Summary:
Just because INC/DEC is a little slow on some processors doesn't mean we shouldn't prefer it when optimizing for size.
This appears to match gcc behavior.
Reviewers: chandlerc, zvi, RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D37177
llvm-svn: 312866
The lxv/stxv instructions require an offset that is 0 % 16. Previously we were
selecting lxv/stxv for loads and stores to the stack where the offset from the
slot was a multiple of 16, but the stack slot was not 16 or more byte aligned.
When the frame gets lowered these transform to r(1|31) + slot + offset.
If slot is not aligned, slot + offset may not be 0 % 16.
Now we require 16 byte or more alignment for select lxv/stxv to stack slots.
Includes a testcase that shows both sufficiently and insufficiently aligned
stack slots.
llvm-svn: 312843
The various scalar bit operations set SCC,
so one is erased or moved it needs to be recomputed.
Not sure why the existing tests don't fail on this.
llvm-svn: 312819
rL312641 Allowed llvm.memcpy/memset/memmove to be tail calls when parent
function return the intrinsics's first argument. However on arm-none-eabi
platform, llvm.memcpy will be expanded to __aeabi_memcpy which doesn't
have return value. The fix is to check the libcall name after expansion
to match "memcpy/memset/memmove" before allowing those intrinsic to be
tail calls.
llvm-svn: 312799
Summary:
This fixes code-gen for XRay in PPC. The regression wasn't caught by
codegen tests which we add in this change.
What happened was the following:
- For tail exits, we used to unconditionally prepend the returns/exits
with a pseudo-instruction that gets lowered to the instrumentation
sled (and leave the actual return/exit instruction as-is).
- Changes to the XRay instrumentation pass caused the tail exits to
suddenly also emit the tail exit pseudo-instruction, since the check
for whether a return instruction was also a call instruction meant it
was a tail exit instruction.
- None of the tests caught the regression either due to non-existent
tests, or the tests being disabled/removed for continuous breakage.
This change re-introduces some of the basic tests and verifies that
we're back to a state that allows the back-end to generate appropriate
XRay instrumented binaries for PPC in the presence of tail exits.
Reviewers: echristo, timshen
Subscribers: nemanjai, kbarton, llvm-commits
Differential Revision: https://reviews.llvm.org/D37570
llvm-svn: 312772
cover the bitwise operators.
Nothing really exciting here, this just stamps out the rest of the core
operations that can RMW memory and set flags.
Still not implemented here: ADC, SBB. Those will require more
interesting logic to channel the flags *in*, and I'm not currently
planning to try to tackle that. It might be interesting for someone who
wants to improve our code generation for bignum implementations.
Differential Revision: https://reviews.llvm.org/D37141
llvm-svn: 312768
operands and used flags to support matching immediate operands.
This is a bit trickier than register operands, and we still want to fall
back on a register operands even for things that appear to be
"immediates" when they won't actually select into the operation's
immediate operand. This also requires us to handle things like selecting
`sub` vs. `add` to minimize the number of bits needed to represent the
immediate, and picking the shortest immediate encoding. In order to
that, we in turn need to scan to make sure that CF isn't used as it will
get inverted.
The end result seems very nice though, and we're now generating
optimal instruction sequences for these patterns IMO.
A follow-up patch will further expand this to other operations with RMW
memory operands. But handing `add` and `sub` are useful starting points
to flesh out the machinery and make sure interesting and complex cases
can be handled.
Thanks to Craig Topper who provided a few fixes and improvements to this
patch in addition to the review!
Differential Revision: https://reviews.llvm.org/D37139
llvm-svn: 312764
For now CUDA-9 is not included in the list of CUDA versions clang
searches for, so the path to CUDA-9 must be explicitly passed
via --cuda-path=.
On LLVM side NVPTX added sm_70 GPU type which bumps required
PTX version to 6.0, but otherwise is equivalent to sm_62 at the moment.
Differential Revision: https://reviews.llvm.org/D37576
llvm-svn: 312734
These don't add any value as they're just compositions of existing
patterns. However, they can confuse the cost logic in ISel, leading to
duplicated vcvt instructions like in PR33199.
llvm-svn: 312724
This patch expands the support of lowerInterleavedload to {8|16|32}x8i stride 3.
LLVM creates suboptimal shuffle code-gen for AVX2. In overall, this patch is a specific fix for the pattern (Strid=3 VF={8|16|32}) and we plan to include the store (deinterleved side).
The patch goal is to optimize the following sequence:
a0 b0 c0 a1 b1 c1 a2 b2
c2 a3 b3 c3 a4 b4 c4 a5
b5 c5 a6 b6 c6 a7 b7 c7
into
a0 a1 a2 a3 a4 a5 a6 a7
b0 b1 b2 b3 b4 b5 b6 b7
c0 c1 c2 c3 c4 c5 c6 c7
Reviewers
1. zvi
2. igor
3. guyblank
4. dorit
5. Ayal
llvm-svn: 312722
Summary:
For large basic blocks with lots of combinable instructions, the
MachineTraceMetrics computations in MachineCombiner can dominate the compile
time, as computing the trace information is quadratic in the number of
instructions in a BB and it's relevant successors/predecessors.
In most cases, knowing the instruction depth should be enough to make
combination decisions. As we already iterate over all instructions in a basic
block, the instruction depth can be computed incrementally. This reduces the
cost of machine-combine drastically in cases where lots of instructions
are combined. The major drawback is that AFAIK, computing the critical path
length cannot be done incrementally. Therefore we only compute
instruction depths incrementally, for basic blocks with more
instructions than inc_threshold. The -machine-combiner-inc-threshold
option can be used to set the threshold and allows for easier
experimenting and checking if using incremental updates for all basic
blocks has any impact on the performance.
Reviewers: sanjoy, Gerolf, MatzeB, efriedma, fhahn
Reviewed By: fhahn
Subscribers: kiranchandramohan, javed.absar, efriedma, llvm-commits
Differential Revision: https://reviews.llvm.org/D36619
llvm-svn: 312719
Summary:
Add patterns for
fptoui <16 x float> to <16 x i8>
fptoui <16 x float> to <16 x i16>
Reviewers: igorb, delena, craig.topper
Reviewed By: craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D37505
llvm-svn: 312704
Globals that are promoted to an ARM constant pool may alias with another
existing constant pool entry. We need to keep a reference to all globals
that were promoted to each constant pool value so that we can emit a
distinct label for each promoted global. These labels are necessary so
that debug info can refer to the promoted global without an undefined
reference during linking.
Patch by Stephen Crane!
llvm-svn: 312692
Tail merging can convert an undef use into a normal one when creating a
common tail. Doing so can make the register live out from a block which
previously contained the undef use. To keep the liveness up-to-date,
insert IMPLICIT_DEFs in such blocks when necessary.
To enable this patch the computeLiveIns() function which used to
compute live-ins for a block and set them immediately is split into new
functions:
- computeLiveIns() just computes the live-ins in a LivePhysRegs set.
- addLiveIns() applies the live-ins to a block live-in list.
- computeAndAddLiveIns() is a convenience function combining the other
two functions and behaving like computeLiveIns() before this patch.
Based on a patch by Krzysztof Parzyszek <kparzysz@codeaurora.org>
Differential Revision: https://reviews.llvm.org/D37034
llvm-svn: 312668
When if-converting a diamond, two separate blocks will be placed back
to back to form a straight line code. To ensure correctness of the
liveness information, any registers that are live in the second block
should not be killed in the first block, even if they were in the
original code.
Additionally, when the two blocks share common instructions at the
beginning, these instructions will not be duplicated, but only placed
once, before both of the blocks. Since the function "isIdenticalTo"
(as used here) ignores kill flags, the common initial code in one
block may have a kill flag for a register that is live in the other
block.
Because the code that removes kill flags only runs for the non-common
parts of the predicated blocks, a kill flag mismatch in the common
code could still lead to a live register being killed prematurely.
llvm-svn: 312654
function return the intrinsics's first argument.
llvm.memcpy/memset/memmove return void but they will return the first
argument after they are expanded as libcalls. Now if the parent function
has any return value, llvm.memcpy cannot be turned into tail call after
expansion.
The patch is to handle that case in SelectionDAGBuilder so when caller
function return the same value as the first argument of llvm.memcpy,
tail call is allowed.
Differential Revision: https://reviews.llvm.org/D37406
llvm-svn: 312641
Summary:
Mesa still uses a hack where empty inline assembly is used as a kind of
optimization barrier. This exposed a problem where not enough wait states
were inserted, because the hazard recognizer implicitly assumed that each
inline assembly "instruction" has at least one wait state.
Reviewers: arsenm
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D37205
llvm-svn: 312635
performing a zext of a register.
On the PR there is discussion of how to more effectively handle this,
but this patch prevents us from miscompiling code.
Differential Revision: https://reviews.llvm.org/D37504
llvm-svn: 312620
Summary:
Most instructions in AVX work “in-lane”, that is, each source element is applied only to other
elements of the same lane, thus a cross lane permutation is costly and needs more than one instrution.
AVX2 includes instructions to perform any-to-any permutation of words over a 256-bit register
and vectorized table lookup.
This should also Fix PR34369
Differential Revision: https://reviews.llvm.org/D37388
llvm-svn: 312608
When packet size equals packet align and is power of 2, transform
__read_pipe* and __write_pipe* to specialized library function.
Differential Revision: https://reviews.llvm.org/D36831
llvm-svn: 312598
Missing these could potentially screw up post-ra scheduling.
Issue found by inspection, so I don't have a real testcase. Included
test just verifies the expected operands after expansion.
Differential Revision: https://reviews.llvm.org/D35156
llvm-svn: 312589
Summary:
This intrinsic represents a label with a list of associated metadata
strings. It is modelled as reading and writing inaccessible memory so
that it won't be removed as dead code. I think the intention is that the
annotation strings should appear at most once in the debug info, so I
marked it noduplicate. We are allowed to inline code with annotations as
long as we strip the annotation, but that can be done later.
Reviewers: majnemer
Subscribers: eraman, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D36904
llvm-svn: 312569
We had already disabled the pattern for SSE4.1 and SSE4.2. But it got re-enabled for AVX and AVX512.
With SSE41 we rely on a separate (v4f32 (X86vzmovl VR128)) pattern to select blendps with a xorps to create zeroess. And a separate (v4f32 (scalar_to_vector FR32X)) to select a COPY_TO_REG_CLASS to move FR32 to VR128
The same thing can happen for AVX with vblendps and those separate patterns already exist.
For AVX512, (v4f32 (X86vzmov VR128)) will select a VMOVSS instruction instead of VBLENDPS due to their not being a EVEX VBLENDPS. This is what we were getting out of the larger pattern anyway. So the larger pattern is unneeded for AVX512 too.
For SSE1-SSSE3 we can rely on (v4f32 (X86vzmov VR128)) selecting a MOVSS similar to AVX512. Again this is what the larger pattern did too.
So the only real change here is that AVX1/2 now properly outputs a VBLENDPS during isel instead of a VMOVSS to match SSE41. Most tests didn't notice because the two address instruction pass knows how to turn VMOVSS into VBLENDPS to get an independent destination register.
llvm-svn: 312564
