This patch adds an optimization in CodeGenPrepare to move an extractelement
right before a store when the target can combine them.
The optimization may promote any scalar operations to vector operations in the
way to make that possible.
** Context **
Some targets use different register files for both vector and scalar operations.
This means that transitioning from one domain to another may incur copy from one
register file to another. These copies are not coalescable and may be expensive.
For example, according to the scheduling model, on cortex-A8 a vector to GPR
move is 20 cycles.
** Motivating Example **
Let us consider an example:
define void @foo(<2 x i32>* %addr1, i32* %dest) {
%in1 = load <2 x i32>* %addr1, align 8
%extract = extractelement <2 x i32> %in1, i32 1
%out = or i32 %extract, 1
store i32 %out, i32* %dest, align 4
ret void
}
As it is, this IR generates the following assembly on armv7:
vldr d16, [r0] @vector load
vmov.32 r0, d16[1] @ cross-register-file copy: 20 cycles
orr r0, r0, #1 @ scalar bitwise or
str r0, [r1] @ scalar store
bx lr
Whereas we could generate much faster code:
vldr d16, [r0] @ vector load
vorr.i32 d16, #0x1 @ vector bitwise or
vst1.32 {d16[1]}, [r1:32] @ vector extract + store
bx lr
Half of the computation made in the vector is useless, but this allows to get
rid of the expensive cross-register-file copy.
** Proposed Solution **
To avoid this cross-register-copy penalty, we promote the scalar operations to
vector operations. The penalty will be removed if we manage to promote the whole
chain of computation in the vector domain.
Currently, we do that only when the chain of computation ends by a store and the
target is able to combine an extract with a store.
Stores are the most likely candidates, because other instructions produce values
that would need to be promoted and so, extracted as some point[1]. Moreover,
this is customary that targets feature stores that perform a vector extract (see
AArch64 and X86 for instance).
The proposed implementation relies on the TargetTransformInfo to decide whether
or not it is beneficial to promote a chain of computation in the vector domain.
Unfortunately, this interface is rather inaccurate for this level of details and
although this optimization may be beneficial for X86 and AArch64, the inaccuracy
will lead to the optimization being too aggressive.
Basically in TargetTransformInfo, everything that is legal has a cost of 1,
whereas, even if a vector type is legal, usually a vector operation is slightly
more expensive than its scalar counterpart. That will lead to too many
promotions that may not be counter balanced by the saving of the
cross-register-file copy. For instance, on AArch64 this penalty is just 4
cycles.
For now, the optimization is just enabled for ARM prior than v8, since those
processors have a larger penalty on cross-register-file copies, and the scope is
limited to basic blocks. Because of these two factors, we limit the effects of
the inaccuracy. Indeed, I did not want to build up a fancy cost model with block
frequency and everything on top of that.
[1] We can imagine targets that can combine an extractelement with other
instructions than just stores. If we want to go into that direction, the current
interfaces must be augmented and, moreover, I think this becomes a global isel
problem.
Differential Revision: http://reviews.llvm.org/D5921
<rdar://problem/14170854>
llvm-svn: 220978
Summary:
The goal is to eventually remove all the code related to getInsertFencesForAtomic
in SelectionDAGBuilder as it is wrong (designed for ARM, not really portable, works
mostly by accident because the backends are overly conservative), and repeats the
same logic that goes in emitLeading/TrailingFence.
In this patch, I make AtomicExpandPass insert the fences as it knows better
where to put them. Because this requires getting the fences and not just
passing an IRBuilder around, I had to change the return type of
emitLeading/TrailingFence.
This code only triggers on ARM for now. Because it is earlier in the pipeline
than SelectionDAGBuilder, it triggers and lowers atomic accesses to atomic so
SelectionDAGBuilder does not add barriers anymore on ARM.
If this patch is accepted I plan to implement emitLeading/TrailingFence for all
backends that setInsertFencesForAtomic(true), which will allow both making them
less conservative and simplifying SelectionDAGBuilder once they are all using
this interface.
This should not cause any functionnal change so the existing tests are used
and not modified.
Test Plan: make check-all, benefits from existing tests of atomics on ARM
Reviewers: jfb, t.p.northover
Subscribers: aemerson, llvm-commits
Differential Revision: http://reviews.llvm.org/D5179
llvm-svn: 218329
Summary:
This patch was originally in D5304 (I could not find a way to reopen that revision).
It was accepted, commited and broke the build bots because the overloading of
the constructor of ArrayRef for braced initializer lists is not supported by all
toolchains. I then reverted it, and propose this fixed version that uses a plain
C array instead in makeDMB (that array is then converted implicitly to an
ArrayRef, but that is not behind an ifdef). Could someone confirm me whether
initialization lists for plain C arrays are supported by every toolchain used
to build llvm ? Otherwise I can just initialize the array in the old way:
args[0] = ...; .. ; args[5] = ...;
Below is the description of the original patch:
```
I had only tested this code for ARMv7 and ARMv8. This patch adds several
fallback paths if the processor does not support dmb ish:
- dmb sy if a cortex-M with support for dmb
- mcr p15, #0, r0, c7, c10, #5 for ARMv6 (special instruction equivalent to a DMB)
These fallback paths were chosen based on the code for fence seq_cst.
Thanks to luqmana for having noticed this bug.
```
Test Plan: Added more cases to atomic-load-store.ll + make check-all
Reviewers: jfb, t.p.northover, luqmana
Subscribers: llvm-commits, aemerson
Differential Revision: http://reviews.llvm.org/D5386
llvm-svn: 218066
It is breaking the build on the buildbots but works fine on my machine, I revert
while trying to understand what happens (it appears to depend on the compiler used
to build, I probably used a C++11 feature that is not perfectly supported by some
of the buildbots).
This reverts commit feb3176c4d006f99af8b40373abd56215a90e7cc.
llvm-svn: 217973
Summary:
I had only tested this code for ARMv7 and ARMv8. This patch adds several
fallback paths if the processor does not support dmb ish:
- dmb sy if a cortex-M with support for dmb
- mcr p15, #0, r0, c7, c10, #5 for ARMv6 (special instruction equivalent to a DMB)
These fallback paths were chosen based on the code for fence seq_cst.
Thanks to luqmana for having noticed this bug.
Test Plan: Added more cases to atomic-load-store.ll + make check-all
Reviewers: jfb, t.p.northover, luqmana
Subscribers: aemerson, llvm-commits
Differential Revision: http://reviews.llvm.org/D5304
llvm-svn: 217965
This required a new hook called hasLoadLinkedStoreConditional to know whether
to expand atomics to LL/SC (ARM, AArch64, in a future patch Power) or to
CmpXchg (X86).
Apart from that, the new code in AtomicExpandPass is mostly moved from
X86AtomicExpandPass. The main result of this patch is to get rid of that
pass, which had lots of code duplicated with AtomicExpandPass.
llvm-svn: 217928
Summary:
Split shouldExpandAtomicInIR() into different versions for Stores/Loads/RMWs/CmpXchgs.
Makes runOnFunction cleaner (no more redundant checking/casting), and will help moving
the X86 backend to this pass.
This requires a way of easily detecting which instructions are atomic.
I followed the pattern of mayReadFromMemory, mayWriteOrReadMemory, etc.. in making
isAtomic() a method of Instruction implemented by a switch on the opcodes.
Test Plan: make check
Reviewers: jfb
Subscribers: mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D5035
llvm-svn: 217080
Fixes two latent bugs:
- There was no fence inserted before expanded seq_cst load (unsound on Power)
- There was only a fence release before seq_cst stores (again unsound, in particular on Power)
It is not even clear if this is correct on ARM swift processors (where release fences are
DMB ishst instead of DMB ish). This behaviour is currently preserved on ARM Swift
as it is not clear whether it is incorrect. I would love to get documentation stating
whether it is correct or not.
These two bugs were not triggered because Power is not (yet) using this pass, and these
behaviours happen to be (mostly?) working on ARM
(although they completely butchered the semantics of the llvm IR).
See:
http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-August/075821.html
for an example of the problems that can be caused by the second of these bugs.
I couldn't see a way of fixing these in a completely target-independent way without
adding lots of unnecessary fences on ARM, hence the target-dependent parts of this
patch.
This patch implements the new target-dependent parts only for ARM (the default
of not doing anything is enough for AArch64), other architectures will use this
infrastructure in later patches.
llvm-svn: 217076
The FPv4-SP floating-point unit is generally referred to as
single-precision only, but it does have double-precision registers and
load, store and GPR<->DPR move instructions which operate on them.
This patch enables the use of these registers, the main advantage of
which is that we now comply with the AAPCS-VFP calling convention.
This partially reverts r209650, which added some AAPCS-VFP support,
but did not handle return values or alignment of double arguments in
registers.
This patch also adds tests for Thumb2 code generation for
floating-point instructions and intrinsics, which previously only
existed for ARM.
llvm-svn: 216172
Add header guards to files that were missing guards. Remove #endif comments
as they don't seem common in LLVM (we can easily add them back if we decide
they're useful)
Changes made by clang-tidy with minor tweaks.
llvm-svn: 215558
Rename to allowsMisalignedMemoryAccess.
On R600, 8 and 16 byte accesses are mostly OK with 4-byte alignment,
and don't need to be split into multiple accesses. Vector loads with
an alignment of the element type are not uncommon in OpenCL code.
llvm-svn: 214055
address of the stack guard was being spilled to the stack.
Previously the address of the stack guard would get spilled to the stack if it
was impossible to keep it in a register. This patch introduces a new target
independent node and pseudo instruction which gets expanded post-RA to a
sequence of instructions that load the stack guard value. Register allocator
can now just remat the value when it can't keep it in a register.
<rdar://problem/12475629>
llvm-svn: 213967
The armv7-windows-itanium environment is nearly identical to the MSVC ABI. It
has a few divergences, mostly revolving around the use of the Itanium ABI for
C++. VLA support is one of the extensions that are amongst the set of the
extensions.
This adds support for proper VLA emission for this environment. This is
somewhat similar to the handling for __chkstk emission on X86 and the large
stack frame emission for ARM. The invocation style for chkstk is still
controlled via the -mcmodel flag to clang.
Make an explicit note that this is an extension.
llvm-svn: 210489
We must validate the value type in TLI::getRegisterByName, because if we
don't and the wrong type was used with the IR intrinsic, then we'll assert
(because we won't be able to find a valid register class with which to
construct the requested copy operation). For PPC64, additionally, the type
information is necessary to decide between the 64-bit register and the 32-bit
subregister.
No functionality change.
llvm-svn: 208508
This patch adds support to ARM for custom lowering of the
llvm.{u|s}add.with.overflow.i32 intrinsics for i32/i64. This is particularly useful
for handling idiomatic saturating math functions as generated by
InstCombineCompare.
Test cases included.
rdar://14853450
llvm-svn: 208435
When using the ARM AAPCS, HFAs (Homogeneous Floating-point Aggregates) must
be passed in a block of consecutive floating-point registers, or on the stack.
This means that unused floating-point registers cannot be back-filled with
part of an HFA, however this can currently happen. This patch, along with the
corresponding clang patch (http://reviews.llvm.org/D3083) prevents this.
llvm-svn: 208413
Handle lowering of global addresses for PIC mode compilation on Windows. Always
use the movw/movt load to load the address as Windows on ARM requires ARMv7+ and
is a pure Thumb environment.
llvm-svn: 208385
This patch implements the infrastructure to use named register constructs in
programs that need access to specific registers (bare metal, kernels, etc).
So far, only the stack pointer is supported as a technology preview, but as it
is, the intrinsic can already support all non-allocatable registers from any
architecture.
llvm-svn: 208104
Still only 32-bit ARM using it at this stage, but the promotion allows
direct testing via opt and is a reasonably self-contained patch on the
way to switching ARM64.
At this point, other targets should be able to make use of it without
too much difficulty if they want. (See ARM64 commit coming soon for an
example).
llvm-svn: 206485
The previous situation where ATOMIC_LOAD_WHATEVER nodes were expanded
at MachineInstr emission time had grown to be extremely large and
involved, to account for the subtly different code needed for the
various flavours (8/16/32/64 bit, cmpxchg/add/minmax).
Moving this transformation into the IR clears up the code
substantially, and makes future optimisations much easier:
1. an atomicrmw followed by using the *new* value can be more
efficient. As an IR pass, simple CSE could handle this
efficiently.
2. Making use of cmpxchg success/failure orderings only has to be done
in one (simpler) place.
3. The common "cmpxchg; did we store?" idiom can be exposed to
optimisation.
I intend to gradually improve this situation within the ARM backend
and make sure there are no hidden issues before moving the code out
into CodeGen to be shared with (at least ARM64/AArch64, though I think
PPC & Mips could benefit too).
llvm-svn: 205525
After some discussion on IRC, emitting a call to the library function seems
like a better default, since it will move from a compiler internal error to
a linker error, that the user can work around until LLVM is fixed.
I'm also adding a note on the responsibility of the user to confirm that
the cache was cleared on platforms where nothing is done.
llvm-svn: 204806
Implementing the LLVM part of the call to __builtin___clear_cache
which translates into an intrinsic @llvm.clear_cache and is lowered
by each target, either to a call to __clear_cache or nothing at all
incase the caches are unified.
Updating LangRef and adding some tests for the implemented architectures.
Other archs will have to implement the method in case this builtin
has to be compiled for it, since the default behaviour is to bail
unimplemented.
A Clang patch is required for the builtin to be lowered into the
llvm intrinsic. This will be done next.
llvm-svn: 204802
Similarly to the vshrn instructions, these are simple zext/sext + trunc
operations. Using normal LLVM IR should allow for better code, and more sharing
with the AArch64 backend.
llvm-svn: 201093
vshrn is just the combination of a right shift and a truncate (and the limits
on the immediate value actually mean the signedness of the shift doesn't
matter). Using that representation allows us to get rid of an ARM-specific
intrinsic, share more code with AArch64 and hopefully get better code out of
the mid-end optimisers.
llvm-svn: 201085
Before this patch we used getIntImmCost from TargetTransformInfo to determine if
a load of a constant should be converted to just a constant, but the threshold
for this was set to an arbitrary value. This value works well for the two
targets (X86 and ARM) that implement this target-hook, but it isn't
target-independent at all.
Now targets have the possibility to decide directly if this optimization should
be performed. The default value is set to false to preserve the current
behavior. The target hook has been moved to TargetLowering, which removed the
last use and need of TargetTransformInfo in SelectionDAG.
llvm-svn: 200271
These are handled almost identically to static mode (and ELF's global address
materialisation), except that a symbol may have "$non_lazy_ptr" appended. This
can be handled by passing appropriate flags along with the instruction instead
of using entirely separate pseudo-instructions.
llvm-svn: 195655
This function-attribute modifies the callee-saved register list and function
epilogue (specifically the return instruction) so that a routine is suitable
for use as an interrupt-handler of the specified type without disrupting
user-mode applications.
rdar://problem/14207019
llvm-svn: 191766
This patch enables calls to __aeabi_idivmod when in EABI mode,
by using the remainder value returned on registers (R1),
enabled by the ARM triple "none-eabi". Note that Darwin and
GNUEABI triples will continue lowering on GNU style, that is,
using the stack for the remainder.
Still need to add SREM/UREM support fix for 64-bit lowering.
llvm-svn: 186390
Currently ARM is the only backend that supports FMA instructions (for at least some subtargets) but does not implement this virtual, so FMAs are never generated except from explicit fma intrinsic calls. Apparently this is due to the fact that it supports both fused (one rounding step) and unfused (two rounding step) multiply + add instructions. This patch clarifies that this the case without changing behavior by implementing the virtual function to simply return false, as the default TargetLoweringBase version does.
It is possible that some cpus perform the fused version faster than the unfused version and vice-versa, so the function implementation should be revisited if hard data is found.
llvm-svn: 185994
Introduction:
In case when stack alignment is 8 and GPRs parameter part size is not N*8:
we add padding to GPRs part, so part's last byte must be recovered at
address K*8-1.
We need to do it, since remained (stack) part of parameter starts from
address K*8, and we need to "attach" "GPRs head" without gaps to it:
Stack:
|---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
[ [padding] [GPRs head] ] [ ------ Tail passed via stack ------ ...
FIX:
Note, once we added padding we need to correct *all* Arg offsets that are going
after padded one. That's why we need this fix: Arg offsets were never corrected
before this patch. See new test-cases included in patch.
We also don't need to insert padding for byval parameters that are stored in GPRs
only. We need pad only last byval parameter and only in case it outsides GPRs
and stack alignment = 8.
Though, stack area, allocated for recovered byval params, must satisfy
"Size mod 8 = 0" restriction.
This patch reduces stack usage for some cases:
We can reduce ArgRegsSaveArea since inner N*4 bytes sized byval params my be
"packed" with alignment 4 in some cases.
llvm-svn: 182237
Now even the small structures could be passed within byval (small enough
to be stored in GPRs).
In regression tests next function prototypes are checked:
PR15293:
%artz = type { i32 }
define void @foo(%artz* byval %s)
define void @foo2(%artz* byval %s, i32 %p, %artz* byval %s2)
foo: "s" stored in R0
foo2: "s" stored in R0, "s2" stored in R2.
Next AAPCS rules are checked:
5.5 Parameters Passing, C.4 and C.5,
"ParamSize" is parameter size in 32bit words:
-- NSAA != 0, NCRN < R4 and NCRN+ParamSize > R4.
Parameter should be sent to the stack; NCRN := R4.
-- NSAA != 0, and NCRN < R4, NCRN+ParamSize < R4.
Parameter stored in GPRs; NCRN += ParamSize.
llvm-svn: 181148
1. VarArgStyleRegisters: functionality that emits "store" instructions for byval regs moved out into separated method "StoreByValRegs". Before this patch VarArgStyleRegisters had confused use-cases. It was used for both variadic functions and for regular functions with byval parameters. In last case it created new stack-frame and registered it as VarArg frame, that is wrong.
This patch replaces VarArgsStyleRegisters usage for byval parameters with StoreByValRegs method.
2. In ARMMachineFunctionInfo, "get/setVarArgsRegSaveSize" was renamed to "get/setArgRegsSaveSize". By the same reason. Sometimes it was used for variadic functions, and sometimes for byval parameters in regular functions. Actually, this property means the size of registers, that keeps arguments, and thats why it was renamed.
3. In ARMISelLowering.cpp, ARMTargetLowering class, in methods computeRegArea and StoreByValRegs, VARegXXXXXX was renamed to ArgRegsXXXXXX still by the same reasons.
llvm-svn: 180774
