017-06-06 13:49:48 +02:00
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2017-09-13 23:15:20 +02:00
|
|
|
#include <cassert>
|
2012-01-11 23:28:30 +01:00
|
|
|
|
|
|
|
using namespace llvm;
|
|
|
|
|
2014-04-22 04:02:50 +02:00
|
|
|
#define DEBUG_TYPE "regalloc"
|
|
|
|
|
2021-02-27 20:41:45 +01:00
|
|
|
STATISTIC(NumNewQueued, "Number of new live ranges queued");
|
2012-01-11 23:28:30 +01:00
|
|
|
|
|
|
|
// Temporary verification option until we can put verification inside
|
|
|
|
// MachineVerifier.
|
|
|
|
static cl::opt<bool, true>
|
2017-12-01 01:53:10 +01:00
|
|
|
VerifyRegAlloc("verify-regalloc", cl::location(RegAllocBase::VerifyEnabled),
|
|
|
|
cl::Hidden, cl::desc("Verify during register allocation"));
|
2012-01-11 23:28:30 +01:00
|
|
|
|
2016-11-18 20:43:18 +01:00
|
|
|
const char RegAllocBase::TimerGroupName[] = "regalloc";
|
|
|
|
const char RegAllocBase::TimerGroupDescription[] = "Register Allocation";
|
2012-01-11 23:28:30 +01:00
|
|
|
bool RegAllocBase::VerifyEnabled = false;
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// RegAllocBase Implementation
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2013-11-19 01:57:56 +01:00
|
|
|
// Pin the vtable to this file.
|
|
|
|
void RegAllocBase::anchor() {}
|
|
|
|
|
2021-02-27 20:41:45 +01:00
|
|
|
void RegAllocBase::init(VirtRegMap &vrm, LiveIntervals &lis,
|
2012-06-21 00:52:29 +02:00
|
|
|
LiveRegMatrix &mat) {
|
2012-01-11 23:28:30 +01:00
|
|
|
TRI = &vrm.getTargetRegInfo();
|
|
|
|
MRI = &vrm.getRegInfo();
|
|
|
|
VRM = &vrm;
|
|
|
|
LIS = &lis;
|
2012-06-21 00:52:29 +02:00
|
|
|
Matrix = &mat;
|
2012-11-28 01:21:29 +01:00
|
|
|
MRI->freezeReservedRegs(vrm.getMachineFunction());
|
2012-01-11 23:28:30 +01:00
|
|
|
RegClassInfo.runOnMachineFunction(vrm.getMachineFunction());
|
|
|
|
}
|
|
|
|
|
|
|
|
// Visit all the live registers. If they are already assigned to a physical
|
|
|
|
// register, unify them with the corresponding LiveIntervalUnion, otherwise push
|
|
|
|
// them on the priority queue for later assignment.
|
|
|
|
void RegAllocBase::seedLiveRegs() {
|
2016-11-18 20:43:18 +01:00
|
|
|
NamedRegionTimer T("seed", "Seed Live Regs", TimerGroupName,
|
|
|
|
TimerGroupDescription, TimePassesIsEnabled);
|
2012-06-20 23:25:05 +02:00
|
|
|
for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
|
2020-11-03 05:07:58 +01:00
|
|
|
Register Reg = Register::index2VirtReg(i);
|
2012-06-20 23:25:05 +02:00
|
|
|
if (MRI->reg_nodbg_empty(Reg))
|
|
|
|
continue;
|
|
|
|
enqueue(&LIS->getInterval(Reg));
|
2012-01-11 23:28:30 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Top-level driver to manage the queue of unassigned VirtRegs and call the
|
|
|
|
// selectOrSplit implementation.
|
|
|
|
void RegAllocBase::allocatePhysRegs() {
|
|
|
|
seedLiveRegs();
|
|
|
|
|
|
|
|
// Continue assigning vregs one at a time to available physical registers.
|
|
|
|
while (LiveInterval *VirtReg = dequeue()) {
|
2020-09-15 23:54:38 +02:00
|
|
|
assert(!VRM->hasPhys(VirtReg->reg()) && "Register already assigned");
|
2012-01-11 23:28:30 +01:00
|
|
|
|
|
|
|
// Unused registers can appear when the spiller coalesces snippets.
|
2020-09-15 23:54:38 +02:00
|
|
|
if (MRI->reg_nodbg_empty(VirtReg->reg())) {
|
2018-05-14 14:53:11 +02:00
|
|
|
LLVM_DEBUG(dbgs() << "Dropping unused " << *VirtReg << '\n');
|
[RegAllocGreedy] Introduce a late pass to repair broken hints.
A broken hint is a copy where both ends are assigned different colors. When a
variable gets evicted in the neighborhood of such copies, it is likely we can
reconcile some of them.
** Context **
Copies are inserted during the register allocation via splitting. These split
points are required to relax the constraints on the allocation problem. When
such a point is inserted, both ends of the copy would not share the same color
with respect to the current allocation problem. When variables get evicted,
the allocation problem becomes different and some split point may not be
required anymore. However, the related variables may already have been colored.
This usually shows up in the assembly with pattern like this:
def A
...
save A to B
def A
use A
restore A from B
...
use B
Whereas we could simply have done:
def B
...
def A
use A
...
use B
** Proposed Solution **
A variable having a broken hint is marked for late recoloring if and only if
selecting a register for it evict another variable. Indeed, if no eviction
happens this is pointless to look for recoloring opportunities as it means the
situation was the same as the initial allocation problem where we had to break
the hint.
Finally, when everything has been allocated, we look for recoloring
opportunities for all the identified candidates.
The recoloring is performed very late to rely on accurate copy cost (all
involved variables are allocated).
The recoloring is simple unlike the last change recoloring. It propagates the
color of the broken hint to all its copy-related variables. If the color is
available for them, the recoloring uses it, otherwise it gives up on that hint
even if a more complex coloring would have worked.
The recoloring happens only if it is profitable. The profitability is evaluated
using the expected frequency of the copies of the currently recolored variable
with a) its current color and b) with the target color. If a) is greater or
equal than b), then it is profitable and the recoloring happen.
** Example **
Consider the following example:
BB1:
a =
b =
BB2:
...
= b
= a
Let us assume b gets split:
BB1:
a =
b =
BB2:
c = b
...
d = c
= d
= a
Because of how the allocation work, b, c, and d may be assigned different
colors. Now, if a gets evicted to make room for c, assuming b and d were
assigned to something different than a.
We end up with:
BB1:
a =
st a, SpillSlot
b =
BB2:
c = b
...
d = c
= d
e = ld SpillSlot
= e
This is likely that we can assign the same register for b, c, and d,
getting rid of 2 copies.
** Performances **
Both ARM64 and x86_64 show performance improvements of up to 3% for the
llvm-testsuite + externals with Os and O3. There are a few regressions too that
comes from the (in)accuracy of the block frequency estimate.
<rdar://problem/18312047>
llvm-svn: 225422
2015-01-08 02:16:39 +01:00
|
|
|
aboutToRemoveInterval(*VirtReg);
|
2020-09-15 23:54:38 +02:00
|
|
|
LIS->removeInterval(VirtReg->reg());
|
2012-01-11 23:28:30 +01:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Invalidate all interference queries, live ranges could have changed.
|
2012-06-21 00:52:29 +02:00
|
|
|
Matrix->invalidateVirtRegs();
|
2012-01-11 23:28:30 +01:00
|
|
|
|
|
|
|
// selectOrSplit requests the allocator to return an available physical
|
|
|
|
// register if possible and populate a list of new live intervals that
|
|
|
|
// result from splitting.
|
2018-05-14 14:53:11 +02:00
|
|
|
LLVM_DEBUG(dbgs() << "\nselectOrSplit "
|
2020-09-15 23:54:38 +02:00
|
|
|
<< TRI->getRegClassName(MRI->getRegClass(VirtReg->reg()))
|
|
|
|
<< ':' << *VirtReg << " w=" << VirtReg->weight() << '\n');
|
2017-09-13 23:15:20 +02:00
|
|
|
|
2020-06-30 17:57:24 +02:00
|
|
|
using VirtRegVec = SmallVector<Register, 4>;
|
2017-09-13 23:15:20 +02:00
|
|
|
|
2012-01-11 23:28:30 +01:00
|
|
|
VirtRegVec SplitVRegs;
|
2020-10-09 19:04:29 +02:00
|
|
|
MCRegister AvailablePhysReg = selectOrSplit(*VirtReg, SplitVRegs);
|
2012-01-11 23:28:30 +01:00
|
|
|
|
|
|
|
if (AvailablePhysReg == ~0u) {
|
|
|
|
// selectOrSplit failed to find a register!
|
|
|
|
// Probably caused by an inline asm.
|
2014-04-14 02:51:57 +02:00
|
|
|
MachineInstr *MI = nullptr;
|
2014-03-13 07:02:25 +01:00
|
|
|
for (MachineRegisterInfo::reg_instr_iterator
|
2020-09-15 23:54:38 +02:00
|
|
|
I = MRI->reg_instr_begin(VirtReg->reg()),
|
|
|
|
E = MRI->reg_instr_end();
|
|
|
|
I != E;) {
|
RegAlloc: try to fail more gracefully when out of registers
Summary:
The emitError path allows the program to continue, unlike report_fatal_error.
This is friendlier to use cases where LLVM is embedded in a larger program,
because the caller may be able to deal with the error somewhat gracefully.
Change the number of requested NOP bytes in the AArch64 and PowerPC
test cases to avoid triggering an unrelated assertion. The compilation
still fails, as verified by the test.
Change-Id: Iafb9ca341002a597b82e59ddc7a1f13c78758e3d
Reviewers: arsenm, MatzeB
Subscribers: qcolombet, nemanjai, wdng, javed.absar, kristof.beyls, kbarton, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61489
llvm-svn: 360786
2019-05-15 19:29:58 +02:00
|
|
|
MI = &*(I++);
|
|
|
|
if (MI->isInlineAsm())
|
2012-01-11 23:28:30 +01:00
|
|
|
break;
|
2014-03-13 07:02:25 +01:00
|
|
|
}
|
2021-01-18 15:53:55 +01:00
|
|
|
|
|
|
|
const TargetRegisterClass *RC = MRI->getRegClass(VirtReg->reg());
|
|
|
|
ArrayRef<MCPhysReg> AllocOrder = RegClassInfo.getOrder(RC);
|
|
|
|
if (AllocOrder.empty())
|
|
|
|
report_fatal_error("no registers from class available to allocate");
|
|
|
|
else if (MI && MI->isInlineAsm()) {
|
2013-10-05 21:33:37 +02:00
|
|
|
MI->emitError("inline assembly requires more registers than available");
|
RegAlloc: try to fail more gracefully when out of registers
Summary:
The emitError path allows the program to continue, unlike report_fatal_error.
This is friendlier to use cases where LLVM is embedded in a larger program,
because the caller may be able to deal with the error somewhat gracefully.
Change the number of requested NOP bytes in the AArch64 and PowerPC
test cases to avoid triggering an unrelated assertion. The compilation
still fails, as verified by the test.
Change-Id: Iafb9ca341002a597b82e59ddc7a1f13c78758e3d
Reviewers: arsenm, MatzeB
Subscribers: qcolombet, nemanjai, wdng, javed.absar, kristof.beyls, kbarton, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61489
llvm-svn: 360786
2019-05-15 19:29:58 +02:00
|
|
|
} else if (MI) {
|
|
|
|
LLVMContext &Context =
|
|
|
|
MI->getParent()->getParent()->getMMI().getModule()->getContext();
|
|
|
|
Context.emitError("ran out of registers during register allocation");
|
|
|
|
} else {
|
2013-10-05 21:33:37 +02:00
|
|
|
report_fatal_error("ran out of registers during register allocation");
|
RegAlloc: try to fail more gracefully when out of registers
Summary:
The emitError path allows the program to continue, unlike report_fatal_error.
This is friendlier to use cases where LLVM is embedded in a larger program,
because the caller may be able to deal with the error somewhat gracefully.
Change the number of requested NOP bytes in the AArch64 and PowerPC
test cases to avoid triggering an unrelated assertion. The compilation
still fails, as verified by the test.
Change-Id: Iafb9ca341002a597b82e59ddc7a1f13c78758e3d
Reviewers: arsenm, MatzeB
Subscribers: qcolombet, nemanjai, wdng, javed.absar, kristof.beyls, kbarton, jsji, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61489
llvm-svn: 360786
2019-05-15 19:29:58 +02:00
|
|
|
}
|
2021-01-18 15:53:55 +01:00
|
|
|
|
2012-01-11 23:28:30 +01:00
|
|
|
// Keep going after reporting the error.
|
2021-01-18 15:53:55 +01:00
|
|
|
VRM->assignVirt2Phys(VirtReg->reg(), AllocOrder.front());
|
2012-01-11 23:28:30 +01:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (AvailablePhysReg)
|
2012-06-21 00:52:29 +02:00
|
|
|
Matrix->assign(*VirtReg, AvailablePhysReg);
|
2012-01-11 23:28:30 +01:00
|
|
|
|
2020-11-03 05:07:58 +01:00
|
|
|
for (Register Reg : SplitVRegs) {
|
2017-07-24 20:07:55 +02:00
|
|
|
assert(LIS->hasInterval(Reg));
|
|
|
|
|
|
|
|
LiveInterval *SplitVirtReg = &LIS->getInterval(Reg);
|
2020-09-15 23:54:38 +02:00
|
|
|
assert(!VRM->hasPhys(SplitVirtReg->reg()) && "Register already assigned");
|
|
|
|
if (MRI->reg_nodbg_empty(SplitVirtReg->reg())) {
|
2017-07-24 20:07:55 +02:00
|
|
|
assert(SplitVirtReg->empty() && "Non-empty but used interval");
|
2018-05-14 14:53:11 +02:00
|
|
|
LLVM_DEBUG(dbgs() << "not queueing unused " << *SplitVirtReg << '\n');
|
[RegAllocGreedy] Introduce a late pass to repair broken hints.
A broken hint is a copy where both ends are assigned different colors. When a
variable gets evicted in the neighborhood of such copies, it is likely we can
reconcile some of them.
** Context **
Copies are inserted during the register allocation via splitting. These split
points are required to relax the constraints on the allocation problem. When
such a point is inserted, both ends of the copy would not share the same color
with respect to the current allocation problem. When variables get evicted,
the allocation problem becomes different and some split point may not be
required anymore. However, the related variables may already have been colored.
This usually shows up in the assembly with pattern like this:
def A
...
save A to B
def A
use A
restore A from B
...
use B
Whereas we could simply have done:
def B
...
def A
use A
...
use B
** Proposed Solution **
A variable having a broken hint is marked for late recoloring if and only if
selecting a register for it evict another variable. Indeed, if no eviction
happens this is pointless to look for recoloring opportunities as it means the
situation was the same as the initial allocation problem where we had to break
the hint.
Finally, when everything has been allocated, we look for recoloring
opportunities for all the identified candidates.
The recoloring is performed very late to rely on accurate copy cost (all
involved variables are allocated).
The recoloring is simple unlike the last change recoloring. It propagates the
color of the broken hint to all its copy-related variables. If the color is
available for them, the recoloring uses it, otherwise it gives up on that hint
even if a more complex coloring would have worked.
The recoloring happens only if it is profitable. The profitability is evaluated
using the expected frequency of the copies of the currently recolored variable
with a) its current color and b) with the target color. If a) is greater or
equal than b), then it is profitable and the recoloring happen.
** Example **
Consider the following example:
BB1:
a =
b =
BB2:
...
= b
= a
Let us assume b gets split:
BB1:
a =
b =
BB2:
c = b
...
d = c
= d
= a
Because of how the allocation work, b, c, and d may be assigned different
colors. Now, if a gets evicted to make room for c, assuming b and d were
assigned to something different than a.
We end up with:
BB1:
a =
st a, SpillSlot
b =
BB2:
c = b
...
d = c
= d
e = ld SpillSlot
= e
This is likely that we can assign the same register for b, c, and d,
getting rid of 2 copies.
** Performances **
Both ARM64 and x86_64 show performance improvements of up to 3% for the
llvm-testsuite + externals with Os and O3. There are a few regressions too that
comes from the (in)accuracy of the block frequency estimate.
<rdar://problem/18312047>
llvm-svn: 225422
2015-01-08 02:16:39 +01:00
|
|
|
aboutToRemoveInterval(*SplitVirtReg);
|
2020-09-15 23:54:38 +02:00
|
|
|
LIS->removeInterval(SplitVirtReg->reg());
|
2012-01-11 23:28:30 +01:00
|
|
|
continue;
|
|
|
|
}
|
2018-05-14 14:53:11 +02:00
|
|
|
LLVM_DEBUG(dbgs() << "queuing new interval: " << *SplitVirtReg << "\n");
|
2020-09-15 23:54:38 +02:00
|
|
|
assert(Register::isVirtualRegister(SplitVirtReg->reg()) &&
|
2012-01-11 23:28:30 +01:00
|
|
|
"expect split value in virtual register");
|
|
|
|
enqueue(SplitVirtReg);
|
|
|
|
++NumNewQueued;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2016-04-13 05:08:27 +02:00
|
|
|
|
|
|
|
void RegAllocBase::postOptimization() {
|
|
|
|
spiller().postOptimization();
|
|
|
|
for (auto DeadInst : DeadRemats) {
|
|
|
|
LIS->RemoveMachineInstrFromMaps(*DeadInst);
|
|
|
|
DeadInst->eraseFromParent();
|
|
|
|
}
|
|
|
|
DeadRemats.clear();
|
|
|
|
}
|
RegAlloc: Allow targets to split register allocation
AMDGPU normally spills SGPRs to VGPRs. Previously, since all register
classes are handled at the same time, this was problematic. We don't
know ahead of time how many registers will be needed to be reserved to
handle the spilling. If no VGPRs were left for spilling, we would have
to try to spill to memory. If the spilled SGPRs were required for exec
mask manipulation, it is highly problematic because the lanes active
at the point of spill are not necessarily the same as at the restore
point.
Avoid this problem by fully allocating SGPRs in a separate regalloc
run from VGPRs. This way we know the exact number of VGPRs needed, and
can reserve them for a second run. This fixes the most serious
issues, but it is still possible using inline asm to make all VGPRs
unavailable. Start erroring in the case where we ever would require
memory for an SGPR spill.
This is implemented by giving each regalloc pass a callback which
reports if a register class should be handled or not. A few passes
need some small changes to deal with leftover virtual registers.
In the AMDGPU implementation, a new pass is introduced to take the
place of PrologEpilogInserter for SGPR spills emitted during the first
run.
One disadvantage of this is currently StackSlotColoring is no longer
used for SGPR spills. It would need to be run again, which will
require more work.
Error if the standard -regalloc option is used. Introduce new separate
-sgpr-regalloc and -vgpr-regalloc flags, so the two runs can be
controlled individually. PBQB is not currently supported, so this also
prevents using the unhandled allocator.
2018-09-27 01:36:28 +02:00
|
|
|
|
|
|
|
void RegAllocBase::enqueue(LiveInterval *LI) {
|
|
|
|
const Register Reg = LI->reg();
|
|
|
|
|
|
|
|
assert(Reg.isVirtual() && "Can only enqueue virtual registers");
|
|
|
|
|
|
|
|
if (VRM->hasPhys(Reg))
|
|
|
|
return;
|
|
|
|
|
|
|
|
const TargetRegisterClass &RC = *MRI->getRegClass(Reg);
|
|
|
|
if (ShouldAllocateClass(*TRI, RC)) {
|
|
|
|
LLVM_DEBUG(dbgs() << "Enqueuing " << printReg(Reg, TRI) << '\n');
|
|
|
|
enqueueImpl(LI);
|
|
|
|
} else {
|
|
|
|
LLVM_DEBUG(dbgs() << "Not enqueueing " << printReg(Reg, TRI)
|
|
|
|
<< " in skipped register class\n");
|
|
|
|
}
|
|
|
|
}
|