Implements the various TTI functions to enable constant hoisting on PPC. The
only significant test-suite change is this:
MultiSource/Benchmarks/VersaBench/bmm/bmm - 20% speedup
(which essentially reverses the slowdown from r206120).
llvm-svn: 206141
We had been using the known-zero values of the operand of the or to construct
the mask for an rlwimi; this is not quite correct, but fine when the mask is
constant. When the mask is constant, then the known zeros of the operand must
be a superset of the zeros in the mask. However, when the mask is not a
constant, then there might be bits in the operand that are not known to be zero
that, at runtime, might be zero in the mask. Therefore, we check that any bits
not known to be zero *are* known to be one in the mask. Otherwise, we can't
fold the mask with the or and shift.
This was revealed as a miscompile of
MultiSource/Benchmarks/BitBench/drop3/drop3 when I started experimenting with
constant hoisting.
llvm-svn: 206136
PPC::isVSLDOIShuffleMask should return -1, not false, when the shuffle
predicate should be false.
Noticed by inspection; no test case (yet).
llvm-svn: 205787
This provides more realistic costs for the insert/extractelement instructions
(which are load/store pairs), accounts for the cheap unaligned Altivec load
sequence, and for unaligned VSX load/stores.
Bad news:
MultiSource/Applications/sgefa/sgefa - 35% slowdown (this will require more investigation)
SingleSource/Benchmarks/McGill/queens - 20% slowdown (we no longer vectorize this, but it was a constant store that was scalarized)
MultiSource/Benchmarks/FreeBench/pcompress2/pcompress2 - 2% slowdown
Good news:
SingleSource/Benchmarks/Shootout/ary3 - 54% speedup
SingleSource/Benchmarks/Shootout-C++/ary - 40% speedup
MultiSource/Benchmarks/Ptrdist/ks/ks - 35% speedup
MultiSource/Benchmarks/FreeBench/neural/neural - 30% speedup
MultiSource/Benchmarks/TSVC/Symbolics-flt/Symbolics-flt - 20% speedup
Unfortunately, estimating the costs of the stack-based scalarization sequences
is hard, and adjusting these costs is like a game of whac-a-mole :( I'll
revisit this again after we have better codegen for vector extloads and
truncstores and unaligned load/stores.
llvm-svn: 205658
gcc inline asm supports specifying "cc" as a clobber of all condition
registers. Add just enough modeling of the full register to make this work.
Fixed PR19326.
llvm-svn: 205630
PPCTTI::getMemoryOpCost will now make use of BasicTTI::getMemoryOpCost to
calculate the base cost of the memory access, and then adjust on top of that.
There is no functionality change from this modification, but it will become
important so that PPCTTI can take advantage of scalarization information for which
BasicTTI::getMemoryOpCost will account in the near future.
llvm-svn: 205476
Just pass a MachineInstr reference rather than an MBB iterator.
Creating a MachineInstr& is the first thing every implementation did
anyway.
llvm-svn: 205453
If we have two unique values for a v2i64 build vector, this will always result
in two vector loads if we expand using shuffles. Only one is necessary.
llvm-svn: 205231
sitofp from v2i32 to v2f64 ends up generating a SIGN_EXTEND_INREG v2i64 node
(and similarly for v2i16 and v2i8). Even though there are no sign-extension (or
algebraic shifts) for v2i64 types, we can handle v2i32 sign extensions by
converting two and from v2i64. The small trick necessary here is to shift the
i32 elements into the right lanes before the i32 -> f64 step. This is because
of the big Endian nature of the system, we need the i32 portion in the high
word of the i64 elements.
For v2i16 and v2i8 we can do the same, but we first use the default Altivec
shift-based expansion from v2i16 or v2i8 to v2i32 (by casting to v4i32) and
then apply the above procedure.
llvm-svn: 205146
v2i64 is a legal type under VSX, however we don't have native vector
comparisons. We can handle eq/ne by casting it to an Altivec type, but
everything else must be expanded.
llvm-svn: 205106
The vector divide and sqrt instructions have high latencies, and the scalar
comparisons are like all of the others. On the P7, permutations take an extra
cycle over purely-simple vector ops.
llvm-svn: 205096
I started trying to fix a small issue, but this code has seen a small fix too
many.
The old code was fairly convoluted. Some of the issues it had:
* It failed to check if a symbol difference was in the some section when
converting a relocation to pcrel.
* It failed to check if the relocation was already pcrel.
* The pcrel value computation was wrong in some cases (relocation-pc.s)
* It was missing quiet a few cases where it should not convert symbol
relocations to section relocations, leaving the backends to patch it up.
* It would not propagate the fact that it had changed a relocation to pcrel,
requiring a quiet nasty work around in ARM.
* It was missing comments.
llvm-svn: 205076
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
v2[fi]64 values need to be explicitly passed in VSX registers. This is because
the code in TRI that finds the minimal register class given a register and a
value type will assert if given an Altivec register and a non-Altivec type.
llvm-svn: 205041
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
First, v2f64 vector extract had not been declared legal (and so the existing
patterns were not being used). Second, the patterns for that, and for
scalar_to_vector, should really be a regclass copy, not a subregister
operation, because the VSX registers directly hold both the vector and scalar data.
llvm-svn: 204971
These operations need to be expanded during legalization so that isel does not
crash. In theory, we might be able to custom lower some of these. That,
however, would need to be follow-up work.
llvm-svn: 204963
This adds back r204781.
Original message:
Aliases are just another name for a position in a file. As such, the
regular symbol resolutions are not applied. For example, given
define void @my_func() {
ret void
}
@my_alias = alias weak void ()* @my_func
@my_alias2 = alias void ()* @my_alias
We produce without this patch:
.weak my_alias
my_alias = my_func
.globl my_alias2
my_alias2 = my_alias
That is, in the resulting ELF file my_alias, my_func and my_alias are
just 3 names pointing to offset 0 of .text. That is *not* the
semantics of IR linking. For example, linking in a
@my_alias = alias void ()* @other_func
would require the strong my_alias to override the weak one and
my_alias2 would end up pointing to other_func.
There is no way to represent that with aliases being just another
name, so the best solution seems to be to just disallow it, converting
a miscompile into an error.
llvm-svn: 204934
These patterns are dead (because v4f32 stores are currently promoted to v4i32
and stored using Altivec instructions), and also are likely not correct
(because they'd store the vector elements in the opposite order from that
assumed by the rest of the Altivec code).
llvm-svn: 204839
These instructions have access to the complete VSX register file. In addition,
they "swap" the order of the elements so that element 0 (the scalar part) comes
first in memory and element 1 follows at a higher address.
llvm-svn: 204838
v2i64 needs to be a legal VSX type because it is the SetCC result type from
v2f64 comparisons. We need to expand all non-arithmetic v2i64 operations.
This fixes the lowering for v2f64 VSELECT.
llvm-svn: 204828
With VSX there is a real vector select instruction, and so we should use it.
Note that VSELECT will still scalarize for v2f64 because the corresponding
SetCC result type (v2i64) is not currently a legal type.
llvm-svn: 204801
This reverts commit r204781.
I will follow up to with msan folks to see what is what they
were trying to do with aliases to weak aliases.
llvm-svn: 204784
Aliases are just another name for a position in a file. As such, the
regular symbol resolutions are not applied. For example, given
define void @my_func() {
ret void
}
@my_alias = alias weak void ()* @my_func
@my_alias2 = alias void ()* @my_alias
We produce without this patch:
.weak my_alias
my_alias = my_func
.globl my_alias2
my_alias2 = my_alias
That is, in the resulting ELF file my_alias, my_func and my_alias are
just 3 names pointing to offset 0 of .text. That is *not* the
semantics of IR linking. For example, linking in a
@my_alias = alias void ()* @other_func
would require the strong my_alias to override the weak one and
my_alias2 would end up pointing to other_func.
There is no way to represent that with aliases being just another
name, so the best solution seems to be to just disallow it, converting
a miscompile into an error.
llvm-svn: 204781
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
TableGen will create a lookup table for the A-type FMA instructions providing
their corresponding M-form opcodes. This will be used by upcoming commits.
llvm-svn: 204746
As a first step towards real little-endian code generation, this patch
changes the PowerPC MC layer to actually generate little-endian object
files. This involves passing the little-endian flag through the various
layers, including down to createELFObjectWriter so we actually get basic
little-endian ELF objects, emitting instructions in little-endian order,
and handling fixups and relocations as appropriate for little-endian.
The bulk of the patch is to update most test cases in test/MC/PowerPC
to verify both big- and little-endian encodings. (The only test cases
*not* updated are those that create actual big-endian ABI code, like
the TLS tests.)
Note that while the object files are now little-endian, the generated
code itself is not yet updated, in particular, it still does not adhere
to the ELFv2 ABI.
llvm-svn: 204634
[PPC64LE] ELFv2 ABI updates for the .opd section
The PPC64 Little Endian (PPC64LE) target supports the ELFv2 ABI, and as
such, does not have a ".opd" section. This is keyed off a _CALL_ELF=2
macro check.
The CALL_ELF check is not clearly documented at this time. The basis
for usage in this patch is from the gcc thread here:
http://gcc.gnu.org/ml/gcc-patches/2013-11/msg01144.html
> Adding comment from Uli:
Looks good to me. I think the old-style JIT doesn't really work
anyway for 64-bit, but at least with this patch LLVM will compile
and link again on a ppc64le host ...
llvm-svn: 204614
I'm under the impression that we used to infer the isCommutable flag from the
instruction-associated pattern. Regardless, we don't seem to do this (at least
by default) any more. I've gone through all of our instruction definitions, and
marked as commutative all of those that should be trivial to commute (by
exchanging the first two operands). There has been special code for the RL*
instructions, and that's not changed.
Before this change, we had the following commutative instructions:
RLDIMI
RLDIMIo
RLWIMI
RLWIMI8
RLWIMI8o
RLWIMIo
XSADDDP
XSMULDP
XVADDDP
XVADDSP
XVMULDP
XVMULSP
After:
ADD4
ADD4o
ADD8
ADD8o
ADDC
ADDC8
ADDC8o
ADDCo
ADDE
ADDE8
ADDE8o
ADDEo
AND
AND8
AND8o
ANDo
CRAND
CREQV
CRNAND
CRNOR
CROR
CRXOR
EQV
EQV8
EQV8o
EQVo
FADD
FADDS
FADDSo
FADDo
FMADD
FMADDS
FMADDSo
FMADDo
FMSUB
FMSUBS
FMSUBSo
FMSUBo
FMUL
FMULS
FMULSo
FMULo
FNMADD
FNMADDS
FNMADDSo
FNMADDo
FNMSUB
FNMSUBS
FNMSUBSo
FNMSUBo
MULHD
MULHDU
MULHDUo
MULHDo
MULHW
MULHWU
MULHWUo
MULHWo
MULLD
MULLDo
MULLW
MULLWo
NAND
NAND8
NAND8o
NANDo
NOR
NOR8
NOR8o
NORo
OR
OR8
OR8o
ORo
RLDIMI
RLDIMIo
RLWIMI
RLWIMI8
RLWIMI8o
RLWIMIo
VADDCUW
VADDFP
VADDSBS
VADDSHS
VADDSWS
VADDUBM
VADDUBS
VADDUHM
VADDUHS
VADDUWM
VADDUWS
VAND
VAVGSB
VAVGSH
VAVGSW
VAVGUB
VAVGUH
VAVGUW
VMADDFP
VMAXFP
VMAXSB
VMAXSH
VMAXSW
VMAXUB
VMAXUH
VMAXUW
VMHADDSHS
VMHRADDSHS
VMINFP
VMINSB
VMINSH
VMINSW
VMINUB
VMINUH
VMINUW
VMLADDUHM
VMULESB
VMULESH
VMULEUB
VMULEUH
VMULOSB
VMULOSH
VMULOUB
VMULOUH
VNMSUBFP
VOR
VXOR
XOR
XOR8
XOR8o
XORo
XSADDDP
XSMADDADP
XSMAXDP
XSMINDP
XSMSUBADP
XSMULDP
XSNMADDADP
XSNMSUBADP
XVADDDP
XVADDSP
XVMADDADP
XVMADDASP
XVMAXDP
XVMAXSP
XVMINDP
XVMINSP
XVMSUBADP
XVMSUBASP
XVMULDP
XVMULSP
XVNMADDADP
XVNMADDASP
XVNMSUBADP
XVNMSUBASP
XXLAND
XXLNOR
XXLOR
XXLXOR
This is a by-inspection change, and I'm not sure how to write a reliable test
case. I would like advice on this, however.
llvm-svn: 204609
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
When VSX is available, these instructions should be used in preference to the
older variants that only have access to the scalar floating-point registers.
llvm-svn: 204559
