Summary:
Backends can use setInsertFencesForAtomic to signal to the middle-end that
montonic is the only memory ordering they can accept for
stores/loads/rmws/cmpxchg. The code lowering those accesses with a stronger
ordering to fences + monotonic accesses is currently living in
SelectionDAGBuilder.cpp. In this patch I propose moving this logic out of it
for several reasons:
- There is lots of redundancy to avoid: extremely similar logic already
exists in AtomicExpand.
- The current code in SelectionDAGBuilder does not use any target-hooks, it
does the same transformation for every backend that requires it
- As a result it is plain *unsound*, as it was apparently designed for ARM.
It happens to mostly work for the other targets because they are extremely
conservative, but Power for example had to switch to AtomicExpand to be
able to use lwsync safely (see r218331).
- Because it produces IR-level fences, it cannot be made sound ! This is noted
in the C++11 standard (section 29.3, page 1140):
```
Fences cannot, in general, be used to restore sequential consistency for atomic
operations with weaker ordering semantics.
```
It can also be seen by the following example (called IRIW in the litterature):
```
atomic<int> x = y = 0;
int r1, r2, r3, r4;
Thread 0:
x.store(1);
Thread 1:
y.store(1);
Thread 2:
r1 = x.load();
r2 = y.load();
Thread 3:
r3 = y.load();
r4 = x.load();
```
r1 = r3 = 1 and r2 = r4 = 0 is impossible as long as the accesses are all seq_cst.
But if they are lowered to monotonic accesses, no amount of fences can prevent it..
This patch does three things (I could cut it into parts, but then some of them
would not be tested/testable, please tell me if you would prefer that):
- it provides a default implementation for emitLeadingFence/emitTrailingFence in
terms of IR-level fences, that mimic the original logic of SelectionDAGBuilder.
As we saw above, this is unsound, but the best that can be done without knowing
the targets well (and there is a comment warning about this risk).
- it then switches Mips/Sparc/XCore to use AtomicExpand, relying on this default
implementation (that exactly replicates the logic of SelectionDAGBuilder, so no
functional change)
- it finally erase this logic from SelectionDAGBuilder as it is dead-code.
Ideally, each target would define its own override for emitLeading/TrailingFence
using target-specific fences, but I do not know the Sparc/Mips/XCore memory model
well enough to do this, and they appear to be dealing fine with the ARM-inspired
default expansion for now (probably because they are overly conservative, as
Power was). If anyone wants to compile fences more agressively on these
platforms, the long comment should make it clear why he should first override
emitLeading/TrailingFence.
Test Plan: make check-all, no functional change
Reviewers: jfb, t.p.northover
Subscribers: aemerson, llvm-commits
Differential Revision: http://reviews.llvm.org/D5474
llvm-svn: 219957
Summary:
Currently, call slot optimization requires that if the destination is an
argument, the argument has the sret attribute. This is to ensure that
the memory access won't trap. In addition to sret, we can also allow the
optimization to happen for arguments that have the new dereferenceable
attribute, which gives the same guarantee.
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5832
llvm-svn: 219950
If a square root call has an FP multiplication argument that can be reassociated,
then we can hoist a repeated factor out of the square root call and into a fabs().
In the simplest case, this:
y = sqrt(x * x);
becomes this:
y = fabs(x);
This patch relies on an earlier optimization in instcombine or reassociate to put the
multiplication tree into a canonical form, so we don't have to search over
every permutation of the multiplication tree.
Because there are no IR-level FastMathFlags for intrinsics (PR21290), we have to
use function-level attributes to do this optimization. This needs to be fixed
for both the intrinsics and in the backend.
Differential Revision: http://reviews.llvm.org/D5787
llvm-svn: 219944
When the constant divisor was larger than 32bits, then the optimized code
generated for the AArch64 backend would emit the wrong code, because the shift
was defined as a shift of a 32bit constant '(1<<Lg2(divisor))' and we would
loose the upper 32bits.
This fixes rdar://problem/18678801.
llvm-svn: 219934
Summary:
In order to support big endian targets for the BuildPairF64 nodes we
just need to swap the low/high pair registers. Additionally, for the
ExtractElementF64 nodes we have to calculate the correct stack offset
with respect to the node's register/operand that we want to extract.
Reviewers: dsanders
Reviewed By: dsanders
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5753
llvm-svn: 219931
Make tail recursion elimination a bit more aggressive. This allows us to get
tail recursion on functions that are just branches to a different function. The
fact that the function takes a byval argument does not restrict it from being
optimised into just a tail call.
llvm-svn: 219899
For pointer-typed function arguments, enhanced alignment can be asserted using
the 'align' attribute. When inlining, if this enhanced alignment information is
not otherwise available, preserve it using @llvm.assume-based alignment
assumptions.
llvm-svn: 219876
In AVX512f we support 64x2 and 32x8 inserts via matching them to 32x4 and 64x4
respectively. These are matched by "Alt" Pat<>'s (Alt stands for alternative
VTs).
Since DQ has native support for these intructions, I peeled off the non-"Alt"
part of the baseclass into vinsert_for_size_no_alt. The DQ instructions are
derived from this multiclass. The "Alt" Pat<>'s are disabled with DQ.
Fixes <rdar://problem/18426089>
llvm-svn: 219874
The SelectDS1Addr1Offset complex pattern always tries to store constant
lds pointers in the offset operand and store a zero value in the addr operand.
Since the addr operand does not accept immediates, the zero value
needs to first be copied to a register.
This newly created zero value will not go through normal instruction
selection, so we need to manually insert a V_MOV_B32_e32 in the complex
pattern.
This bug was hidden by the fact that if there was another zero value
in the DAG that had not been selected yet, then the CSE done by the DAG
would use the unselected node for the addr operand rather than the one
that was just created. This would lead to the zero value being selected
and the DAG automatically inserting a V_MOV_B32_e32 instruction.
llvm-svn: 219848
If x is known to have the range [a, b) in a loop predicated by (icmp
ne x, a), its range can be sharpened to [a + 1, b). Get
ScalarEvolution and hence IndVars to exploit this fact.
This change triggers an optimization to widen-loop-comp.ll, so it had
to be edited to get it to pass.
phabricator: http://reviews.llvm.org/D5639
llvm-svn: 219834
Truncate the operands of a switch instruction to a narrower type if the upper
bits are known to be all ones or zeros.
rdar://problem/17720004
llvm-svn: 219832
This is mostly a copy of the existing FastISel GEP code, but we have to
duplicate it for AArch64, because otherwise we would bail out even for simple
cases. This is because the standard fastEmit functions don't cover MUL at all
and ADD is lowered very inefficientily.
The original commit had a bug in the add emit logic, which has been fixed.
llvm-svn: 219831
When processing assembly like
.long .text
we were creating a new undefined symbol .text. GAS on the other hand would
handle that as a reference to the .text section.
This patch implements that by creating the section symbols earlier so that
they are visible during asm parsing.
The patch also updates llvm-readobj to print the symbol number in the relocation
dump so that the test can differentiate between two sections with the same name.
llvm-svn: 219829
The SLP vectorizer should not vectorize ephemeral values. These are used to
express information to the optimizer, and vectorizing them does not lead to
faster code (because the ephemeral values are dropped prior to code generation,
vectorized or not), and obscures the information the instructions are
attempting to communicate (the logic that interprets the arguments to
@llvm.assume generically does not understand vectorized conditions).
Also, uses by ephemeral values are free (because they, and the necessary
extractelement instructions, will be dropped prior to code generation).
llvm-svn: 219816
Summary:
Currently an error is thrown if bundle alignment mode is set more than once
per module (either via the API or the .bundle_align_mode directive). This
change allows setting it multiple times as long as the alignment doesn't
change.
Also nested bundle_lock groups are currently not allowed. This change allows
them, with the effect that the group stays open until all nests are exited,
and if any of the bundle_lock directives has the align_to_end flag, the
group becomes align_to_end.
These changes make the bundle aligment simpler to use in the compiler, and
also better match the corresponding support in GNU as.
Reviewers: jvoung, eliben
Differential Revision: http://reviews.llvm.org/D5801
llvm-svn: 219811
Summary:
Fixes a FIXME in MachineSinking. Instead of using the simple heuristics in
isPostDominatedBy, use the real MachinePostDominatorTree and MachineLoopInfo.
The old heuristics caused instructions to sink unnecessarily, and might create
register pressure.
This is the second try of the fix. The first one (D4814) caused a performance
regression due to failing to sink instructions out of loops (PR21115). This
patch fixes PR21115 by sinking an instruction from a deeper loop to a shallower
one regardless of whether the target block post-dominates the source.
Thanks Alexey Volkov for reporting PR21115!
Test Plan:
Added a NVPTX codegen test to verify that our change prevents the backend from
over-sinking. It also shows the unnecessary register pressure caused by
over-sinking.
Added an X86 test to verify we can sink instructions out of loops regardless of
the dominance relationship. This test is reduced from Alexey's test in PR21115.
Updated an affected test in X86.
Also ran SPEC CINT2006 and llvm-test-suite for compilation time and runtime
performance. Results are attached separately in the review thread.
Reviewers: Jiangning, resistor, hfinkel
Reviewed By: hfinkel
Subscribers: hfinkel, bruno, volkalexey, llvm-commits, meheff, eliben, jholewinski
Differential Revision: http://reviews.llvm.org/D5633
llvm-svn: 219773
Peephole optimization that generates a single conditional branch
for csinc-branch sequences like in the examples below. This is
possible when the csinc sets or clears a register based on a condition
code and the branch checks that register. Also the condition
code may not be modified between the csinc and the original branch.
Examples:
1. Convert csinc w9, wzr, wzr, <CC>;tbnz w9, #0, 0x44
to b.<invCC>
2. Convert csinc w9, wzr, wzr, <CC>; tbz w9, #0, 0x44
to b.<CC>
rdar://problem/18506500
llvm-svn: 219742
A few minor changes to prevent @llvm.assume from interfering with loop
vectorization. First, treat @llvm.assume like the lifetime intrinsics, which
are scalarized (but don't otherwise interfere with the legality checking).
Second, ignore the cost of ephemeral instructions in the loop (these will go
away anyway during CodeGen).
Alignment assumptions and other uses of @llvm.assume can often end up inside of
loops that should be vectorized (this is not uncommon for assumptions generated
by __attribute__((align_value(n))), for example).
llvm-svn: 219741
Patch to provide shuffle decodes and asm comments for the sse pslldq/psrldq SSE2/AVX2 byte shift instructions.
Differential Revision: http://reviews.llvm.org/D5598
llvm-svn: 219738
Thumb1 has legitimate reasons for preferring 32-bit alignment of types
i1/i8/i16, since the 16-bit encoding of "add rD, sp, #imm" requires #imm to be
a multiple of 4. However, this is a trade-off betweem code size and RAM usage;
the DataLayout string is not the best place to represent it even if desired.
So this patch removes the extra Thumb requirements, hopefully making ARM and
Thumb completely compatible in this respect.
llvm-svn: 219734
There's no hard requirement on LLVM to align local variable to 32-bits, so the
Thumb1 frame handling needs to be able to deal with variables that are only
naturally aligned without falling over.
llvm-svn: 219733
This is mostly a copy of the existing FastISel GEP code, but on AArch64 we bail
out even for simple cases, because the standard fastEmit functions don't cover
MUL and ADD is lowered inefficientily.
llvm-svn: 219726
Before, ARM and Thumb mode code had different preferred alignments, which could
lead to some rather unexpected results. There's justification for reducing it
from the default 64-bits (wasted space), but I don't think there is for going
below 32-bits.
There's no actual ABI change here, just to reassure people.
llvm-svn: 219719
The CFL-AA implementation was missing a visit* routine for va_arg instructions,
causing it to assert when run on a function that had one. For now, handle these
in a conservative way.
Fixes PR20954.
llvm-svn: 219718
Eliminate library calls and intrinsic calls to fabs when the input
is a squared value.
Note that no unsafe-math / fast-math assumptions are needed for
this optimization.
Differential Revision: http://reviews.llvm.org/D5777
llvm-svn: 219717
Sign-/zero-extend folding depended on the load and the integer extend to be
both selected by FastISel. This cannot always be garantueed and SelectionDAG
might interfer. This commit adds additonal checks to load and integer extend
lowering to catch this.
Related to rdar://problem/18495928.
llvm-svn: 219716
We assumed that A must be greater than B because the right hand side of
a remainder operator must be nonzero.
However, it is possible for A to be less than B if Pow2 is a power of
two greater than 1.
Take for example:
i32 %A = 0
i32 %B = 31
i32 Pow2 = 2147483648
((Pow2 << 0) >>u 31) is non-zero but A is less than B.
This fixes PR21274.
llvm-svn: 219713
This effectively reverts revert 219707. After fixing the test to work with
new function name format and renamed intrinsic.
Reviewed-by: Tom Stellard <tom@stellard.net>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 219710
Reapply r216913, a fix for PR20832 by Andrea Di Biagio. The commit was reverted
because of buildbot failures, and credit goes to Ulrich Weigand for isolating
the underlying issue (which can be confirmed by Valgrind, which does helpfully
light up like the fourth of July). Uli explained the problem with the original
patch as:
It seems the problem is calling multiplySignificand with an addend of category
fcZero; that is not expected by this routine. Note that for fcZero, the
significand parts are simply uninitialized, but the code in (or rather, called
from) multiplySignificand will unconditionally access them -- in effect using
uninitialized contents.
This version avoids using a category == fcZero addend within
multiplySignificand, which avoids this problem (the Valgrind output is also now
clean).
Original commit message:
[APFloat] Fixed a bug in method 'fusedMultiplyAdd'.
When folding a fused multiply-add builtin call, make sure that we propagate the
correct result in the case where the addend is zero, and the two other operands
are finite non-zero.
Example:
define double @test() {
%1 = call double @llvm.fma.f64(double 7.0, double 8.0, double 0.0)
ret double %1
}
Before this patch, the instruction simplifier wrongly folded the builtin call
in function @test to constant 'double 7.0'.
With this patch, method 'fusedMultiplyAdd' correctly evaluates the multiply and
propagates the expected result (i.e. 56.0).
Added test fold-builtin-fma.ll with the reproducible from PR20832 plus extra
test cases to verify the behavior of method 'fusedMultiplyAdd' in the presence
of NaN/Inf operands.
This fixes PR20832.
llvm-svn: 219708
v2: Add SI lowering
Add test
v3: Place work dimensions after the kernel arguments.
v4: Calculate offset while lowering arguments
v5: rebase
v6: change prefix to AMDGPU
Reviewed-by: Tom Stellard <tom@stellard.net>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 219705
When LazyValueInfo uses @llvm.assume intrinsics to provide edge-value
constraints, we should check for intrinsics that dominate the edge's branch,
not just any potential context instructions. An assumption that dominates the
edge's branch represents a truth on that edge. This is specifically useful, for
example, if multiple predecessors assume a pointer to be nonnull, allowing us
to simplify a later null comparison.
The test case, and an initial patch, were provided by Philip Reames. Thanks!
llvm-svn: 219688
