IR: Remove unnecessary fields from MDTemplateParameter
I noticed this fields were never used in r228607, but I neglected to
propagate that into `MDTemplateParameter` until now. This really should
have been done before commit in r228640; sorry for the churn.
David Majnemer [Tue, 10 Feb 2015 00:57:42 +0000 (00:57 +0000)]
X86: Emit an ABI compliant prologue and epilogue for Win64
Win64 has specific contraints on what valid prologues and epilogues look
like. This constraint is born from the flexibility and descriptiveness
of Win64's unwind opcodes.
Prologues previously emitted by LLVM could not be represented by the
unwind opcodes, preventing operations powered by stack unwinding to
successfully work.
Add specialized debug info metadata nodes that match the `DIDescriptor`
wrappers (used by `DIBuilder`) closely. Assembly and bitcode support to
follow soon (it'll mostly just be obvious), but this sketches in today's
schema. This is the first big commit (well, the only *big* one aside
from the testcase changes that'll come when I move this into place) for
PR22464.
I've marked a bunch of obvious changes as `TODO`s in the source; I plan
to make those changes promptly after this hierarchy is moved underneath
`DIDescriptor`, but for now I'm aiming mostly to match the status quo.
Philip Reames [Tue, 10 Feb 2015 00:04:53 +0000 (00:04 +0000)]
Adjust how we avoid poll insertion inside the poll function (NFC)
I realized that my early fix for this was overly complicated. Rather than scatter checks around in a bunch of places, just exit early when we visit the poll function itself.
Thinking about it a bit, the whole inlining mechanism used with gc.safepoint_poll could probably be cleaned up a bit. Originally, poll insertion was fused with gc relocation rewriting. It might be worth going back to see if we can simplify the chain of events now that these two are seperated. As one thought, maybe it makes sense to rewrite calls inside the helper function before inlining it to the many callers. This would require us to visit the poll function before any other functions though..
Adrian Prantl [Mon, 9 Feb 2015 23:57:22 +0000 (23:57 +0000)]
Debug info: When updating debug info during SROA, do not emit debug info
for any padding introduced by SROA. In particular, do not emit debug info
for an alloca that represents only the padding introduced by a previous
iteration.
Adrian Prantl [Mon, 9 Feb 2015 23:57:15 +0000 (23:57 +0000)]
Debug info: Use DW_OP_bit_piece instead of DW_OP_piece in the
intermediate representation. This
- increases consistency by using the same granularity everywhere
- allows for pieces < 1 byte
- DW_OP_piece didn't actually allow storing an offset.
I just realized that the specialized metadata node patch I'm about to
commit won't compile on old compilers. Bump `hash_combine()`'s support
for non-variadic templates to 18 (I tested this by reversing the logic
in the #ifdef).
Aaron Ballman [Mon, 9 Feb 2015 23:11:39 +0000 (23:11 +0000)]
On Windows, we now use RaiseException to generate the kind of trap we require (one which calls our vectored exception handler), and fall back to using a volatile write to simulate a trap elsewhere.
[Statepoint] Improve two asserts, fix some style (NFC)
Summary:
It's important that our users immediately know what gc.safepoint_poll
is. Also fix the style of the declaration of CreateGCStatepoint, in
preparation for another change that will wrap it.
IR: Take uint64_t in DIBuilder::createExpression()
`DIExpression` deals with `uint64_t`, so it doesn't make sense that
`createExpression()` is created from `int64_t`. Switch to `uint64_t` to
unify them.
I've temporarily left in the `int64_t` version, which forwards to the
`uint64_t` version. I'll delete it once I've updated the callers.
MemDerefPrinter: Require DataLayoutPass for higher accuracy
Without a valid data layout, deferenceable(N) doesn't get parsed or
propagated. Since this is the key item we are testing, add a dependency
on the pass.
Philip Reames [Mon, 9 Feb 2015 21:48:05 +0000 (21:48 +0000)]
Add basic tests for PlaceSafepoints
This is just adding really simple tests which should have been part of the original submission. When doing so, I discovered that I'd mistakenly removed required pieces when preparing the patch for upstream submission. I fixed two such bugs in this submission.
isDereferenceablePointer: look through gc.relocate calls
While a theoretical GC might change dereferenceability on collection,
there is no such known collector and no need to account for the case
with a flag yet.
Ben Langmuir [Mon, 9 Feb 2015 20:34:24 +0000 (20:34 +0000)]
Reduce the LockFileManager timeout, and provide unsafeRemoveLockFile
5 minutes is an eternity, so try to strike a better balance between
waiting long enough for any reasonable module build and not so long that
users kill the process because they think it's hanging.
Also give the client a way to delete the lock file after a timeout.
Sanjoy Das [Mon, 9 Feb 2015 19:39:00 +0000 (19:39 +0000)]
Address post-commit review for rL228587: make it explicit that the
<NW> bit of a SCEVAddRecExpr does not depend on the sign of the step
and the start value of the step.
Sanjoy Das [Mon, 9 Feb 2015 18:34:55 +0000 (18:34 +0000)]
Bugfix: SCEV incorrectly marks certain add recurrences as nsw
When creating a scev for sext({X,+,Y}), scev checks if the expression
is equivalent to {sext X,+,zext Y}. If it can prove that, it also
tags the original {X,+,Y} as <nsw>, which is not correct.
In the test case I run `-scalar-evolution` twice because the bug
manifests only once SCEV has run through and seen the `sext`
expressions (and then does a in-place mutation on {X,+,Y}).
Kit Barton [Mon, 9 Feb 2015 17:03:18 +0000 (17:03 +0000)]
This change implements the following three logical vector operations:
veqv (vector equivalence)
vnand
vorc
I increased the AddedComplexity for these instructions to 500 to ensure they are generated instead of issuing other VSX instructions.
For the attached test case different types are used in the ICmpInst
and SelectInst that represent the min/max expressions. However, if the
ICmpInst type is smaller a comparison with the sign/zero extended
operands would have yielded the same result. This situation might
arise after the instruction combination pass was applied.
Lang Hames [Mon, 9 Feb 2015 07:22:56 +0000 (07:22 +0000)]
[Orc] Tweak lambda capture lists to try to avoid an ICE on gcc-4.7.2. NFC.
Apparently gcc-4.7.2 is touchy about 'this' appearing in a lambda capture list
along with other captures. I've rewritten my captures to try to avoid the issue.
Akira Hatanaka [Mon, 9 Feb 2015 06:38:23 +0000 (06:38 +0000)]
Fix a bug in DemoteRegToStack where a reload instruction was inserted into the
wrong basic block.
This would happen when the result of an invoke was used by a phi instruction
in the invoke's normal destination block. An instruction to reload the invoke's
value would get inserted before the critical edge was split and a new basic
block (which is the correct insertion point for the reload) was created. This
commit fixes the bug by splitting the critical edge before all the reload
instructions are inserted.
Also, hoist up the code which computes the insertion point to the only place
that need that computation.
Tim Northover [Mon, 9 Feb 2015 01:21:00 +0000 (01:21 +0000)]
DeadArgElim: fix mismatch in accounting of array return types.
Some parts of DeadArgElim were only considering the individual fields
of StructTypes separately, but others (where insertvalue &
extractvalue instructions occur) also looked into ArrayTypes.
This one is an actual bug; the mismatch can lead to an argument being
considered used by a return sub-value that isn't being tracked (and
hence is dead by default). It then gets incorrectly eliminated.
Tim Northover [Mon, 9 Feb 2015 01:20:53 +0000 (01:20 +0000)]
DeadArgElim: assess uses of entire return value aggregate.
Previously, a non-extractvalue use of an aggregate return value meant
the entire return was considered live (the algorithm gave up
entirely). This was correct, but conservative. It's better to actually
look at that Use, making the analysis results apply to all sub-values
under consideration.
Lang Hames [Mon, 9 Feb 2015 01:20:51 +0000 (01:20 +0000)]
[Orc] Add a JITSymbol class to the Orc APIs, refactor APIs, update clients.
This patch refactors a key piece of the Orc APIs: It removes the
*::getSymbolAddress and *::lookupSymbolAddressIn methods, which returned target
addresses (uint64_ts), and replaces them with *::findSymbol and *::findSymbolIn
respectively, which return instances of the new JITSymbol type. Unlike the old
methods, calling findSymbol or findSymbolIn does not cause the symbol to be
immediately materialized when found. Instead, the symbol will be materialized
if/when the getAddress method is called on the returned JITSymbol. This allows
us to query for the existence of symbols without actually materializing them. In
the future I expect more information to be attached to the JITSymbol class, for
example whether the returned symbol is a weak or strong definition. This will
allow us to properly handle weak symbols and multiple definitions.
Sanjoy Das [Sun, 8 Feb 2015 22:52:17 +0000 (22:52 +0000)]
Bugfix: ScalarEvolution incorrectly assumes that the start of certain
add recurrences don't overflow.
This change makes the optimization more restrictive. It still assumes
that an overflowing `add nsw` is undefined behavior; and this change
will need revisiting once we have a consistent semantics for poison
values.
Craig Topper [Sun, 8 Feb 2015 22:38:25 +0000 (22:38 +0000)]
[X86] Remove the remaining uses of memop from AVX and AVX2 instruction patterns. AVX and AVX2 can handle unaligned loads being folded so we can just use 'load'
Zachary Turner [Sun, 8 Feb 2015 20:58:09 +0000 (20:58 +0000)]
DebugInfoPDB: Make the symbol base case hold an IPDBSession ref.
Dumping a symbol often requires access to data that isn't inside
the symbol hierarchy, but which is only accessible through the
top-level session. This patch is a pure interface change to give
symbols a reference to the session.
Correctly combine alias.scope metadata by a union instead of intersecting
Summary:
The alias.scope metadata represents sets of things an instruction might
alias with. When generically combining the metadata from two
instructions the result must be the union of the original sets, because
the new instruction might alias with anything any of the original
instructions aliased with.
Gather and Scatter are new introduced intrinsics, comming after recently implemented masked load and store.
This is the first patch for Gather and Scatter intrinsics. It includes only the syntax, parsing and verification.
Gather and Scatter intrinsics allow to perform multiple memory accesses (read/write) in one vector instruction.
The intrinsics are not target specific and will have the following syntax:
Gather:
declare <16 x i32> @llvm.masked.gather.v16i32(<16 x i32*> <vector of ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x i32> <passthru>)
declare <8 x float> @llvm.masked.gather.v8f32(<8 x float*><vector of ptrs>, i32 <alignment>, <8 x i1> <mask>, <8 x float><passthru>)
Scatter:
declare void @llvm.masked.scatter.v8i32(<8 x i32><vector value to be stored> , <8 x i32*><vector of ptrs> , i32 <alignment>, <8 x i1> <mask>)
declare void @llvm.masked.scatter.v16i32(<16 x i32> <vector value to be stored> , <16 x i32*> <vector of ptrs>, i32 <alignment>, <16 x i1><mask> )
Vector of ptrs - a set of source/destination addresses, to load/store the value.
Mask - switches on/off vector lanes to prevent memory access for switched-off lanes
vector of ptrs, value and mask should have the same vector width.
These are code examples where gather / scatter should be used and will allow function vectorization
;void foo1(int * restrict A, int * restrict B, int * restrict C) {
; for (int i=0; i<SIZE; i++) {
; A[i] = B[C[i]];
; }
;}
;void foo3(int * restrict A, int * restrict B) {
; for (int i=0; i<SIZE; i++) {
; A[B[i]] = i+5;
; }
;}
Tests will come in the following patches, with CodeGen and Vectorizer.
Tim Northover [Sun, 8 Feb 2015 00:50:47 +0000 (00:50 +0000)]
ARM & AArch64: teach LowerVSETCC that output type size may differ from input.
While various DAG combines try to guarantee that a vector SETCC
operation will have the same output size as input, there's nothing
intrinsic to either creation or LegalizeTypes that actually guarantees
it, so the function needs to be ready to handle a mismatch.
Fortunately this is easy enough, just extend or truncate the naturally
compared result.
I couldn't reproduce the failure in other backends that I know have
SIMD, so it's probably only an issue for these two due to shared
heritage.
Zachary Turner [Sun, 8 Feb 2015 00:29:29 +0000 (00:29 +0000)]
Some cleanup for libpdb.
This patch implements a few of the optional suggestions from the
initial patch comitting libpdb. In particular, it implements a
virtual function out of line for each of the concrete classes.
A few other minor cleanups exist as well, such as using override
instead of virtual, etc.
projects / llvm / log