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.
llvm-svn: 205075
As explained in r204976, because of how the allocation of VSX registers
interacts with the call-lowering code, we sometimes end up generating self VSX
copies. Specifically, things like this:
%VSL2<def> = COPY %F2, %VSL2<imp-use,kill>
(where %F2 is really a sub-register of %VSL2, and so this copy is a nop)
This adds a small cleanup pass to remove these prior to post-RA scheduling.
llvm-svn: 204980
Because of how the allocation of VSX registers interacts with the call-lowering
code, we sometimes end up generating self VSX copies. Specifically, things like
this:
%VSL2<def> = COPY %F2, %VSL2<imp-use,kill>
(where %F2 is really a sub-register of %VSL2, and so this copy is a nop)
The problem is that ExpandPostRAPseudos always assumes that *some* instruction
has been inserted, and adds implicit defs to it. This is a problem if no copy
was inserted because it can cause subtle problems during post-RA scheduling.
These self copies will have to be removed some other way.
llvm-svn: 204976
The VSX instruction set has two types of FMA instructions: A-type (where the
addend is taken from the output register) and M-type (where one of the product
operands is taken from the output register). This adds a small pass that runs
just after MI scheduling (and, thus, just before register allocation) that
mutates A-type instructions (that are created during isel) into M-type
instructions when:
1. This will eliminate an otherwise-necessary copy of the addend
2. One of the product operands is killed by the instruction
The "right" moment to make this decision is in between scheduling and register
allocation, because only there do we know whether or not one of the product
operands is killed by any particular instruction. Unfortunately, this also
makes the implementation somewhat complicated, because the MIs are not in SSA
form and we need to preserve the LiveIntervals analysis.
As a simple example, if we have:
%vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9
%vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16,
%RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16
...
%vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19,
%RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19
...
We can eliminate the copy by changing from the A-type to the
M-type instruction. This means:
%vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16,
%RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16
is replaced by:
%vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9,
%RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9
and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9
llvm-svn: 204768
Although the first two operands are the ones that can be swapped, the tied
input operand is listed before them, so we need to adjust for that.
I have a test case for this, but it goes along with an upcoming commit (so it
will come soon).
llvm-svn: 204748
I've done some experimentation with this, and it looks like using the
lower-latency (but lower throughput) copy instruction is essentially always the
right thing to do.
My assumption is that, in order to be relatively sure that the higher-latency
copy will increase throughput, we'd want to have it unlikely to be in-flight
with its use. On the P7, the global completion table (GCT) can hold a maximum
of 120 instructions, shared among all active threads (up to 4), giving 30
instructions per thread. So specifically, I'd require at least that many
instructions between the copy and the use before the high-latency variant is
used.
Trying this, however, over the entire test suite resulted in zero cases where
the high-latency form would be preferable. This may be a consequence of the
fact that the scheduler views copies as free, and so they tend to end up close
to their uses. For this experiment I created a function:
unsigned chooseVSXCopy(MachineBasicBlock &MBB,
MachineBasicBlock::iterator I,
unsigned DestReg, unsigned SrcReg,
unsigned StartDist = 1,
unsigned Depth = 3) const;
with an implementation like:
if (!Depth)
return PPC::XXLOR;
const unsigned MaxDist = 30;
unsigned Dist = StartDist;
for (auto J = I, JE = MBB.end(); J != JE && Dist <= MaxDist; ++J) {
if (J->isTransient() && !J->isCopy())
continue;
if (J->isCall() || J->isReturn() || J->readsRegister(DestReg, TRI))
return PPC::XXLOR;
++Dist;
}
// We've exceeded the required distance for the high-latency form, use it.
if (Dist > MaxDist)
return PPC::XVCPSGNDP;
// If this is only an exit block, use the low-latency form.
if (MBB.succ_empty())
return PPC::XXLOR;
// We've reached the end of the block, check the successor blocks (up to some
// depth), and use the high-latency form if that is okay with all successors.
for (auto J = MBB.succ_begin(), JE = MBB.succ_end(); J != JE; ++J) {
if (chooseVSXCopy(**J, (*J)->begin(), DestReg, SrcReg,
Dist, --Depth) == PPC::XXLOR)
return PPC::XXLOR;
}
// All of our successor blocks seem okay with the high-latency variant, so
// we'll use it.
return PPC::XVCPSGNDP;
and then changed the copy opcode selection from:
Opc = PPC::XXLOR;
to:
Opc = chooseVSXCopy(MBB, std::next(I), DestReg, SrcReg);
In conclusion, I'm removing the FIXME from the comment, because I believe that
there is, at least absent other examples, nothing to fix.
llvm-svn: 204591
operator* on the by-operand iterators to return a MachineOperand& rather than
a MachineInstr&. At this point they almost behave like normal iterators!
Again, this requires making some existing loops more verbose, but should pave
the way for the big range-based for-loop cleanups in the future.
llvm-svn: 203865
VSX is an ISA extension supported on the POWER7 and later cores that enhances
floating-point vector and scalar capabilities. Among other things, this adds
<2 x double> support and generally helps to reduce register pressure.
The interesting part of this ISA feature is the register configuration: there
are 64 new 128-bit vector registers, the 32 of which are super-registers of the
existing 32 scalar floating-point registers, and the second 32 of which overlap
with the 32 Altivec vector registers. This makes things like vector insertion
and extraction tricky: this can be free but only if we force a restriction to
the right register subclass when needed. A new "minipass" PPCVSXCopy takes care
of this (although it could do a more-optimal job of it; see the comment about
unnecessary copies below).
Please note that, currently, VSX is not enabled by default when targeting
anything because it is not yet ready for that. The assembler and disassembler
are fully implemented and tested. However:
- CodeGen support causes miscompiles; test-suite runtime failures:
MultiSource/Benchmarks/FreeBench/distray/distray
MultiSource/Benchmarks/McCat/08-main/main
MultiSource/Benchmarks/Olden/voronoi/voronoi
MultiSource/Benchmarks/mafft/pairlocalalign
MultiSource/Benchmarks/tramp3d-v4/tramp3d-v4
SingleSource/Benchmarks/CoyoteBench/almabench
SingleSource/Benchmarks/Misc/matmul_f64_4x4
- The lowering currently falls back to using Altivec instructions far more
than it should. Worse, there are some things that are scalarized through the
stack that shouldn't be.
- A lot of unnecessary copies make it past the optimizers, and this needs to
be fixed.
- Many more regression tests are needed.
Normally, I'd fix these things prior to committing, but there are some
students and other contributors who would like to work this, and so it makes
sense to move this development process upstream where it can be subject to the
regular code-review procedures.
llvm-svn: 203768
This change enables tracking i1 values in the PowerPC backend using the
condition register bits. These bits can be treated on PowerPC as separate
registers; individual bit operations (and, or, xor, etc.) are supported.
Tracking booleans in CR bits has several advantages:
- Reduction in register pressure (because we no longer need GPRs to store
boolean values).
- Logical operations on booleans can be handled more efficiently; we used to
have to move all results from comparisons into GPRs, perform promoted
logical operations in GPRs, and then move the result back into condition
register bits to be used by conditional branches. This can be very
inefficient, because the throughput of these CR <-> GPR moves have high
latency and low throughput (especially when other associated instructions
are accounted for).
- On the POWER7 and similar cores, we can increase total throughput by using
the CR bits. CR bit operations have a dedicated functional unit.
Most of this is more-or-less mechanical: Adjustments were needed in the
calling-convention code, support was added for spilling/restoring individual
condition-register bits, and conditional branch instruction definitions taking
specific CR bits were added (plus patterns and code for generating bit-level
operations).
This is enabled by default when running at -O2 and higher. For -O0 and -O1,
where the ability to debug is more important, this feature is disabled by
default. Individual CR bits do not have assigned DWARF register numbers,
and storing values in CR bits makes them invisible to the debugger.
It is critical, however, that we don't move i1 values that have been promoted
to larger values (such as those passed as function arguments) into bit
registers only to quickly turn around and move the values back into GPRs (such
as happens when values are returned by functions). A pair of target-specific
DAG combines are added to remove the trunc/extends in:
trunc(binary-ops(binary-ops(zext(x), zext(y)), ...)
and:
zext(binary-ops(binary-ops(trunc(x), trunc(y)), ...)
In short, we only want to use CR bits where some of the i1 values come from
comparisons or are used by conditional branches or selects. To put it another
way, if we can do the entire i1 computation in GPRs, then we probably should
(on the POWER7, the GPR-operation throughput is higher, and for all cores, the
CR <-> GPR moves are expensive).
POWER7 test-suite performance results (from 10 runs in each configuration):
SingleSource/Benchmarks/Misc/mandel-2: 35% speedup
MultiSource/Benchmarks/Prolangs-C++/city/city: 21% speedup
MultiSource/Benchmarks/MiBench/automotive-susan: 23% speedup
SingleSource/Benchmarks/CoyoteBench/huffbench: 13% speedup
SingleSource/Benchmarks/Misc-C++/Large/sphereflake: 13% speedup
SingleSource/Benchmarks/Misc-C++/mandel-text: 10% speedup
SingleSource/Benchmarks/Misc-C++-EH/spirit: 10% slowdown
MultiSource/Applications/lemon/lemon: 8% slowdown
llvm-svn: 202451
As part of the cleanup done to enable the disassembler, the PPC instructions
now have a valid Size description field. This can now be used to replace some
custom logic in a few places to compute instruction sizes.
Patch by David Wiberg!
llvm-svn: 200623
GPRC_NOR0 is not a subclass of GPRC (because it also contains the ZERO pseudo
register). As a result, we also need to check for it in the spilling code.
llvm-svn: 200288
subsequent changes are easier to review. About to fix some layering
issues, and wanted to separate out the necessary churn.
Also comment and sink the include of "Windows.h" in three .inc files to
match the usage in Memory.inc.
llvm-svn: 198685
Without this, MachineCSE is powerless to handle redundant operations with truncated source operands.
This required fixing the 2-addr pass to handle tied subregisters. It isn't clear what combinations of subregisters can legally be tied, but the simple case of truncated source operands is now safely handled:
%vreg11<def> = COPY %vreg1:sub_32bit; GR32:%vreg11 GR64:%vreg1
%vreg12<def> = COPY %vreg2:sub_32bit; GR32:%vreg12 GR64:%vreg2
%vreg13<def,tied1> = ADD32rr %vreg11<tied0>, %vreg12<kill>, %EFLAGS<imp-def>
Test case: cse-add-with-overflow.ll.
This exposed an existing bug in
PPCInstrInfo::commuteInstruction. Thanks to Rafael for the test case:
PowerPC/crash.ll.
llvm-svn: 197465
Aside from a few minor latency corrections, the major change here is a new
hazard recognizer which focuses on better dispatch-group formation on the
POWER7. As with the PPC970's hazard recognizer, the most important thing it
does is avoid load-after-store hazards within the same dispatch group. It uses
the POWER7's special dispatch-group-terminating nop instruction (instead of
inserting multiple regular nop instructions). This new hazard recognizer makes
use of the scheduling dependency graph itself, built using AA information, to
robustly detect the possibility of load-after-store hazards.
significant test-suite performance changes (the error bars are 99.5% confidence
intervals based on 5 test-suite runs both with and without the change --
speedups are negative):
speedups:
MultiSource/Benchmarks/FreeBench/pcompress2/pcompress2
-0.55171% +/- 0.333168%
MultiSource/Benchmarks/TSVC/CrossingThresholds-dbl/CrossingThresholds-dbl
-17.5576% +/- 14.598%
MultiSource/Benchmarks/TSVC/Reductions-dbl/Reductions-dbl
-29.5708% +/- 7.09058%
MultiSource/Benchmarks/TSVC/Reductions-flt/Reductions-flt
-34.9471% +/- 11.4391%
SingleSource/Benchmarks/BenchmarkGame/puzzle
-25.1347% +/- 11.0104%
SingleSource/Benchmarks/Misc/flops-8
-17.7297% +/- 9.79061%
SingleSource/Benchmarks/Shootout-C++/ary3
-35.5018% +/- 23.9458%
SingleSource/Regression/C/uint64_to_float
-56.3165% +/- 25.4234%
SingleSource/UnitTests/Vectorizer/gcc-loops
-18.5309% +/- 6.8496%
regressions:
MultiSource/Benchmarks/ASCI_Purple/SMG2000/smg2000
18.351% +/- 12.156%
SingleSource/Benchmarks/Shootout-C++/methcall
27.3086% +/- 14.4733%
llvm-svn: 197099
For one predicate to subsume another, they must both check the same condition
register. Failure to check this prerequisite was causing miscompiles.
Fixes PR18003.
llvm-svn: 197089
PPCScoreboardHazardRecognizer was a subclass of ScoreboardHazardRecognizer
which did only one thing: filtered out nodes in EmitInstruction for which
DAG->getInstrDesc(SU) returned NULL. This used to be the case for PPC pseudo
instructions. As far as I can tell, this is no longer true, and so we can use
ScoreboardHazardRecognizer directly.
llvm-svn: 196171
This patch removes most of the trivial cases of weak vtables by pinning them to
a single object file. The memory leaks in this version have been fixed. Thanks
Alexey for pointing them out.
Differential Revision: http://llvm-reviews.chandlerc.com/D2068
Reviewed by Andy
llvm-svn: 195064
This change is incorrect. If you delete virtual destructor of both a base class
and a subclass, then the following code:
Base *foo = new Child();
delete foo;
will not cause the destructor for members of Child class. As a result, I observe
plently of memory leaks. Notable examples I investigated are:
ObjectBuffer and ObjectBufferStream, AttributeImpl and StringSAttributeImpl.
llvm-svn: 194997
This patch removes most of the trivial cases of weak vtables by pinning them to
a single object file.
Differential Revision: http://llvm-reviews.chandlerc.com/D2068
Reviewed by Andy
llvm-svn: 194865
PPCInstrInfo::insertSelect and PPCInstrInfo::canInsertSelect were computing the
common subclass of the true and false inputs, and then selecting either the
32-bit or the 64-bit isel variant based on the result of calling
PPC::GPRCRegClass.hasSubClassEq(RC) and PPC::G8RCRegClass.hasSubClassEq(RC)
(where RC is the common subclass). Unfortunately, this is not quite right: if
we have something like this:
%vreg8<def> = SELECT_CC_I8 %vreg4<kill>, %vreg7<kill>, %vreg6<kill>, 76;
G8RC_and_G8RC_NOX0:%vreg8 CRRC:%vreg4 G8RC_NOX0:%vreg7,%vreg6
then the common subclass of G8RC_and_G8RC_NOX0 and G8RC_NOX0 is G8RC_NOX0, and
G8RC_NOX0 is not a subclass of G8RC (because it also contains the ZERO8
pseudo-register). As a result, we also need to check the common subclass
against GPRC_NOR0 and G8RC_NOX0 explicitly.
This had not been a problem for clients of insertSelect that called
canInsertSelect first (because it had a compensating mistake), but insertSelect
is also used by the PPC pseudo-instruction expander, and this error was causing
a problem in that context.
This problem was found by csmith.
llvm-svn: 186343
Frame index handling is now target-agnostic, so delete the target hooks
for creation & asm printing of target-specific addressing in DBG_VALUEs
and any related functions.
llvm-svn: 184067
The floating-point record forms on PPC don't set the condition register bits
based on a comparison with zero (like the integer record forms do), but rather
based on the exception status bits.
llvm-svn: 181423
The getSwappedPredicate function can be used in other places (such as in
improvements to the PPCCTRLoops pass). Instead of trapping it as a static
function in PPCInstrInfo, move it into PPCPredicates with other
predicate-related things.
No functionality change intended.
llvm-svn: 179926
When matching a compare with a subtract where the arguments of the compare are
swapped w.r.t. the arguments of the subtract, we need to negate the predicates
(or CR bit indices) of the users. This, however, is not the same as inverting
the predicate (negating LT -> GT, but inverting LT -> GE, for example). The ARM
backend seems to do this correctly, but when I adapted the code for the PPC
backend, I introduced an error in this logic.
Comparison optimization is now enabled again by default.
llvm-svn: 179899
Many PPC instructions have a so-called 'record form' which stores to a specific
condition register the result of comparing the result of the instruction with
zero (always as a signed comparison). For integer operations on PPC64, this is
always a 64-bit comparison.
This implementation is derived from the implementation in the ARM backend;
there are some differences because PPC condition registers are allocatable
virtual registers (although the record forms always use a specific one), and we
look for a matching subtraction instruction after the compare (but before the
first use) in addition to before it.
llvm-svn: 179802
This is prep. work for the implementation of optimizeCompare. Many PPC
instructions have 'record' forms (in almost all cases, this means that the RC
bit is set) that cause the result of the instruction to be compared with zero,
and the result of that comparison saved in a predefined condition register. In
order to add the record forms of the instructions without too much
copy-and-paste, the relevant functions have been refactored into multiclasses
which define both the record and normal forms.
Also, two TableGen-generated mapping functions have been added which allow
querying the instruction code for the record form given the normal form (and
vice versa).
No functionality change intended.
llvm-svn: 179356
Because of how predication in implemented on PPC (only for branches), I think
that this is the right thing to do. No functionality change intended.
llvm-svn: 179252
I've not seen this happen in practice, and probably can't until we start
allowing decrement-counter-based conditional branches to be double predicated,
but just in case, don't allow predication of a diamond in which both sides have
ctr-defining branches. Even though the branching behavior of these can be
predicated, the counter-decrementing behavior cannot be.
llvm-svn: 179199
This adds in-principle support for if-converting the bctr[l] instructions.
These instructions are used for indirect branching. It seems, however, that the
current if converter will never actually predicate these. To do so, it would
need the ability to hoist a few setup insts. out of the conditionally-executed
block. For example, code like this:
void foo(int a, int (*bar)()) { if (a != 0) bar(); }
becomes:
...
beq 0, .LBB0_2
std 2, 40(1)
mr 12, 4
ld 3, 0(4)
ld 11, 16(4)
ld 2, 8(4)
mtctr 3
bctrl
ld 2, 40(1)
.LBB0_2:
...
and it would be safe to do all of this unconditionally with a predicated
beqctrl instruction.
llvm-svn: 179156
This enables us to form predicated branches (which are the same conditional
branches we had before) and also a larger set of predicated returns (including
instructions like bdnzlr which is a conditional return and loop-counter
decrement all in one).
At the moment, if conversion does not capture all possible opportunities. A
simple example is provided in early-ret2.ll, where if conversion forms one
predicated return, and then the PPCEarlyReturn pass picks up the other one. So,
at least for now, we'll keep both mechanisms.
llvm-svn: 179134
Some general cleanup and only scan the end of a BB for branches (once we're
done with the terminators and debug values, then there should not be any other
branches). These address post-commit review suggestions by Bill Schmidt.
No functionality change intended.
llvm-svn: 179112
PowerPC has a conditional branch to the link register (return) instruction: BCLR.
This should be used any time when we'd otherwise have a conditional branch to a
return. This adds a small pass, PPCEarlyReturn, which runs just prior to the
branch selection pass (and, importantly, after block placement) to generate
these conditional returns when possible. It will also eliminate unconditional
branches to returns (these happen rarely; most of the time these have already
been tail duplicated by the time PPCEarlyReturn is invoked). This is a nice
optimization for small functions that do not maintain a stack frame.
llvm-svn: 179026
There are certain PPC instructions into which we can fold a zero immediate
operand. We can detect such cases by looking at the register class required
by the using operand (so long as it is not otherwise constrained).
llvm-svn: 178961
On cores for which we know the misprediction penalty, and we have
the isel instruction, we can profitably perform early if conversion.
This enables us to replace some small branch sequences with selects
and avoid the potential stalls from mispredicting the branches.
Enabling this feature required implementing canInsertSelect and
insertSelect in PPCInstrInfo; isel code in PPCISelLowering was
refactored to use these functions as well.
llvm-svn: 178926
These functions should have the same list of load/store instructions. Now that
all load/store forms have been normalized (to single instructions or pseudos)
they can be resynchronized.
Found by inspection, although hopefully this will improve optimization. I've
also added some comments.
llvm-svn: 178180
As Bill Schmidt pointed out to me, only on Darwin do we need to spill/restore
VRSAVE in the SjLj code. For non-Darwin, don't spill/restore VRSAVE (and I've
added some asserts to make sure that we're not).
As it turns out, we're not currently handling the Darwin case correctly (I've
added a FIXME in the test case). I've tried adding various implied register
definitions/uses to force the spill without success, so I'll need to address
this later.
llvm-svn: 178096
In preparation for using the new register scavenger capability for providing
more than one register simultaneously, specifically note functions that have
spilled VRSAVE (currently, this can happen only in functions that use the
setjmp intrinsic). As with CR spilling, such functions will need to provide two
emergency spill slots to the scavenger.
No functionality change intended.
llvm-svn: 177832
The LR register is unconditionally reserved, and its spilling and restoration
is handled by the prologue/epilogue code. As a result, it is never explicitly
spilled by the register allocator.
No functionality change intended.
llvm-svn: 177823
We currently have a duplicated set of call instruction patterns depending
on the ABI to be followed (Darwin vs. Linux). This is a bit odd; while the
different ABIs will result in different instruction sequences, the actual
instructions themselves ought to be independent of the ABI. And in fact it
turns out that the only nontrivial difference between the two sets of
patterns is that in the PPC64 Linux ABI, the instruction used for indirect
calls is marked to take X11 as extra input register (which is indeed used
only with that ABI to hold an incoming environment pointer for nested
functions). However, this does not need to be hard-coded at the .td
pattern level; instead, the C++ code expanding calls can simply add that
use, just like it adds uses for argument registers anyway.
No change in generated code expected.
llvm-svn: 177735
Thanks to Jakob for isolating the underlying problem from the
test case in r177423. The original commit had introduced
asymmetric copy operations, but these turned out to be a work-around
to the real problem (the use of == instead of hasSubClassEq in PPCCTRLoops).
llvm-svn: 177679
Although there is only one Altivec VRSAVE register, it is a member of
a register class, and we need the ability to spill it. Because this
register is normally callee-preserved and handled by special code this
has never before been necessary. However, this capability will be required by
a forthcoming commit adding SjLj support.
llvm-svn: 177654
Currently the PPC r0 register is unconditionally reserved. There are two reasons
for this:
1. r0 is treated specially (as the constant 0) by certain instructions, and so
cannot be used with those instructions as a regular register.
2. r0 is used as a temporary register in the CR-register spilling process
(where, under some circumstances, we require two GPRs).
This change addresses the first reason by introducing a restricted register
class (without r0) for use by those instructions that treat r0 specially. These
register classes have a new pseudo-register, ZERO, which represents the r0-as-0
use. This has the side benefit of making the existing target code simpler (and
easier to understand), and will make it clear to the register allocator that
uses of r0 as 0 don't conflict will real uses of the r0 register.
Once the CR spilling code is improved, we'll be able to allocate r0.
Adding these extra register classes, for some reason unclear to me, causes
requests to the target to copy 32-bit registers to 64-bit registers. The
resulting code seems correct (and causes no test-suite failures), and the new
test case covers this new kind of asymmetric copy.
As r0 is still reserved, no functionality change intended.
llvm-svn: 177423
