Roman Lebedev [Thu, 29 Aug 2019 10:50:09 +0000 (10:50 +0000)]
[X86][CodeGen][NFC] Delay `combineIncDecVector()` from DAGCombine to X86DAGToDAGISel
Summary:
We were previously doing it in DAGCombine.
But we also want to do `sub %x, C` -> `add %x, (sub 0, C)` for vectors in DAGCombine.
So if we had `sub %x, -1`, we'll transform it to `add %x, 1`,
which `combineIncDecVector()` will immediately transform back into `sub %x, -1`,
and here we go again...
I've marked this as NFC since not a single test changes,
but since that 'changes' DAGCombine, probably this isn't fully NFC.
Roman Lebedev [Thu, 29 Aug 2019 10:26:23 +0000 (10:26 +0000)]
[InstCombine] Shift amount reassociation in bittest: trunc-of-lshr (PR42399)
Summary:
Finally, the fold i was looking forward to :)
The legality check is muddy, i doubt i've groked the full generalization,
but it handles all the cases i care about, and can come up with:
https://rise4fun.com/Alive/26j
I.e. we can perform the fold if **any** of the following is true:
* The shift amount is either zero or one less than widest bitwidth
* Either of the values being shifted has at most lowest bit set
* The value that is being shifted by `shl` (which is not truncated) should have no less leading zeros than the total shift amount;
* The value that is being shifted by `lshr` (which **is** truncated) should have no less leading zeros than the widest bit width minus total shift amount minus one
I strongly suspect there is some better generalization, but i'm not aware of it as of right now.
For now i also avoided using actual `computeKnownBits()`, but restricted it to constants.
Martin Storsjo [Thu, 29 Aug 2019 08:59:41 +0000 (08:59 +0000)]
[COFF] Fix error handling in ResourceSectionRef
Previously, the expression (Reader.readFoo()) was expanded twice,
triggering asserts as one of the Error types ends up not checked
(and as it was expanded twice, the method would end up called twice
if it failed first).
Craig Topper [Thu, 29 Aug 2019 05:48:48 +0000 (05:48 +0000)]
[X86] Add a DAG combine to combine INSERTPS and VBROADCAST of a scalar load. Remove corresponding isel patterns.
We had an isel pattern to perform this, but its better to
do it in DAG combine as a simplification. This also fixes the lack
of patterns for AVX512 targets.
Florian Hahn [Thu, 29 Aug 2019 04:26:29 +0000 (04:26 +0000)]
[LoopUnroll] Use Lazy strategy for DTU used for MergeBlockIntoPredecessor.
We do not access the DT in the loop, so we do not have to apply updates
eagerly. We can apply them lazyly and flush them after we are done
merging blocks.
As follow-up work, we might be able to use the DTU above as well,
instead of manually updating the DT.
This brings the example from PR43134 from ~100s to ~4s for a relase +
assertions build on my machine.
Matt Arsenault [Thu, 29 Aug 2019 01:13:47 +0000 (01:13 +0000)]
AMDGPU: Don't use frame virtual registers
SGPR spills aren't really handled after SILowerSGPRSpills. In order to
directly control what happens if the scavenger needs to spill, the
scavenger needs to be used directly. There is an alternative to
spilling in these contexts anyway since the frame register can be
increment and restored.
This does present another possible issue if spilling is needed for the
unused carry out if an add is needed. I think this can be avoided by
using a scalar add (although that clobbers SCC, which happens anyway).
Simon Atanasyan [Wed, 28 Aug 2019 22:32:10 +0000 (22:32 +0000)]
[mips] Fix 64-bit address loading in case of applying 32-bit mask to the result
If result of 64-bit address loading combines with 32-bit mask, LLVM
tries to optimize the code and remove "redundant" loading of upper
32-bits of the address. It leads to incorrect code on MIPS64 targets.
MIPS backend creates the following chain of commands to load 64-bit
address in the `MipsTargetLowering::getAddrNonPICSym64` method:
```
(add (shl (add (shl (add %highest(sym), %higher(sym)),
16),
%hi(sym)),
16),
%lo(%sym))
```
If the mask presents, LLVM decides to optimize the chain of commands. It
really does not make sense to load upper 32-bits because the 0x0fffffff
mask anyway clears them. After removing redundant commands we get this
chain:
```
(add (shl (%hi(sym), 16), %lo(%sym))
```
There is no patterns matched `(MipsHi (i64 symbol))`. Due a bug in `SYM_32`
predicate definition, backend incorrectly selects a pattern for a 32-bit
symbols and uses the `lui` instruction for loading `%hi(sym)`.
As a result we get incorrect set of instructions with unnecessary 16-bit
left shifting:
```
lui at,0x0
R_MIPS_HI16 foo
dsll at,at,0x10
daddiu at,at,0
R_MIPS_LO16 foo
```
This patch resolves two problems:
- Fix `SYM_32/SYM_64` predicates to prevent selection of patterns dedicated
to 32-bit symbols in case of using N64 ABI.
- Add missed patterns for 64-bit symbols for `%hi/%lo`.
Artur Pilipenko [Wed, 28 Aug 2019 21:27:50 +0000 (21:27 +0000)]
Fix for "DICompileUnit not listed in llvm.dbg.cu" verification error after ...
...cloning a function from a different module
Currently when a function with debug info is cloned from a different module, the
cloned function may have hanging DICompileUnits, so that the module with the
cloned function fails debug info verification.
The proposed fix inserts all DICompileUnits reachable from the cloned function
to "llvm.dbg.cu" metadata operands of the cloned function module.
Julian Lettner [Wed, 28 Aug 2019 20:40:55 +0000 (20:40 +0000)]
[ASan] Make insertion of version mismatch guard configurable
By default ASan calls a versioned function
`__asan_version_mismatch_check_vXXX` from the ASan module constructor to
check that the compiler ABI version and runtime ABI version are
compatible. This ensures that we get a predictable linker error instead
of hard-to-debug runtime errors.
Sometimes, however, we want to skip this safety guard. This new command
line option allows us to do just that.
James Y Knight [Wed, 28 Aug 2019 20:35:50 +0000 (20:35 +0000)]
Ignore object files that lack coverage information.
Before this change, if multiple binary files were presented, all of them must have been instrumented or the load would fail with coverage_map_error::no_data_found.
Reuse the logic for INSERT_SUBREG to also import SUBREG_TO_REG patterns.
- Split `inferSuperRegisterClass` into two functions, one which tries to use
an existing TreePatternNode (`inferSuperRegisterClassForNode`), and one that
doesn't. SUBREG_TO_REG doesn't have a node to leverage, which is the cause
for the split.
- Rename GlobalISelEmitterInsertSubreg.td to GlobalISelEmitterSubreg.td and
update it.
- Update impacted tests in the AArch64 and X86 backends.
This is kind of a hit/miss for code size improvements/regressions. E.g. in
add-ext.ll, we now get some identity copies. This isn't really anything the
importer can handle, since it's caused by a later pass introducing the copy for
the sake of correctness.
[Attributor] Regularly clear dependences to remove spurious ones
As dependences between abstract attributes can become stale, e.g., if
one was sufficient to imply another one at some point but it has since
been wakened to the point it is not usable for the formerly implied one.
To weed out spurious dependences, and thereby eliminate unneeded
updates, we introduce an option to determine how often the dependence
cache is cleared and recomputed during the fixpoint iteration.
Note that the initial value was determined such that we see a positive
result on our tests.
Kevin P. Neal [Wed, 28 Aug 2019 16:33:36 +0000 (16:33 +0000)]
[FPEnv] Add fptosi and fptoui constrained intrinsics.
This implements constrained floating point intrinsics for FP to signed and
unsigned integers.
Quoting from D32319:
The purpose of the constrained intrinsics is to force the optimizer to
respect the restrictions that will be necessary to support things like the
STDC FENV_ACCESS ON pragma without interfering with optimizations when
these restrictions are not needed.
Reviewed by: Andrew Kaylor, Craig Topper, Hal Finkel, Cameron McInally, Roman Lebedev, Kit Barton
Approved by: Craig Topper
Differential Revision: http://reviews.llvm.org/D63782
Ryan Taylor [Wed, 28 Aug 2019 15:00:45 +0000 (15:00 +0000)]
[AMDGPU] Adjust number of SGPRs available in Calling Convention
This reduces the number of SGPRs due to some concerns about running
out of SGPRs if you make all the SGPRs that aren't reserved available
for the calling convention.
[Attributor] Restrict liveness and return information to functions
Summary:
Until we have proper call-site information we should not recompute
liveness and return information for each call site. This patch directly
uses the function versions and introduces TODOs at the usage sites.
The required iterations to get to the fixpoint are most of the time
reduced by this change and we always avoid work duplication.
Simon Atanasyan [Wed, 28 Aug 2019 12:35:53 +0000 (12:35 +0000)]
[mips] Use less registers to load address of TargetExternalSymbol
There is no pattern matched `add hi, (MipsLo texternalsym)`. As a result,
loading an address of 32-bit symbol requires two registers and one more
additional instruction:
```
addiu $1, $zero, %lo(foo)
lui $2, %hi(foo)
addu $25, $2, $1
```
This patch adds the missed pattern and enables generation more effective
set of instructions:
```
lui $1, %hi(foo)
addiu $25, $1, %lo(foo)
```
Amaury Sechet [Wed, 28 Aug 2019 12:00:06 +0000 (12:00 +0000)]
[TargetLowering] Add buildLegalVectorShuffle facility to help build legal shuffles
Summary: There are at least 2 ways to express the same shuffle. Various pieces of code explicit check for both option, but other places do not when they would benefit from doing it. This patches refactor the codebase to use buildLegalVectorShuffle in order to make that behavior more consistent.
David Green [Wed, 28 Aug 2019 10:13:23 +0000 (10:13 +0000)]
[MVE] VMOVX patterns
This adds fp16 VMOVX patterns, using the same patterns as rL362482 with some
adjustments for MVE. It allows us to move fp16 registers without going into and
out of gprs.
VMOVX is able to move the top bits from a fp16 in a fp reg into the bottom bits
of another register, zeroing the rest. This can be used for odd MVE register
lanes. The top bits are not read by fp16 instructions, so no move is required
there if we are dealing with even lanes.
Hans Wennborg [Wed, 28 Aug 2019 09:21:56 +0000 (09:21 +0000)]
[LLVM-C] Fix ByVal Attribute crashing
With the introduction of the typed byval attribute change there was no
way that the LLVM-C API could create the correct class Attribute. If a
program that uses the C API creates a ByVal attribute and annotates a
function with that attribute LLVM will crash when it assembles or write
that module containing the function out as bitcode.
This change is a minimal fix to at least allow code to work, this is
because the byval change is on the 9.0 and I don't want to introduce new
LLVM-C API this late in the release cycle.
Ayal Zaks [Wed, 28 Aug 2019 09:02:23 +0000 (09:02 +0000)]
[LV] Fold tail by masking - handle reductions
Allow vectorizing loops that have reductions when tail is folded by masking.
A select is introduced in VPlan, choosing between the last value carried by the
loop-exit/live-out instruction of the reduction, and the penultimate value
carried by the reduction phi, according to the "i < n" mask of fold-tail.
This select replaces the last value as the live-out value of the loop.
Sam Parker [Wed, 28 Aug 2019 08:51:13 +0000 (08:51 +0000)]
[ARM][ParallelDSP] Change search for muls
rL369567 reverted a couple of recent changes made to ARMParallelDSP
because of a miscompilation error: PR43073.
The issue stemmed from an underlying bug that was caused by adding
muls into a reduction before it was proved that they could be executed
in parallel with another mul.
Most of the changes here are from the previously reverted commits.
The additional changes have been made area:
1) The Search function now doesn't insert any muls into the Reduction
object. That now happens once the search has successfully finished.
2) For any muls added into the reduction but that weren't paired, we
accumulate their values as an input into the smlad.
Amara Emerson [Tue, 27 Aug 2019 19:54:27 +0000 (19:54 +0000)]
[GlobalISel] Replace hard coded dynamic alloca handling with G_DYN_STACKALLOC.
This change moves the actual stack pointer manipulation into the legalizer,
available to targets via lower(). The codegen is slightly different because
we're using explicit masks instead of G_PTRMASK, and using G_SUB rather than
adding a negative amount via G_GEP.
Philip Reames [Tue, 27 Aug 2019 19:34:43 +0000 (19:34 +0000)]
[Loads/SROA] Remove blatantly incorrect code and fix a bug revealed in the process
The code we had isSafeToLoadUnconditionally was blatantly wrong. This function takes a "Size" argument which is supposed to describe the span loaded from. Instead, the code use the size of the pointer passed (which may be unrelated!) and only checks that span. For any Size > LoadSize, this can and does lead to miscompiles.
Worse, the generic code just a few lines above correctly handles the cases which *are* valid. So, let's delete said code.
Removing this code revealed two issues:
1) As noted by jdoerfert the removed code incorrectly handled external globals. The test update in SROA is to stop testing incorrect behavior.
2) SROA was confusing bytes and bits, but this wasn't obvious as the Size parameter was being essentially ignored anyway. Fixed.
Matt Arsenault [Tue, 27 Aug 2019 19:05:33 +0000 (19:05 +0000)]
DAG: computeNumSignBits for MUL
Copied directly from the IR version.
Most of the testcases I've added for this are somewhat problematic
because they really end up testing the yet to be implemented version
for MUL_I24/MUL_U24.
Jason Liu [Tue, 27 Aug 2019 18:54:46 +0000 (18:54 +0000)]
[XCOFF][AIX] Generate symbol table entries with llvm-readobj
Summary:
This patch implements main entry and auxiliary entries of symbol table generation for llvm-readobj on AIX.
The source code of aix_xcoff_xlc_test8.o (compile with xlc) is:
-bash-4.2$ cat test8.c
extern int i;
extern int TestforXcoff;
extern int fun(int i);
static int static_i;
char* p="abcd";
int fun1(int j) {
static_i++;
j++;
j=j+*p;
return j;
}
int main() {
i++;
fun(i);
return fun1(i);
}
The opt-viewer tests don't pass after this change, and fixing them isn't
trivial. opt-viewer.py imports optmap, which requires adjusting
pythonpath, which is more work than I'm willing to do to fix forward.
Summary:
This patch introduces, SequenceBBQuery - new heuristic to find likely next callable functions it tries to find the blocks with calls in order of execution sequence of Blocks.
It still uses BlockFrequencyAnalysis to find high frequency blocks. For a handful of hottest blocks (plan to customize), the algorithm traverse and discovered the caller blocks along the way to Entry Basic Block and Exit Basic Block. It uses Block Hint, to stop traversing the already visited blocks in both direction. It implicitly assumes that once the block is visited during discovering entry or exit nodes, revisiting them again does not add much. It also branch probability info (cached result) to traverse only hot edges (planned to customize) from hot blocks. Without BPI, the algorithm mostly return's all the blocks in the CFG with calls.
It also changes the heuristic queries, so they don't maintain states. Hence it is safe to call from multiple threads.
It also implements, new instrumentation to avoid jumping into JIT on every call to the function with the help _orc_speculate.decision.block and _orc_speculate.block.
"Speculator Registration Mechanism is also changed" - kudos to @lhames
Open to review, mostly looking to change implementation of SequeceBBQuery heuristics with good data structure choices.
Andrea Di Biagio [Tue, 27 Aug 2019 18:20:34 +0000 (18:20 +0000)]
[Tblgen][MCA] Add the ability to mark groups as LoadQueue and StoreQueue. NFCI
Before this patch, users were not allowed to optionally mark processor resource
groups as load/store queues. That is because tablegen class MemoryQueue was
originally declared as expecting a ProcResource template argument (instead of a
more generic ProcResourceKind).
That was an oversight, since the original intention from D54957 was to let user
mark any processor resource as either load/store queue. This patch adds the
ability to use processor resource groups in MemoryQueue definitions. This is not
a user visible change.
projects / llvm / log