The initial values were arbitrary. I want them to be more
conservative. This represents the number of latency cycles hidden by
OOO execution. In practice, I think it should be within a small factor
of the complex floating point operation latency so the scheduler can
make some attempt to hide latency even for smallish blocks.
These are by no means the best values, just a starting point for
tuning heuristics. Some benchmarks such as TSVC run faster with this
lower value for SandyBridge. I haven't run anything on Haswell, but
it's shouldn't be 2x SB.
llvm-svn: 179450
I need to handle this for the test case in my following scheduler
commit.
Work is already under way to redesign the mechanism for node order
propagation because this case by case approach is unmaintainable.
llvm-svn: 179448
immediate displacement. Specifically, add support for generating the proper IR.
We've been able to parse this for some time now. Test case to be added on the
clang side.
Part of rdar://13453209
llvm-svn: 179393
can build up the identifier string. No test case as support for looking up
these type of identifiers hasn't been implemented on the clang side.
Part of rdar://13499009
llvm-svn: 179336
As packed comparisons in AVX/SSE produce all 0s or all 1s in each SIMD lane,
vector select could be simplified to AND/OR or removed if one or both values
being selected is all 0s or all 1s.
llvm-svn: 179267
As these two instructions in AVX extension are privileged instructions for
special purpose, it's only expected to be used in inlined assembly.
llvm-svn: 179266
This patch is revised based on patch from Victor Umansky
<victor.umansky@intel.com>. More cases are handled in X86's bool
simplification, i.e.
- SETCC_CARRY
- value is truncated to i1 with AND
As a by-product, PR5443 is also fixed.
llvm-svn: 179265
Add support for the COFF relocation types IMAGE_REL_I386_DIR32NB and
IMAGE_REL_AMD64_ADDR32NB for 32- and 64-bit respectively. These are
similar to normal 4-byte relocations except that they do not include
the base address of the image.
Image-relative relocations are used for debug information (32-bit) and
SEH unwind tables (64-bit).
A new MCSymbolRef variant called 'VK_COFF_IMGREL32' is introduced to
specify such relocations. For AT&T assembly, this variant can be accessed
using the symbol suffix '@imgrel'.
llvm-svn: 179240
into the operand array of the start of the memory reference descriptor.
Additional code in EncodeInstruction provides an additional adjustment.
This patch places that additional code in a separate function,
called getOperandBias, so that any caller of getMemoryOperandNo
can also call getOperandBias.
llvm-svn: 179211
wasn't always the start of the operand. If there was a symbol reference, then
Start pointed to that token. It's very likely there are other places that need
to be updated.
llvm-svn: 179210
Test cases that regressed due to r179115, plus a few more, were added in
r179182. Original commit message below:
[ms-inline asm] Use parsePrimaryExpr in lieu of parseExpression if we need to
parse an identifier. Otherwise, parseExpression may parse multiple tokens,
which makes it impossible to properly compute an immediate displacement.
An example of such a case is the source operand (i.e., [Symbol + ImmDisp]) in
the below example:
__asm mov eax, [Symbol + ImmDisp]
Part of rdar://13611297
llvm-svn: 179187
parse an identifier. Otherwise, parseExpression may parse multiple tokens,
which makes it impossible to properly compute an immediate displacement.
An example of such a case is the source operand (i.e., [Symbol + ImmDisp]) in
the below example:
__asm mov eax, [Symbol + ImmDisp]
The existing test cases exercise this patch.
rdar://13611297
llvm-svn: 179115
rather than deriving the StringRef from the Start and End SMLocs.
Using the Start and End SMLocs works fine for operands such as [Symbol], but
not for operands such as [Symbol + ImmDisp]. All existing test cases that
reference a variable exercise this patch.
rdar://13602265
llvm-svn: 179109
The costs are overfitted so that I can still use the legalization factor.
For example the following kernel has about half the throughput vectorized than
unvectorized when compiled with SSE2. Before this patch we would vectorize it.
unsigned short A[1024];
double B[1024];
void f() {
int i;
for (i = 0; i < 1024; ++i) {
B[i] = (double) A[i];
}
}
radar://13599001
llvm-svn: 179033
During LTO, the target options on functions within the same Module may
change. This would necessitate resetting some of the back-end. Do this for X86,
because it's a Friday afternoon.
llvm-svn: 178917
memory operands.
Essentially, this layers an infix calculator on top of the parsing state
machine. The scale on the index register is still expected to be an immediate
__asm mov eax, [eax + ebx*4]
and will not work with more complex expressions. For example,
__asm mov eax, [eax + ebx*(2*2)]
The plus and minus binary operators assume the numeric value of a register is
zero so as to not change the displacement. Register operands should never
be an operand for a multiply or divide operation; the scale*indexreg
expression is always replaced with a zero on the operand stack to prevent
such a case.
rdar://13521380
llvm-svn: 178881
SSE2 has efficient support for shifts by a scalar. My previous change of making
shifts expensive did not take this into account marking all shifts as expensive.
This would prevent vectorization from happening where it is actually beneficial.
With this change we differentiate between shifts of constants and other shifts.
radar://13576547
llvm-svn: 178808
On certain architectures we can support efficient vectorized version of
instructions if the operand value is uniform (splat) or a constant scalar.
An example of this is a vector shift on x86.
We can efficiently support
for (i = 0 ; i < ; i += 4)
w[0:3] = v[0:3] << <2, 2, 2, 2>
but not
for (i = 0; i < ; i += 4)
w[0:3] = v[0:3] << x[0:3]
This patch adds a parameter to getArithmeticInstrCost to further qualify operand
values as uniform or uniform constant.
Targets can then choose to return a different cost for instructions with such
operand values.
A follow-up commit will test this feature on x86.
radar://13576547
llvm-svn: 178807
The default logic does not correctly identify costs of casts because they are
marked as custom on x86.
For some cases, where the shift amount is a scalar we would be able to generate
better code. Unfortunately, when this is the case the value (the splat) will get
hoisted out of the loop, thereby making it invisible to ISel.
radar://13130673
radar://13537826
llvm-svn: 178703
qualifiers.
This patch only adds support for parsing these identifiers in the
X86AsmParser. The front-end interface isn't capable of looking up
these identifiers at this point in time. The end result is the
compiler now errors during object file emission, rather than at
parse time. Test case coming shortly.
Part of rdar://13499009 and PR13340
llvm-svn: 178566
Buffered means a later divide may be executed out-of-order while a
prior divide is sitting (buffered) in a reservation station.
You can tell it's not pipelined, because operations that use it
reserve it for more than one cycle:
def : WriteRes<WriteIDiv, [HWPort0, HWDivider]> {
let Latency = 25;
let ResourceCycles = [1, 10];
}
We don't currently distinguish between an unpipeline operation and one
that is split into multiple micro-ops requiring the same unit. Except
that the later may have NumMicroOps > 1 if they also consume
issue/dispatch resources.
llvm-svn: 178519
'@SECREL' is what is used by the Microsoft assembler, but GNU as expects '@SECREL32'.
With the patch, the MC-generated code works fine in combination with a recent GNU as (2.23.51.20120920 here).
Patch by David Nadlinger!
Differential Revision: http://llvm-reviews.chandlerc.com/D429
llvm-svn: 178427
- RDRAND always clears the destination value when a random value is not
available (i.e. CF == 0). This value is truncated or zero-extended as
the false boolean value to be returned. Boolean simplification needs
to skip this 'zext' or 'trunc' node.
llvm-svn: 178312
To enable a load of a call address to be folded with that call, this
load is moved from outside of callseq into callseq. Such a moving
adds a non-glued node (that load) into a glued sequence. This non-glue
load is only removed when DAG selection folds them into a memory form
call instruction. When such instruction selection is disabled, it breaks
DAG schedule.
To prevent that, such moving is disabled when target favors register
indirect call.
Previous workaround disabling CALL32m/CALL64m insn selection is removed.
llvm-svn: 178308
form of call in preference to memory indirect on Atom.
In this case, the patch applies the optimization to the code for reloading
spilled registers.
The patch also includes changes to sibcall.ll and movgs.ll, which were
failing on the Atom buildbot after the first patch was applied.
This patch by Sriram Murali.
llvm-svn: 178193
indirect through a memory address is to load the memory address into
a register and then call indirect through the register.
This patch implements this improvement by modifying SelectionDAG to
force a function address which is a memory reference to be loaded
into a virtual register.
Patch by Sriram Murali.
llvm-svn: 178171
All Intel CPUs since Yonah look a lot alike, at least at the granularity
of the scheduling models. We can add more accurate models for
processors that aren't Sandy Bridge if required. Haswell will probably
need its own.
The Atom processor and anything based on NetBurst is completely
different. So are the non-Intel chips.
llvm-svn: 178080
Now all x86 instructions that have itinerary classes also have SchedRW
lists. This is required before the new scheduling models can be used.
There are still unannotated instructions remaining, but they don't have
itinerary classes either.
llvm-svn: 178051
- It's still considered aligned when the specified alignment is larger
than the natural alignment;
- The new alignment for the high 128-bit vector should be min(16,
alignment) as the pointer is advanced by 16, a power-of-2 offset.
llvm-svn: 177947
All the instructions tagged with IIC_DEFAULT had nothing in common, and
we already have a NoItineraries class to represent untagged
instructions.
llvm-svn: 177937
- After moving logic recognizing vector shift with scalar amount from
DAG combining into DAG lowering, we declare to customize all vector
shifts even vector shift on AVX is legal. As a result, the cost model
needs special tuning to identify these legal cases.
llvm-svn: 177586
an X86Operand, but also performs a Sema lookup and adds the sizing directive
when appropriate. Use this when parsing a bracketed statement. This is
necessary to get the instruction matching correct as well. Test case coming
on clang side.
rdar://13455408
llvm-svn: 177439
def : Pat<(load (i64 (X86Wrapper tglobaltlsaddr :$dst))),
(MOV64rm tglobaltlsaddr :$dst)>;
This pattern is invalid because the MOV64rm instruction expects a
source operand of type "i64mem", which is a subclass of X86MemOperand
and thus actually consists of five MI operands, but the Pat provides
only a single MI operand ("tglobaltlsaddr" matches an SDnode of
type ISD::TargetGlobalTLSAddress and provides a single output).
Thus, if the pattern were ever matched, subsequent uses of the MOV64rm
instruction pattern would access uninitialized memory. In addition,
with the TableGen patch I'm about to check in, this would actually be
reported as a build-time error.
Fortunately, the pattern does in fact never match, for at least two
independent reasons.
First, the code generator actually never generates a pattern of the
form (load (X86Wrapper (tglobaltlsaddr))). For most combinations of
TLS and code models, (tglobaltlsaddr) represents just an offset that
needs to be added to some base register, so it is never directly
dereferenced. The only exception is the initial-exec model, where
(tglobaltlsaddr) refers to the (pc-relative) address of a GOT slot,
which *is* in fact directly dereferenced: but in that case, the
X86WrapperRIP node is used, not X86Wrapper, so the Pat doesn't match.
Second, even if some patterns along those lines *were* ever generated,
we should not need an extra Pat pattern to match it. Instead, the
original MOV64rm instruction pattern ought to match directly, since
it uses an "addr" operand, which is implemented via the SelectAddr
C++ routine; this routine is supposed to accept the full range of
input DAGs that may be implemented by a single mov instruction,
including those cases involving ISD::TargetGlobalTLSAddress (and
actually does so e.g. in the initial-exec case as above).
To avoid build breaks (due to the above-mentioned error) after the
TableGen patch is checked in, I'm removing this Pat here.
llvm-svn: 177426
We hitch a ride with the existing OpndItins class that was used to add
instruction itinerary classes in the many multiclasses in this file.
Use the link provided by the X86FoldableSchedWrite.Folded to find the
right SchedWrite for folded loads.
llvm-svn: 177326
This new-style scheduling information is going to replace the
instruction iteneraries.
This also serves as a test case for Andy's fix in r177317.
llvm-svn: 177323
MinGW is almost completely compatible to MSVC, with the exception of the _tls_array global not being available.
Patch by David Nadlinger!
llvm-svn: 177257
Since almost all X86 instructions can fold loads, use a multiclass to
define register/memory pairs of SchedWrites.
An X86FoldableSchedWrite represents the register version of an
instruction. It holds a reference to the SchedWrite to use when the
instruction folds a load.
This will be used inside multiclasses that define rr and rm instruction
versions together.
llvm-svn: 177210
The new InstrSchedModel is easier to use than the instruction
itineraries. It will be used to model instruction latency and throughput
in modern Intel microarchitectures like Sandy Bridge.
InstrSchedModel should be able to coexist with instruction itinerary
classes, but for cleanliness we should switch the Atom processor model
to the new InstrSchedModel as well.
llvm-svn: 177122
LegalizeDAG.cpp uses the value of the comparison operands when checking
the legality of BR_CC, so DAGCombiner should do the same.
v2:
- Expand more BR_CC value types for NVPTX
v3:
- Expand correct BR_CC value types for Hexagon, Mips, and XCore.
llvm-svn: 176694
That can usually be lowered efficiently and is common in sandybridge code.
It would be nice to do this in DAGCombiner but we can't insert arbitrary
BUILD_VECTORs this late.
Fixes PR15462.
llvm-svn: 176634
- Phi nodes should be replaced/updated after lowering CMOV into branch
because 'mainMBB' updating operand in Phi node is changed.
- Add EFLAGS in livein before lowering the 2nd CMOV. It's necessary as
we will reuse the EFLAGS generated before the 1st lowered CMOV, which
won't clobber EFLAGS. However, we need explicitly specify that.
- '-attr=-cmov' test case are added.
llvm-svn: 176598
- Clear 'mayStore' flag when loading from the atomic variable before the
spin loop
- Clear kill flag from one use to multiple use in registers forming the
address to that atomic variable
- don't use a physical register as live-in register in BB (neither entry
nor landing pad.) by copying it into virtual register
(patch by Cameron Zwarich)
llvm-svn: 176538
one-byte NOPs. If the processor actually executes those NOPs, as it sometimes
does with aligned bundling, this can have a performance impact. From my
micro-benchmarks run on my one machine, a 15-byte NOP followed by twelve
one-byte NOPs is about 20% worse than a 15 followed by a 12. This patch
changes NOP emission to emit as many 15-byte (the maximum) as possible followed
by at most one shorter NOP.
llvm-svn: 176464
* Only apply divide bypass optimization when not optimizing for size.
* Fixed bug caused by constant for 0 value of type Int32,
used dividend type to generate the constant instead.
* For atom x86-64 apply the divide bypass to use 16-bit divides instead of
64-bit divides when operand values are small enough.
* Added lit tests for 64-bit divide bypass.
Patch by Tyler Nowicki!
llvm-svn: 176442
This matters for example in following matrix multiply:
int **mmult(int rows, int cols, int **m1, int **m2, int **m3) {
int i, j, k, val;
for (i=0; i<rows; i++) {
for (j=0; j<cols; j++) {
val = 0;
for (k=0; k<cols; k++) {
val += m1[i][k] * m2[k][j];
}
m3[i][j] = val;
}
}
return(m3);
}
Taken from the test-suite benchmark Shootout.
We estimate the cost of the multiply to be 2 while we generate 9 instructions
for it and end up being quite a bit slower than the scalar version (48% on my
machine).
Also, properly differentiate between avx1 and avx2. On avx-1 we still split the
vector into 2 128bits and handle the subvector muls like above with 9
instructions.
Only on avx-2 will we have a cost of 9 for v4i64.
I changed the test case in test/Transforms/LoopVectorize/X86/avx1.ll to use an
add instead of a mul because with a mul we now no longer vectorize. I did
verify that the mul would be indeed more expensive when vectorized with 3
kernels:
for (i ...)
r += a[i] * 3;
for (i ...)
m1[i] = m1[i] * 3; // This matches the test case in avx1.ll
and a matrix multiply.
In each case the vectorized version was considerably slower.
radar://13304919
llvm-svn: 176403
- ISD::SHL/SRL/SRA must have either both scalar or both vector operands
but TLI.getShiftAmountTy() so far only return scalar type. As a
result, backend logic assuming that breaks.
- Rename the original TLI.getShiftAmountTy() to
TLI.getScalarShiftAmountTy() and re-define TLI.getShiftAmountTy() to
return target-specificed scalar type or the same vector type as the
1st operand.
- Fix most TICG logic assuming TLI.getShiftAmountTy() a simple scalar
type.
llvm-svn: 176364
fewer scalar integer (i32 or i64) arguments. It completely eliminates the need
for SDISel for trivial functions.
Also, add the new llc -fast-isel-abort-args option, which is similar to
-fast-isel-abort option, but for formal argument lowering.
llvm-svn: 176052
to TargetFrameLowering, where it belongs. Incidentally, this allows us
to delete some duplicated (and slightly different!) code in TRI.
There are potentially other layering problems that can be cleaned up
as a result, or in a similar manner.
The refactoring was OK'd by Anton Korobeynikov on llvmdev.
Note: this touches the target interfaces, so out-of-tree targets may
be affected.
llvm-svn: 175788
exists solely to enable it to call itself for i8 with some registers.
The proposed patch simplifies the function somewhat to make the High
bit only meaningful for the i8 mode, which makes sense. No functional
difference (getX86SubSuperRegister is not getting called from anywhere
outside with i64 and High=true).
llvm-svn: 175762
sext <4 x i1> to <4 x i64>
sext <4 x i8> to <4 x i64>
sext <4 x i16> to <4 x i64>
I'm running Combine on SIGN_EXTEND_IN_REG and revert SEXT patterns:
(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) -> (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
The sext_in_reg (v4i32 x) may be lowered to shl+sar operations.
The "sar" does not exist on 64-bit operation, so lowering sext_in_reg (v4i64 x) has no vector solution.
I also added a cost of this operations to the AVX costs table.
llvm-svn: 175619
MS-style inline assembly.
This is a follow-on to r175334. Forcing a FP to be emitted doesn't ensure it
will be used. Therefore, force the base pointer as well. We now treat MS
inline assembly in the same way we treat functions with dynamic stack
realignment and VLAs. This guarantees the BP will be used to reference
parameters and locals.
rdar://13218191
llvm-svn: 175576
If the frame pointer is omitted, and any stack changes occur in the inline
assembly, e.g.: "pusha", then any C local variable or C argument references
will be incorrect.
I pass no judgement on anyone who would do such a thing. ;)
rdar://13218191
llvm-svn: 175334
If two functions require different features (e.g., `-mno-sse' vs. `-msse') then
we want to honor that, especially during LTO. We can do that by resetting the
subtarget's features depending upon the 'target-feature' attribute.
llvm-svn: 175314
blocks. We still don't have consensus if we should try to change clang or
the standard, but llvm should work with compilers that implement the current
standard and mangle those functions.
llvm-svn: 175267
This happens when there is both stack realignment and a dynamic alloca in the
function. If we overwrite %esi (rep;movsl uses fixed registers) we'll lose the
base pointer and the next register spill will write into oblivion.
Fixes PR15249 and unbreaks firefox on i386/freebsd. Mozilla uses dynamic allocas
and freebsd a 4 byte stack alignment.
llvm-svn: 175057
account. Atoms use LEA for updating SP in prologs/epilogs, and the
exact LEA opcode depends on the data model.
Also reapplying the test case which was added and then reverted
(because of Atom failures), this time specifying explicitly the CPU in
addition to the triple. The test case now checks all variations (data
mode, cpu Atom vs. Core).
llvm-svn: 174542
pointer in function prologs/epilogs. The opcodes should depend on the
data model (LP64 vs. ILP32) rather than the architecture bit-ness.
llvm-svn: 174446
This change lets us bootstrap LLVM/Clang under ASan and MSan. It contains
fixes for 2 issues:
- X86JIT reads return address from stack, which MSan does not know is
initialized.
- bugpoint tests run binaries with RLIMIT_AS. This does not work with certain
Sanitizers.
We are no longer including config.h in Compiler.h with this change.
llvm-svn: 174306
1) allows the use of RIP-relative addressing in 32-bit LEA instructions under
x86-64 (ILP32 and LP64)
2) separates the size of address registers in 64-bit LEA instructions from
control by ILP32/LP64.
llvm-svn: 174208
conditions are met:
1. They share the same operand and are in the same BB.
2. Both outputs are used.
3. The target has a native instruction that maps to ISD::FSINCOS node or
the target provides a sincos library call.
Implemented the generic optimization in sdisel and enabled it for
Mac OSX. Also added an additional optimization for x86_64 Mac OSX by
using an alternative entry point __sincos_stret which returns the two
results in xmm0 / xmm1.
rdar://13087969
PR13204
llvm-svn: 173755
This catches many cases where we can emit a more efficient shuffle for a
specific mask or when the mask contains undefs. Once the splat is lowered to
unpacks we can't do that anymore.
There is a possibility of moving the promotion after pshufb matching, but I'm
not sure if pshufb with a mask loaded from memory is faster than 3 shuffles, so
I avoided that for now.
llvm-svn: 173569
(defined by the x32 ABI) mode, in which case its pointers are 32-bits
in size. This knowledge is also added to X86RegisterInfo that now
returns the appropriate registers in getPointerRegClass.
There are many outcomes to this change. In order to keep the patches
separate and manageable, we start by focusing on some simple testable
cases. The patch adds a test with passing a pointer to a function -
focusing on the difference between the two data models for x86-64.
Another test is added for handling of 'sret' arguments (and
functionality is added in X86ISelLowering to make it work).
A note on naming: the "x32 ABI" document refers to the AMD64
architecture (in LLVM it's distinguished by being is64Bits() in the
x86 subtarget) with two variations: the LP64 (default) data model, and
the ILP32 data model. This patch adds predicates to the subtarget
which are consistent with this naming scheme.
llvm-svn: 173503
- Add list of physical registers clobbered in pseudo atomic insts
Physical registers are clobbered when pseudo atomic instructions are
expanded. Add them in clobber list to prevent DAG scheduler to
mis-schedule them after these insns are declared side-effect free.
- Add test case from Michael Kuperstein <michael.m.kuperstein@intel.com>
llvm-svn: 173200
Add the x32 environment kind to the triple, and separate the concept of
pointer size and callee save stack slot size, since they're not equal
on x32.
llvm-svn: 173175
Previously we tried to infer it from the bit width size, with an added
IsIEEE argument for the PPC/IEEE 128-bit case, which had a default
value. This default value allowed bugs to creep in, where it was
inappropriate.
llvm-svn: 173138
The optimization handles esoteric cases but adds a lot of complexity both to the X86 backend and to other backends.
This optimization disables an important canonicalization of chains of SEXT nodes and makes SEXT and ZEXT asymmetrical.
Disabling the canonicalization of consecutive SEXT nodes into a single node disables other DAG optimizations that assume
that there is only one SEXT node. The AVX mask optimizations is one example. Additionally this optimization does not update the cost model.
llvm-svn: 172968
Move the early if-conversion pass into this group.
ILP optimizations usually need to find the right balance between
register pressure and ILP using the MachineTraceMetrics analysis to
identify critical paths and estimate other costs. Such passes should run
together so they can share dominator tree and loop info analyses.
Besides if-conversion, future passes to run here here could include
expression height reduction and ARM's MLxExpansion pass.
llvm-svn: 172687
Moving the X86CostTable to a common place, so that other back-ends
can share the code. Also simplifying it a bit and commoning up
tables with one and two types on operations.
llvm-svn: 172658
PR 14848. The lowered sequence is based on the existing sequence the target-independent
DAG Combiner creates for the scalar case.
Patch by Zvi Rackover.
llvm-svn: 171953
This was an experimental option, but needs to be defined
per-target. e.g. PPC A2 needs to aggressively hide latency.
I converted some in-order scheduling tests to A2. Hal is working on
more test cases.
llvm-svn: 171946
The current Intel Atom microarchitecture has a feature whereby
when a function returns early then it is slightly faster to execute
a sequence of NOP instructions to wait until the return address is ready,
as opposed to simply stalling on the ret instruction until
the return address is ready.
When compiling for X86 Atom only, this patch will run a pass,
called "X86PadShortFunction" which will add NOP instructions where less
than four cycles elapse between function entry and return.
It includes tests.
This patch has been updated to address Nadav's review comments
- Optimize only at >= O1 and don't do optimization if -Os is set
- Stores MachineBasicBlock* instead of BBNum
- Uses DenseMap instead of std::map
- Fixes placement of braces
Patch by Andy Zhang.
llvm-svn: 171879
This is necessary not only for representing empty ranges, but for handling
multibyte characters in the input. (If the end pointer in a range refers to
a multibyte character, should it point to the beginning or the end of the
character in a char array?) Some of the code in the asm parsers was already
assuming this anyway.
llvm-svn: 171765
a TargetMachine to construct (and thus isn't always available), to an
analysis group that supports layered implementations much like
AliasAnalysis does. This is a pretty massive change, with a few parts
that I was unable to easily separate (sorry), so I'll walk through it.
The first step of this conversion was to make TargetTransformInfo an
analysis group, and to sink the nonce implementations in
ScalarTargetTransformInfo and VectorTargetTranformInfo into
a NoTargetTransformInfo pass. This allows other passes to add a hard
requirement on TTI, and assume they will always get at least on
implementation.
The TargetTransformInfo analysis group leverages the delegation chaining
trick that AliasAnalysis uses, where the base class for the analysis
group delegates to the previous analysis *pass*, allowing all but tho
NoFoo analysis passes to only implement the parts of the interfaces they
support. It also introduces a new trick where each pass in the group
retains a pointer to the top-most pass that has been initialized. This
allows passes to implement one API in terms of another API and benefit
when some other pass above them in the stack has more precise results
for the second API.
The second step of this conversion is to create a pass that implements
the TargetTransformInfo analysis using the target-independent
abstractions in the code generator. This replaces the
ScalarTargetTransformImpl and VectorTargetTransformImpl classes in
lib/Target with a single pass in lib/CodeGen called
BasicTargetTransformInfo. This class actually provides most of the TTI
functionality, basing it upon the TargetLowering abstraction and other
information in the target independent code generator.
The third step of the conversion adds support to all TargetMachines to
register custom analysis passes. This allows building those passes with
access to TargetLowering or other target-specific classes, and it also
allows each target to customize the set of analysis passes desired in
the pass manager. The baseline LLVMTargetMachine implements this
interface to add the BasicTTI pass to the pass manager, and all of the
tools that want to support target-aware TTI passes call this routine on
whatever target machine they end up with to add the appropriate passes.
The fourth step of the conversion created target-specific TTI analysis
passes for the X86 and ARM backends. These passes contain the custom
logic that was previously in their extensions of the
ScalarTargetTransformInfo and VectorTargetTransformInfo interfaces.
I separated them into their own file, as now all of the interface bits
are private and they just expose a function to create the pass itself.
Then I extended these target machines to set up a custom set of analysis
passes, first adding BasicTTI as a fallback, and then adding their
customized TTI implementations.
The fourth step required logic that was shared between the target
independent layer and the specific targets to move to a different
interface, as they no longer derive from each other. As a consequence,
a helper functions were added to TargetLowering representing the common
logic needed both in the target implementation and the codegen
implementation of the TTI pass. While technically this is the only
change that could have been committed separately, it would have been
a nightmare to extract.
The final step of the conversion was just to delete all the old
boilerplate. This got rid of the ScalarTargetTransformInfo and
VectorTargetTransformInfo classes, all of the support in all of the
targets for producing instances of them, and all of the support in the
tools for manually constructing a pass based around them.
Now that TTI is a relatively normal analysis group, two things become
straightforward. First, we can sink it into lib/Analysis which is a more
natural layer for it to live. Second, clients of this interface can
depend on it *always* being available which will simplify their code and
behavior. These (and other) simplifications will follow in subsequent
commits, this one is clearly big enough.
Finally, I'm very aware that much of the comments and documentation
needs to be updated. As soon as I had this working, and plausibly well
commented, I wanted to get it committed and in front of the build bots.
I'll be doing a few passes over documentation later if it sticks.
Commits to update DragonEgg and Clang will be made presently.
llvm-svn: 171681
cvtsi2* should parse with an 'l' or 'q' suffix or no suffix at all. No suffix should be treated the same as 'l' suffix. Printing should always print a suffix. Previously we didn't parse or print an 'l' suffix.
cvtt*2si/cvt*2si should parse with an 'l' or 'q' suffix or not suffix at all. No suffix should use the destination register size to choose encoding. Printing should not print a suffix.
Original 'l' suffix issue with cvtsi2* pointed out by Michael Kuperstein.
llvm-svn: 171668
URL: http://llvm.org/viewvc/llvm-project?rev=171524&view=rev
Log:
The current Intel Atom microarchitecture has a feature whereby when a function
returns early then it is slightly faster to execute a sequence of NOP
instructions to wait until the return address is ready,
as opposed to simply stalling on the ret instruction
until the return address is ready.
When compiling for X86 Atom only, this patch will run a pass, called
"X86PadShortFunction" which will add NOP instructions where less than four
cycles elapse between function entry and return.
It includes tests.
Patch by Andy Zhang.
llvm-svn: 171603
returns early then it is slightly faster to execute a sequence of NOP
instructions to wait until the return address is ready,
as opposed to simply stalling on the ret instruction
until the return address is ready.
When compiling for X86 Atom only, this patch will run a pass, called
"X86PadShortFunction" which will add NOP instructions where less than four
cycles elapse between function entry and return.
It includes tests.
Patch by Andy Zhang.
llvm-svn: 171524
1. Add code to estimate register pressure.
2. Add code to select the unroll factor based on register pressure.
3. Add bits to TargetTransformInfo to provide the number of registers.
llvm-svn: 171469
In order to cost subvector insertion and extraction, we need to know
the type of the subvector being extracted.
No functionality change.
llvm-svn: 171453
into their new header subdirectory: include/llvm/IR. This matches the
directory structure of lib, and begins to correct a long standing point
of file layout clutter in LLVM.
There are still more header files to move here, but I wanted to handle
them in separate commits to make tracking what files make sense at each
layer easier.
The only really questionable files here are the target intrinsic
tablegen files. But that's a battle I'd rather not fight today.
I've updated both CMake and Makefile build systems (I think, and my
tests think, but I may have missed something).
I've also re-sorted the includes throughout the project. I'll be
committing updates to Clang, DragonEgg, and Polly momentarily.
llvm-svn: 171366
directly.
This is in preparation for removing the use of the 'Attribute' class as a
collection of attributes. That will shift to the AttributeSet class instead.
llvm-svn: 171253
register. In most cases we actually compare or select YMM-sized registers
and mixing the two types creates horrible code. This commit optimizes
some of the transition sequences.
PR14657.
llvm-svn: 171148
The vector truncs were scalarized during LegalizeVectorOps, later vectorized again by some DAGCombine optimization
and finally, lowered by a dagcombing optimization. Now, they are properly lowered during LegalizeVectorOps.
No new testcase because the original testcases still work.
llvm-svn: 171146
When these instructions are encoded in VEX (on AVX) there is no such requirement. This changes the folding
tables and removes the alignment restrictions from VEX-encoded instructions.
llvm-svn: 171024
pmuludq is slow, but it turns out that all the unpacking and packing of the
scalarized mul is even slower. 10% speedup on loop-vectorized paq8p.
llvm-svn: 170985
The only way to read the eflags is using push and pop. If we don't
adjust the stack then we run over the first frame index. This is
not something that we want to do, so we have to make sure that
our machine function does not copy the flags. If it does then
we have to emit the prolog that adjusts the stack.
rdar://12896831
llvm-svn: 170961
This function is often used to decorate dangling instructions, so a
context reference is required to allocate memory for the operands.
Also add a corresponding MachineInstrBuilder method.
llvm-svn: 170797
MC disassembler clients (LLDB) are interested in querying if an
instruction may affect control flow other than by virtue of being
an explicit branch instruction. For example, instructions which
write directly to the PC on some architectures.
llvm-svn: 170610
Use the version that also takes an MF reference instead.
It would technically be possible to extract an MF reference from the MI
as MI->getParent()->getParent(), but that would not work for MIs that
are not inserted into any basic block.
Given the reasonably small number of places this constructor was used at
all, I preferred the compile time check to a run time assertion.
llvm-svn: 170588
When the least bit of C is greater than V, (x&C) must be greater than V
if it is not zero, so the comparison can be simplified.
Although this was suggested in Target/X86/README.txt, it benefits any
architecture with a directly testable form of AND.
Patch by Kevin Schoedel
llvm-svn: 170576
We match the pattern "x >= y ? x-y : 0" into "subus x, y" and two special cases
if y is a constant. DAGCombiner canonicalizes those so we first have to undo the
canonicalization for those cases. The pattern occurs in gzip when the loop
vectorizer is enabled. Part of PR14613.
llvm-svn: 170273
Not all chips targeted by x86_64 have this feature, but a dramatically
increasing number do. Specifying a chip-specific tuning parameter will
continue to turn the feature on or off as appropriate for that
particular chip, but the generic flag should try to achieve the best
performance on the most widely available hardware. Today, the number of
chips with fast UA access dwarfs those without in the x86-64 space.
Note that this also brings LLVM's code generation for this '-march' flag
more in line with that of modern GCCs. Reviewed by Dan Gohman.
llvm-svn: 170269
mention the inline memcpy / memset expansion code is a mess?
This patch split the ZeroOrLdSrc argument into two: IsMemset and ZeroMemset.
The first indicates whether it is expanding a memset or a memcpy / memmove.
The later is whether the memset is a memset of zero. It's totally possible
(likely even) that targets may want to do different things for memcpy and
memset of zero.
llvm-svn: 169959
Also added more comments to explain why it is generally ok to return true.
- Rename getOptimalMemOpType argument IsZeroVal to ZeroOrLdSrc. It's meant to
be true for loaded source (memcpy) or zero constants (memset). The poor name
choice is probably some kind of legacy issue.
llvm-svn: 169954
This shouldn't affect codegen for -O0 compiles as tail call markers are not
emitted in unoptimized compiles. Testing with the external/internal nightly
test suite reveals no change in compile time performance. Testing with -O1,
-O2 and -O3 with fast-isel enabled did not cause any compile-time or
execution-time failures. All tests were performed on my x86 machine.
I'll monitor our arm testers to ensure no regressions occur there.
In an upcoming clang patch I will be marking the objc_autoreleaseReturnValue
and objc_retainAutoreleaseReturnValue as tail calls unconditionally. While
it's theoretically true that this is just an optimization, it's an
optimization that we very much want to happen even at -O0, or else ARC
applications become substantially harder to debug.
Part of rdar://12553082
llvm-svn: 169796
1. Teach it to use overlapping unaligned load / store to copy / set the trailing
bytes. e.g. On 86, use two pairs of movups / movaps for 17 - 31 byte copies.
2. Use f64 for memcpy / memset on targets where i64 is not legal but f64 is. e.g.
x86 and ARM.
3. When memcpy from a constant string, do *not* replace the load with a constant
if it's not possible to materialize an integer immediate with a single
instruction (required a new target hook: TLI.isIntImmLegal()).
4. Use unaligned load / stores more aggressively if target hooks indicates they
are "fast".
5. Update ARM target hooks to use unaligned load / stores. e.g. vld1.8 / vst1.8.
Also increase the threshold to something reasonable (8 for memset, 4 pairs
for memcpy).
This significantly improves Dhrystone, up to 50% on ARM iOS devices.
rdar://12760078
llvm-svn: 169791
Summary:
Not all chips targeted by x86_64 have this feature, but a dramatically
increasing number do. Specifying a chip-specific tuning parameter will
continue to turn the feature on or off as appropriate for that
particular chip, but the generic flag should try to achieve the best
performance on the most widely available hardware. Today, the number of
chips with fast UA access dwarfs those without in the x86-64 space.
Note that this also brings LLVM's code generation for this '-march' flag
more in line with that of modern GCCs.
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D195
llvm-svn: 169740
Intel chips.
The model number rules were determined by inspecting Intel's
documentation for their newer chip model numbers. My understanding is
that all of the newer Intel chips have fast unaligned memory access, but
if anyone is concerned about a particular chip, just shout.
No tests updated; it's not clear we have dedicated tests for the chips'
various features, but if anyone would like tests (or can point me at
some existing ones), I'm happy to oblige.
llvm-svn: 169730
There are still bugs in this pass, as well as other issues that are
being worked on, but the bugs are crashers that occur pretty easily in
the wild. Test cases have been sent to the original commit's review
thread.
This reverts the commits:
r169671: Fix a logic error.
r169604: Move the popcnt tests to an X86 subdirectory.
r168931: Initial commit adding the pass.
llvm-svn: 169683
understand target implementation of any_extend / extload, just generate
zero_extend in place of any_extend for liveouts when the target knows the
zero_extend will be implicit (e.g. ARM ldrb / ldrh) or folded (e.g. x86 movz).
rdar://12771555
llvm-svn: 169536
This is for the lldb team so most of but not all of the values are
to be printed as hex with this option. Some small values like the
scale in an X86 address were requested to printed in decimal
without the leading 0x.
There may be some tweaks need to places that may still be in
decimal that they want in hex. Specially for arm. I made my best
guess. Any tweaks from here should be simple.
I also did the best I know now with help from the C++ gurus
creating the cleanest formatImm() utility function and containing
the changes. But if someone has a better idea to make something
cleaner I'm all ears and game for changing the implementation.
rdar://8109283
llvm-svn: 169393
textually as NativeClient. Also added a link to the native client project for
readers unfamiliar with it.
A Clang patch will follow shortly.
llvm-svn: 169291
missed in the first pass because the script didn't yet handle include
guards.
Note that the script is now able to handle all of these headers without
manual edits. =]
llvm-svn: 169224
Sooooo many of these had incorrect or strange main module includes.
I have manually inspected all of these, and fixed the main module
include to be the nearest plausible thing I could find. If you own or
care about any of these source files, I encourage you to take some time
and check that these edits were sensible. I can't have broken anything
(I strictly added headers, and reordered them, never removed), but they
may not be the headers you'd really like to identify as containing the
API being implemented.
Many forward declarations and missing includes were added to a header
files to allow them to parse cleanly when included first. The main
module rule does in fact have its merits. =]
llvm-svn: 169131
This revision attempts to recognize following population-count pattern:
while(a) { c++; ... ; a &= a - 1; ... },
where <c> and <a>could be used multiple times in the loop body.
TODO: On X8664 and ARM, __buildin_ctpop() are not expanded to a efficent
instruction sequence, which need to be improved in the following commits.
Reviewed by Nadav, really appreciate!
llvm-svn: 168931
when the destination register is wider than the memory load.
These load instructions load from m32 or m64 and set the upper bits to zero,
while the folded instructions may accept m128.
rdar://12721174
llvm-svn: 168710
This pass was conservative in that it always reserved the FP to enable dynamic
stack realignment, which allowed the RA to use aligned spills for vector
registers. This happens even when spills were not necessary. The RA has
since been improved to use unaligned spills when necessary.
The new behavior is to realign the stack if the frame pointer was already
reserved for some other reason, but don't reserve the frame pointer just
because a function contains vector virtual registers.
Part of rdar://12719844
llvm-svn: 168627
Chad Rosier