OSDN Git Service

android-x86/external-llvm.git
6 years ago[SystemZ] Bugfix for handling of subregisters in getRegAllocationHints().
Jonas Paulsson [Mon, 20 Nov 2017 14:54:03 +0000 (14:54 +0000)]
[SystemZ]  Bugfix for handling of subregisters in getRegAllocationHints().

The 32 bit subreg indices of GR128 registers must also be checked for in
getRC32().

Review: Ulrich Weigand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318652 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[PPC] Heuristic to choose between a X-Form VSX ld/st vs a X-Form FP ld/st.
Tony Jiang [Mon, 20 Nov 2017 14:38:30 +0000 (14:38 +0000)]
[PPC] Heuristic to choose between a X-Form VSX ld/st vs a X-Form FP ld/st.

The VSX versions have the advantage of a full 64-register target whereas the FP
ones have the advantage of lower latency and higher throughput. So what we’re
after is using the faster instructions in low register pressure situations and
using the larger register file in high register pressure situations.

The heuristic chooses between the following 7 pairs of instructions.
PPC::LXSSPX vs PPC::LFSX
PPC::LXSDX vs PPC::LFDX
PPC::STXSSPX vs PPC::STFSX
PPC::STXSDX vs PPC::STFDX
PPC::LXSIWAX vs PPC::LFIWAX
PPC::LXSIWZX vs PPC::LFIWZX
PPC::STXSIWX vs PPC::STFIWX

Differential Revision: https://reviews.llvm.org/D38486

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318651 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AArch64][TableGen] Skip tied result operands for InstAlias
Sander de Smalen [Mon, 20 Nov 2017 14:36:40 +0000 (14:36 +0000)]
[AArch64][TableGen] Skip tied result operands for InstAlias

Summary:
This patch fixes an issue so that the right alias is printed when the instruction has tied operands. It checks the number of operands in the resulting instruction as opposed to the alias, and then skips over tied operands that should not be printed in the alias.

This allows to generate the preferred assembly syntax for the AArch64 'ins' instruction, which should always be displayed as 'mov' according to the ARM Architecture Reference Manual. Several unit tests have changed as a result, but only to reflect the preferred disassembly. Some other InstAlias patterns (movk/bic/orr) needed a slight adjustment to stop them becoming the default and breaking other unit tests.

Please note that the patch is mostly the same as https://reviews.llvm.org/D29219 which was reverted because of an issue found when running TableGen with the Address Sanitizer. That issue has been addressed in this iteration of the patch.

Reviewers: rengolin, stoklund, huntergr, SjoerdMeijer, rovka

Reviewed By: rengolin, SjoerdMeijer

Subscribers: fhahn, aemerson, javed.absar, kristof.beyls, llvm-commits

Differential Revision: https://reviews.llvm.org/D40030

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318650 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Partial ILP scheduler port from SelectionDAG to SchedulingDAG (experimental)
Valery Pykhtin [Mon, 20 Nov 2017 14:35:53 +0000 (14:35 +0000)]
AMDGPU: Partial ILP scheduler port from SelectionDAG to SchedulingDAG (experimental)

Differential revision: https://reviews.llvm.org/D39897

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318649 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LV] Model masking in VPlan, introducing VPInstructions
Gil Rapaport [Mon, 20 Nov 2017 12:01:47 +0000 (12:01 +0000)]
[LV] Model masking in VPlan, introducing VPInstructions

This patch adds a new abstraction layer to VPlan and leverages it to model the planned
instructions that manipulate masks (AND, OR, NOT), introduced during predication.

The new VPValue and VPUser classes model how data flows into, through and out
of a VPlan, forming the vertices of a planned Def-Use graph. The new
VPInstruction class is a generic single-instruction Recipe that models a
planned instruction along with its opcode, operands and users. See
VectorizationPlan.rst for more details.

Differential Revision: https://reviews.llvm.org/D38676

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318645 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM GlobalISel] Add test for RSBri. NFC
Diana Picus [Mon, 20 Nov 2017 11:05:31 +0000 (11:05 +0000)]
[ARM GlobalISel] Add test for RSBri. NFC

Add instruction selector test for RSBri, which is derived from
AsI1_rbin_irs, and make sure it doesn't get mistaken for SUBri, which is
derived from the very similar AsI1_bin_irs pattern.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318643 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM GlobalISel] Clean up binary operator tests. NFC
Diana Picus [Mon, 20 Nov 2017 10:35:35 +0000 (10:35 +0000)]
[ARM GlobalISel] Clean up binary operator tests. NFC

Remove some of the instruction selector tests for binary operators (and,
or, xor). These are all derived from the same kind of TableGen pattern,
AsI1_bin_irs, so there's no point in testing all of them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318642 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LV][X86] Support of AVX2 Gathers code generation and update the LV with this
Mohammed Agabaria [Mon, 20 Nov 2017 08:18:12 +0000 (08:18 +0000)]
[LV][X86] Support of AVX2 Gathers code generation and update the LV with this

This patch depends on: https://reviews.llvm.org/D35348

Support of pattern selection of masked gathers of AVX2 (X86\AVX2 code gen)
Update LoopVectorize to generate gathers for AVX2 processors.

Reviewers: delena, zvi, RKSimon, craig.topper, aaboud, igorb

Reviewed By: delena, RKSimon

Differential Revision: https://reviews.llvm.org/D35772

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318641 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Regenerate a test with broadcast comments. NFC
Craig Topper [Mon, 20 Nov 2017 08:15:04 +0000 (08:15 +0000)]
[X86] Regenerate a test with broadcast comments. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318640 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[IRCE] Smart range intersection
Max Kazantsev [Mon, 20 Nov 2017 06:07:57 +0000 (06:07 +0000)]
[IRCE] Smart range intersection

In rL316552, we ban intersection of unsigned latch range with signed range check and vice
versa, unless the entire range check iteration space is known positive. It was a correct
functional fix that saved us from dealing with ambiguous values, but it also appeared
to be a very restrictive limitation. In particular, in the following case:

  loop:
    %iv = phi i32 [ 0, %preheader ], [ %iv.next, %latch]
    %iv.offset = add i32 %iv, 10
    %rc = icmp slt i32 %iv.offset, %len
    br i1 %rc, label %latch, label %deopt

  latch:
    %iv.next = add i32 %iv, 11
    %cond = icmp i32 ult %iv.next, 100
    br it %cond, label %loop, label %exit

Here, the unsigned iteration range is `[0, 100)`, and the safe range for range
check is `[-10, %len - 10)`. For unsigned iteration spaces, we use unsigned
min/max functions for range intersection. Given this, we wanted to avoid dealing
with `-10` because it is interpreted as a very big unsigned value. Semantically, range
check's safe range goes through unsigned border, so in fact it is two disjoint
ranges in IV's iteration space. Intersection of such ranges is not trivial, so we prohibited
this case saying that we are not allowed to intersect such ranges.

What semantics of this safe range actually means is that we can start from `-10` and go
up increasing the `%iv` by one until we reach `%len - 10` (for simplicity let's assume that
`%len - 10`  is a reasonably big positive value).

In particular, this safe iteration space includes `0, 1, 2, ..., %len - 11`. So if we were able to return
safe iteration space `[0, %len - 10)`, we could safely intersect it with IV's iteration space. All
values in this range are non-negative, so using signed/unsigned min/max for them is unambiguous.

In this patch, we alter the algorithm of safe range calculation so that it returnes a subset of the
original safe space which is represented by one continuous range that does not go through wrap.
In order to reach this, we use modified SCEV substraction function. It can be imagined as a function
that substracts by `1` (or `-1`) as long as the further substraction does not cause a wrap in IV iteration
space. This allows us to perform IRCE in many situations when we deal with IV space and range check
of different types (in terms of signed/unsigned).

We apply this approach for both matching and not matching types of IV iteration space and the
range check. One implication of this is that now IRCE became smarter in detection of empty safe
ranges. For example, in this case:
  loop:
    %iv = phi i32 [ %begin, %preheader ], [ %iv.next, %latch]
    %iv.offset = sub i32 %iv, 10
    %rc = icmp ult i32 %iv.offset, %len
    br i1 %rc, label %latch, label %deopt

  latch:
    %iv.next = add i32 %iv, 11
    %cond = icmp i32 ult %iv.next, 100
    br it %cond, label %loop, label %exit

If `%len` was less than 10 but SCEV failed to trivially prove that `%begin - 10 >u %len- 10`,
we could end up executing entire loop in safe preloop while the main loop was still generated,
but never executed. Now, cutting the ranges so that if both `begin - 10` and `%len - 10` overflow,
we have a trivially empty range of `[0, 0)`. This in some cases prevents us from meaningless optimization.

Differential Revision: https://reviews.llvm.org/D39954

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318639 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CGP] Fix the crash caused by enable of complex addr mode
Serguei Katkov [Mon, 20 Nov 2017 05:42:36 +0000 (05:42 +0000)]
[CGP] Fix the crash caused by enable of complex addr mode

We must collect all AddModes even if they are the same.
This is due to Original value is different but we need all original
values collected as they are used as anchors in common phi finding.

Reviewers: john.brawn, reames
Reviewed By: john.brawn
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D40166

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318638 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] clzero check in getHostCPUFeatures should use getX86CpuIDAndInfo not getX86CpuI...
Craig Topper [Sun, 19 Nov 2017 23:49:19 +0000 (23:49 +0000)]
[X86] clzero check in getHostCPUFeatures should use getX86CpuIDAndInfo not getX86CpuIDAndInfoEx.

This leaf doesn't take an additional argument.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318634 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Reorder and reformat the feature bit checks in getHostCPUFeatues to keep the...
Craig Topper [Sun, 19 Nov 2017 23:30:22 +0000 (23:30 +0000)]
[X86] Reorder and reformat the feature bit checks in getHostCPUFeatues to keep the bits in order per register and encourage future additions to be in order too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318633 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[x86] add sqrt tests for partially-inline-libcalls (PR31455)
Sanjay Patel [Sun, 19 Nov 2017 17:31:37 +0000 (17:31 +0000)]
[x86] add sqrt tests for partially-inline-libcalls (PR31455)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318630 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LibCallSimplifier] allow splat vectors for pow(x, 0.5) -> sqrt() transforms
Sanjay Patel [Sun, 19 Nov 2017 16:42:27 +0000 (16:42 +0000)]
[LibCallSimplifier] allow splat vectors for pow(x, 0.5) -> sqrt() transforms

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318629 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LibCallSimplifier] partly fix pow(x, 0.5) -> sqrt() transforms
Sanjay Patel [Sun, 19 Nov 2017 16:13:14 +0000 (16:13 +0000)]
[LibCallSimplifier] partly fix pow(x, 0.5) -> sqrt() transforms

As the first test shows, we could transform an llvm intrinsic which never sets errno
into a libcall which could set errno (even though it's marked readnone?), so that's
not ideal.

It's possible that we can also transform a libcall which could set errno to an
intrinsic given the fast-math-flags constraint, but that's deferred to determine
exactly which set of FMF are needed.

Differential Revision: https://reviews.llvm.org/D40150

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318628 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstSimplify] fold and/or of fcmp ord/uno when operand is known nnan
Sanjay Patel [Sun, 19 Nov 2017 15:34:27 +0000 (15:34 +0000)]
[InstSimplify] fold and/or of fcmp ord/uno when operand is known nnan

The 'ord' and 'uno' predicates have a logic operation for NAN built into their definitions:

FCMP_ORD   =  7,  ///< 0 1 1 1    True if ordered (no nans)
FCMP_UNO   =  8,  ///< 1 0 0 0    True if unordered: isnan(X) | isnan(Y)

So we can simplify patterns like this:

(fcmp ord (known NNAN), X) && (fcmp ord X, Y) --> fcmp ord X, Y
(fcmp uno (known NNAN), X) || (fcmp uno X, Y) --> fcmp uno X, Y

It might be better to split this into (X uno 0) | (Y uno 0) as a canonicalization, but that
would be another patch.

Differential Revision: https://reviews.llvm.org/D40130

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318627 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Add test cases for rndscaless/sd intrinsics.
Craig Topper [Sun, 19 Nov 2017 06:24:26 +0000 (06:24 +0000)]
[X86] Add test cases for rndscaless/sd intrinsics.

Also fix the memop in the ins for these instructions. Not sure what effect this has.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318624 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Improve load folding of scalar rcp28 and rsqrt28 instructions using sse_load_f3...
Craig Topper [Sun, 19 Nov 2017 05:42:54 +0000 (05:42 +0000)]
[X86] Improve load folding of scalar rcp28 and rsqrt28 instructions using sse_load_f32/f64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318623 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[bpf] remove unused variable
Alexei Starovoitov [Sun, 19 Nov 2017 02:41:53 +0000 (02:41 +0000)]
[bpf] remove unused variable

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318615 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[bpf] allow direct and indirect calls
Alexei Starovoitov [Sun, 19 Nov 2017 01:35:00 +0000 (01:35 +0000)]
[bpf] allow direct and indirect calls

kernel verifier is becoming smarter and soon will support
direct and indirect function calls.
Remove obsolete error from BPF backend.
Make call to use PCRel_4 fixup.
'bpf to bpf' calls are distinguished from 'bpf to kernel' calls
by insn->src_reg == BPF_PSEUDO_CALL == 1 which is used as relocation
indicator similar to ld_imm64->src_reg == BPF_PSEUDO_MAP_FD == 1
The actual 'call' instruction remains the same for both
'bpf to kernel' and 'bpf to bpf' calls.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318614 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Switch cannonlake to use the SkylakeServer scheduling model instead of Haswell.
Craig Topper [Sun, 19 Nov 2017 01:25:30 +0000 (01:25 +0000)]
[X86] Switch cannonlake to use the SkylakeServer scheduling model instead of Haswell.

Cannonlake comes after skylake and supports avx512 so this is probably a closer model for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318613 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Add skeleton support for icelake CPU.
Craig Topper [Sun, 19 Nov 2017 01:12:00 +0000 (01:12 +0000)]
[X86] Add skeleton support for icelake CPU.

There are several patches out for review right now to implement Icelake features. This adds a CPU to collect them under.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318612 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Fix 80 column violation and remove trailing whitespace. NFC
Craig Topper [Sun, 19 Nov 2017 01:11:58 +0000 (01:11 +0000)]
[X86] Fix 80 column violation and remove trailing whitespace. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318611 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LIT] Fix testing out-of-tree Clang builds
Eric Fiselier [Sun, 19 Nov 2017 00:00:49 +0000 (00:00 +0000)]
[LIT] Fix testing out-of-tree Clang builds

Summary:
Currently, LIT configures the LLVM binary path before the Clang binary path. However this breaks testing out-of-tree Clang builds (where the LLVM binary path includes a copy of Clang).

This patch reverses the order of the paths when looking for Clang, putting the Clang binary directory first.

Reviewers: zturner, beanz, chapuni, modocache, EricWF

Reviewed By: EricWF

Subscribers: mgorny, cfe-commits, llvm-commits

Differential Revision: https://reviews.llvm.org/D40217

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318607 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MC][X86] Add test case from PR19251
Simon Pilgrim [Sat, 18 Nov 2017 23:23:25 +0000 (23:23 +0000)]
[MC][X86] Add test case from PR19251

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318605 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MC][X86] Add teet case from PR32807
Simon Pilgrim [Sat, 18 Nov 2017 23:06:42 +0000 (23:06 +0000)]
[MC][X86] Add teet case from PR32807

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318603 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix use of config.h in public headers.
Eric Fiselier [Sat, 18 Nov 2017 22:42:26 +0000 (22:42 +0000)]
Fix use of config.h in public headers.

The CodeGenCoverage.h header is installed, but it references
the build-only header "llvm/Config/config.h". This breaks use
of the CodeGenCoverage.h header once it is installed, because config.h isn't
available.

This patch fixes the error by moving the config.h include from
the CodeGenCoverage.h header (where it's not needed), to the
CodeGenCoverage.cpp source file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318602 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][AVX512VL] Add AVX512VL tests to the vselect packss tests.
Simon Pilgrim [Sat, 18 Nov 2017 19:47:59 +0000 (19:47 +0000)]
[X86][AVX512VL] Add AVX512VL tests to the vselect packss tests.

PR34553 has gone, adding tests to ensure it doesn't come back.

vselect_packss_v16i64 still has some awful codegen on AVX512 targets....

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318599 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Add another gather test with v8i8 sign extended indices.
Craig Topper [Sat, 18 Nov 2017 19:25:35 +0000 (19:25 +0000)]
[X86] Add another gather test with v8i8 sign extended indices.

This requires the indices to be legalized and sign extended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318597 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Simplify the gather/scatter isel predicates.
Craig Topper [Sat, 18 Nov 2017 19:05:12 +0000 (19:05 +0000)]
[X86] Simplify the gather/scatter isel predicates.

We don't need a dyn_cast, the predicate already specified the base node. We only need to check the type of the index, the base ptr is guaranteed to be scalar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318596 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Qualify a few places with ExperimentalVectorWideningLegalization.
Craig Topper [Sat, 18 Nov 2017 18:49:16 +0000 (18:49 +0000)]
[X86] Qualify a few places with ExperimentalVectorWideningLegalization.

I'm playing around with this flag and these places cause errors if not qualified.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318595 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Add todo comment for TRUNC(SUB(X,C)) -> SUB(TRUNC(X),C')
Simon Pilgrim [Sat, 18 Nov 2017 18:33:07 +0000 (18:33 +0000)]
[X86] Add todo comment for TRUNC(SUB(X,C)) -> SUB(TRUNC(X),C')

As discussed on PR35295, but it causes regressions in combineSubToSubus which need to be addressed first

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318594 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CallSiteSplitting] Remove some indirection (NFC).
Florian Hahn [Sat, 18 Nov 2017 18:14:13 +0000 (18:14 +0000)]
[CallSiteSplitting] Remove some indirection (NFC).

Summary:
With this patch I tried to reduce the complexity of the code sightly, by
removing some indirection. Please let me know what you think.

Reviewers: junbuml, mcrosier, davidxl

Reviewed By: junbuml

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D40037

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318593 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[x86] add tests for unnecessary shuffling; NFC
Sanjay Patel [Sat, 18 Nov 2017 16:25:38 +0000 (16:25 +0000)]
[x86] add tests for unnecessary shuffling; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318592 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Output cfi directives for saved XMM registers even if no GPRs are saved
Martin Storsjo [Sat, 18 Nov 2017 06:23:48 +0000 (06:23 +0000)]
[X86] Output cfi directives for saved XMM registers even if no GPRs are saved

This makes sure that functions that only clobber xmm registers
(on win64) also get the right cfi directives, if dwarf exceptions
are enabled.

Differential Revision: https://reviews.llvm.org/D40191

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318591 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Fix typo in variable name. NFC
Craig Topper [Sat, 18 Nov 2017 05:09:55 +0000 (05:09 +0000)]
[X86] Fix typo in variable name. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318590 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AArch64] Map G_LOAD on FPR when the definition goes to a copy to FPR
Quentin Colombet [Sat, 18 Nov 2017 04:28:59 +0000 (04:28 +0000)]
[AArch64] Map G_LOAD on FPR when the definition goes to a copy to FPR

We used to detect loads feeding fp instructions, but we were
failing to take into account cases where this happens through copies.
For instance, loads can fed copies coming from the ABI lowering
of floating point arguments/results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318589 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AArch64] Map G_STORE on FPR when the source comes from a FPR copy
Quentin Colombet [Sat, 18 Nov 2017 04:28:58 +0000 (04:28 +0000)]
[AArch64] Map G_STORE on FPR when the source comes from a FPR copy

We used to detect that stores were fed by fp instructions, but we were
failing to take into account cases where this happens through copies.
For instance, stores can be fed by copies coming from the ABI lowering
of floating point arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318588 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[RegisterBankInfo] Relax the assert of having matching type sizes on default mappings
Quentin Colombet [Sat, 18 Nov 2017 04:28:58 +0000 (04:28 +0000)]
[RegisterBankInfo] Relax the assert of having matching type sizes on default mappings

Instead of asserting that the type sizes are exactly equal, we check
that the new size is big enough to contain the original type.
We have to relax this constrain because, right now, we sometimes
specify that things that are smaller than a storage type are legal
instead of widening everything to the size of a storage type.
E.g., we say that G_AND s16 is legal and we map that on GPR32.

This is something we may revisit in the future (either by changing
the legalization process or keeping track separately of the storage
size and the size of the type), but let us reflect the reality of
the situation for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318587 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AArch64][RegisterBankInfo] Teach instruction mapping about gpr32 -> fpr16 cross...
Quentin Colombet [Sat, 18 Nov 2017 04:28:56 +0000 (04:28 +0000)]
[AArch64][RegisterBankInfo] Teach instruction mapping about gpr32 -> fpr16 cross copies

Turns out this copies can actually occur because of the way we lower the
ABI for half.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318586 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoReorder static functions. NFC.
Rafael Espindola [Sat, 18 Nov 2017 02:12:53 +0000 (02:12 +0000)]
Reorder static functions. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318584 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSplit realPathFromHandle in two.
Rafael Espindola [Sat, 18 Nov 2017 02:05:59 +0000 (02:05 +0000)]
Split realPathFromHandle in two.

By having an UTF-16 version we avoid some code duplication in calling
GetFinalPathNameByHandleW.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318583 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[asan] Add a full redzone after every stack variable
Walter Lee [Sat, 18 Nov 2017 01:13:18 +0000 (01:13 +0000)]
[asan] Add a full redzone after every stack variable

We were not doing that for large shadow granularity.  Also add more
stack frame layout tests for large shadow granularity.

Differential Revision: https://reviews.llvm.org/D39475

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318581 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert "[asan] Use dynamic shadow on 32-bit Android" and 3 more.
Evgeniy Stepanov [Sat, 18 Nov 2017 00:22:34 +0000 (00:22 +0000)]
Revert "[asan] Use dynamic shadow on 32-bit Android" and 3 more.

Revert the following commits:
  r318369 [asan] Fallback to non-ifunc dynamic shadow on android<22.
  r318235 [asan] Prevent rematerialization of &__asan_shadow.
  r317948 [sanitizer] Remove unnecessary attribute hidden.
  r317943 [asan] Use dynamic shadow on 32-bit Android.

MemoryRangeIsAvailable() reads /proc/$PID/maps into an mmap-ed buffer
that may overlap with the address range that we plan to use for the
dynamic shadow mapping. This is causing random startup crashes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318575 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[globalisel][tablegen] Generalize pointer-type inference by introducing ptypeN. NFC
Daniel Sanders [Sat, 18 Nov 2017 00:16:44 +0000 (00:16 +0000)]
[globalisel][tablegen] Generalize pointer-type inference by introducing ptypeN. NFC

ptypeN is functionally the same as typeN except that it informs the
SelectionDAG importer that an operand should be treated as a pointer even
if it was written as iN. This is important for patterns that use iN instead
of iPTR to represent pointers. E.g.:
  (set GPR64:$dst, (load GPR64:$addr))

Previously, this was handled as a hardcoded special case for the appropriate
operands to G_LOAD and G_STORE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318574 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[lit] Try to improve Ctrl-C behavior on Windows
Reid Kleckner [Fri, 17 Nov 2017 23:52:33 +0000 (23:52 +0000)]
[lit] Try to improve Ctrl-C behavior on Windows

This functionality was broken during a refactor a while back because
'pool' is no longer in scope.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318572 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake][runtimes] Don't passthrough prefixes for non-default targets
Petr Hosek [Fri, 17 Nov 2017 23:51:53 +0000 (23:51 +0000)]
[CMake][runtimes] Don't passthrough prefixes for non-default targets

The passthrough is useful for setting up the options for the default
build, but we already have a different mechanism to pass CMake flags
to builds for builtins and runtimes targets so this is not really
needed there. Furthermore, when the flags are set for the default
build, with the prefix passthrough set we have to explicitly override
all options in other targets which can be cumbersome.

Differential Revision: https://reviews.llvm.org/D39988

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318571 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix an issue with llvm lit tool substitutions.
Zachary Turner [Fri, 17 Nov 2017 22:51:43 +0000 (22:51 +0000)]
Fix an issue with llvm lit tool substitutions.

When using an installed clang with an in-tree llvm, we were not
searching in the right paths for the tools.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318564 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MC] Fix regression tests on Windows when git “core.autocrlf” is set to true.
Zhen Cao [Fri, 17 Nov 2017 21:59:43 +0000 (21:59 +0000)]
[MC] Fix regression tests on Windows when git “core.autocrlf” is set to true.

Differential Revision: https://reviews.llvm.org/D39737

This is the second attempt to commit this. The test was broken on Linux in the first attempt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318560 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Move hazard avoidance out of waitcnt pass.
Matt Arsenault [Fri, 17 Nov 2017 21:35:32 +0000 (21:35 +0000)]
AMDGPU: Move hazard avoidance out of waitcnt pass.

This is mostly moving VMEM clause breaking into
the hazard recognizer. Also move another hazard
currently handled in the waitcnt pass.

Also stops breaking clauses unless xnack is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318557 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-profdata] Don't treat non-fatal merge errors as fatal
Vedant Kumar [Fri, 17 Nov 2017 21:18:32 +0000 (21:18 +0000)]
[llvm-profdata] Don't treat non-fatal merge errors as fatal

This fixes an issue seen on the coverage bot:
http://lab.llvm.org:8080/green/view/Experimental/job/clang-stage2-coverage-R/1930

Profile merging shouldn't fail if a single counter mismatch is detected.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318555 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agollvm-demangle-fuzzer: Link in Support
Jonas Hahnfeld [Fri, 17 Nov 2017 20:46:55 +0000 (20:46 +0000)]
llvm-demangle-fuzzer: Link in Support

The refactoring in r318407 transiently includes abi-breaking.h
which defines EnableABIBreakingChecks. This breaks my Debug
build because this fuzzer did not link in Support with the symbol.

Differential Revision: https://reviews.llvm.org/D40190

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318553 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LICM] Fix PR35342
Jun Bum Lim [Fri, 17 Nov 2017 20:38:25 +0000 (20:38 +0000)]
[LICM] Fix PR35342

Summary: This change fix PR35342 by replacing only the current use with undef in unreachable blocks.

Reviewers: efriedma, mcrosier, igor-laevsky

Reviewed By: efriedma

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D40184

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318551 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUse TempFile in the implementation of LockFileManager.
Rafael Espindola [Fri, 17 Nov 2017 20:06:41 +0000 (20:06 +0000)]
Use TempFile in the implementation of LockFileManager.

This move some of the complexity over to the lower level TempFile.

It also makes it a bit more explicit where errors are ignored since we
now have a call to consumeError.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318550 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[PM/Unswitch] Teach SimpleLoopUnswitch to do non-trivial unswitching,
Chandler Carruth [Fri, 17 Nov 2017 19:58:36 +0000 (19:58 +0000)]
[PM/Unswitch] Teach SimpleLoopUnswitch to do non-trivial unswitching,
making it no longer even remotely simple.

The pass will now be more of a "full loop unswitching" pass rather than
anything substantively simpler than any other approach. I plan to rename
it accordingly once the dust settles.

The key ideas of the new loop unswitcher are carried over for
non-trivial unswitching:
1) Fully unswitch a branch or switch instruction from inside of a loop to
   outside of it.
2) Update the CFG and IR. This avoids needing to "remember" the
   unswitched branches as well as avoiding excessively cloning and
   reliance on complex parts of simplify-cfg to cleanup the cfg.
3) Update the analyses (where we can) rather than just blowing them away
   or relying on something else updating them.

Sadly, #3 is somewhat compromised here as the dominator tree updates
were too complex for me to want to reason about. I will need to make
another attempt to do this now that we have a nice dynamic update API
for dominators. However, we do adhere to #3 w.r.t. LoopInfo.

This approach also adds an important principls specific to non-trivial
unswitching: not *all* of the loop will be duplicated when unswitching.
This fact allows us to compute the cost in terms of how much *duplicate*
code is inserted rather than just on raw size. Unswitching conditions
which essentialy partition loops will work regardless of the total loop
size.

Some remaining issues that I will be addressing in subsequent commits:
- Handling unstructured control flow.
- Unswitching 'switch' cases instead of just branches.
- Moving to the dynamic update API for dominators.

Some high-level, interesting limitationsV that folks might want to push
on as follow-ups but that I don't have any immediate plans around:
- We could be much more clever about not cloning things that will be
  deleted. In fact, we should be able to delete *nothing* and do
  a minimal number of clones.
- There are many more interesting selection criteria for which branch to
  unswitch that we might want to look at. One that I'm interested in
  particularly are a set of conditions which all exit the loop and which
  can be merged into a single unswitched test of them.

Differential revision: https://reviews.llvm.org/D34200

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318549 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMIRParser: Avoid reading uninitialized memory on generic vregs
Justin Bogner [Fri, 17 Nov 2017 18:51:20 +0000 (18:51 +0000)]
MIRParser: Avoid reading uninitialized memory on generic vregs

If a vreg's bank is specified in the registers block and one of its
defs or uses also specifies the bank, we end up checking that the
RegBank is equal to diagnose conflicting banks. The problem comes up
for generic vregs, where we weren't fully initializing the VRegInfo
when parsing the registers block, so we'd end up comparing a null
pointer to uninitialized memory.

This fixes a non-deterministic failure when round tripping through MIR
with generic vregs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318543 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Merge scheduling tests for SHLD/SHRD
Simon Pilgrim [Fri, 17 Nov 2017 18:35:49 +0000 (18:35 +0000)]
[X86] Merge scheduling tests for SHLD/SHRD

Reduces spsce used and makes it easier to compare the 2 values for the equivalent instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318541 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ThinLTO] Remove too aggressive assertion in building function call graph.
Volodymyr Sapsai [Fri, 17 Nov 2017 18:28:05 +0000 (18:28 +0000)]
[ThinLTO] Remove too aggressive assertion in building function call graph.

The assertion was introduced in r317853 but there are cases when a call
isn't handled either as direct or indirect. In this case we add a
reference graph edge but not a call graph edge.

Reviewers: tejohnson

Reviewed By: tejohnson

Subscribers: mehdi_amini, inglorion, eraman, hiraditya, efriedma, llvm-commits

Differential Revision: https://reviews.llvm.org/D40056

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318540 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTry to fix the windows build.
Rafael Espindola [Fri, 17 Nov 2017 17:46:39 +0000 (17:46 +0000)]
Try to fix the windows build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318535 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUse TempFile in dsymutil.
Rafael Espindola [Fri, 17 Nov 2017 17:33:09 +0000 (17:33 +0000)]
Use TempFile in dsymutil.

I don't think there is any functionality change, but the code is
easier to understand IMHO.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318534 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert "[MC] Fix regression tests on Windows when git “core.autocrlf” is set to true."
Rafael Espindola [Fri, 17 Nov 2017 17:31:20 +0000 (17:31 +0000)]
Revert "[MC] Fix regression tests on Windows when git “core.autocrlf” is set to true."

This reverts commit r318528.

MC/AsmParser/preserve-comments-crlf.s fails on linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318533 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake] Allow for Solaris ld -V output on stdout
Michal Gorny [Fri, 17 Nov 2017 17:12:14 +0000 (17:12 +0000)]
[CMake] Allow for Solaris ld -V output on stdout

In recent versions of Solaris 11.4 (previously 12), ld -V output went to
stdout instead of stderr. Since AddLLVM.cmake only expects it on stderr,
Solaris ld wasn't properly detected and options not understood by it are
passed during the build.

The following patch fixes this, allowing for both variants.

Tested on i386-pc-solaris2.11.4 (on top of D35755 which is needed for
proper Solaris support).

Patch by Rainer Orth.

Differential Revision: https://reviews.llvm.org/D39601

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318532 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AArch64] Adjust the cost model for Exynos M1 and M2
Evandro Menezes [Fri, 17 Nov 2017 16:42:15 +0000 (16:42 +0000)]
[AArch64] Adjust the cost model for Exynos M1 and M2

Improve the accuracy of the model by specifying the proper number of uops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318531 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MC] Fix regression tests on Windows when git “core.autocrlf” is set to true.
Zhen Cao [Fri, 17 Nov 2017 16:17:56 +0000 (16:17 +0000)]
[MC] Fix regression tests on Windows when git “core.autocrlf” is set to true.

Differential Revision: https://reviews.llvm.org/D39737

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318528 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AMDGPU][MC][GFX9][disassembler] Corrected decoding of op_sel_hi for v_mad_mix*
Dmitry Preobrazhensky [Fri, 17 Nov 2017 15:15:40 +0000 (15:15 +0000)]
[AMDGPU][MC][GFX9][disassembler] Corrected decoding of op_sel_hi for v_mad_mix*

See bug 35148: https://bugs.llvm.org//show_bug.cgi?id=35148

Reviewers: tamazov, SamWot, arsenm

Differential Revision: https://reviews.llvm.org/D39492

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318526 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Support][CachePruning] Fix regression in pruning interval
Ben Dunbobbin [Fri, 17 Nov 2017 14:42:18 +0000 (14:42 +0000)]
[Support][CachePruning] Fix regression in pruning interval

Fixed broken comparison.
borked by: rL284966 (see: https://reviews.llvm.org/D25730).

Differential Revision: https://reviews.llvm.org/D40119

This is a second attempt to commit this.
The first attempt broke lld and gold tests that had been written against
the incorrect behaivour.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318524 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[gold] Fix broken thin LTO cache test NFC
Andrew Ng [Fri, 17 Nov 2017 12:00:57 +0000 (12:00 +0000)]
[gold] Fix broken thin LTO cache test NFC

Fix test as it is assuming that the cache pruning is always being
performed by default. Explicitly set prune interval to 0s to ensure
pruning is always performed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318520 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] Use dwarf exception handling on MinGW
Martin Storsjo [Fri, 17 Nov 2017 08:04:40 +0000 (08:04 +0000)]
[ARM] Use dwarf exception handling on MinGW

Enabling and using dwarf exceptions seems like an easier path
to take, than to make the COFF/ARM backend output EHABI directives.
Previously, no EH model was enabled at all on this target.

There's no point in setting UseIntegratedAssembler to false since
GNU binutils doesn't support Windows on ARM, and since we don't
need to support external assembler, we don't need to use register
numbers in cfi directives.

Differential Revision: https://reviews.llvm.org/D39532

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318510 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SelectionDAG] Allow custom vector widening through ReplaceNodeResults to handle...
Craig Topper [Fri, 17 Nov 2017 07:03:57 +0000 (07:03 +0000)]
[SelectionDAG] Allow custom vector widening through ReplaceNodeResults to handle nodes with chain outputs.

Previously we were assuming all results were vectors and calling SetWidenedVector, but if its a chain result we should just replace uses instead.

This fixes an error found by expensive checks after r318368.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318509 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[IRCE] Remove folding of two range checks into RANGE_CHECK_BOTH
Max Kazantsev [Fri, 17 Nov 2017 06:49:26 +0000 (06:49 +0000)]
[IRCE] Remove folding of two range checks into RANGE_CHECK_BOTH

The logic of replacing of a couple `RANGE_CHECK_LOWER + RANGE_CHECK_UPPER`
into `RANGE_CHECK_BOTH` in fact duplicates the logic of range intersection which
happens when we calculate safe iteration space. Effectively, the result of intersection of
these ranges doesn't differ from the range of merged range check.

We chose to remove duplicating logic in favor of code simplicity.

Differential Revision: https://reviews.llvm.org/D39589

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318508 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Replace list of SMEM buffer opcodes
Matt Arsenault [Fri, 17 Nov 2017 04:18:26 +0000 (04:18 +0000)]
AMDGPU: Replace list of SMEM buffer opcodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318506 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Fix breaking SMEM clauses
Matt Arsenault [Fri, 17 Nov 2017 04:18:24 +0000 (04:18 +0000)]
AMDGPU: Fix breaking SMEM clauses

This was completely ignoring subregisters,
so was not very useful. Also only break them
if xnack is actually enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318505 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-profdata] Fix a dangling reference to an error string
Vedant Kumar [Fri, 17 Nov 2017 02:58:23 +0000 (02:58 +0000)]
[llvm-profdata] Fix a dangling reference to an error string

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318502 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[GISel]: DCE copy instructions during legalization
Aditya Nandakumar [Fri, 17 Nov 2017 02:44:55 +0000 (02:44 +0000)]
[GISel]: DCE copy instructions during legalization

We might have instructions such as ext(copy(trunc)), and while cleaning
up legalization artifacts, we can also dce the copies that are in
between legalization artifacts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318501 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SelectionDAG] Consolidate (t|T)ransferDbgValues methods, NFC (reapply)
Vedant Kumar [Fri, 17 Nov 2017 01:48:33 +0000 (01:48 +0000)]
[SelectionDAG] Consolidate (t|T)ransferDbgValues methods, NFC (reapply)

TransferDbgValues (capital 'T') is wired into ReplaceAllUsesWith, and
transferDbgValues (lowercase 't') is used elsewhere (e.g in Legalize).

Both functions should be doing the exact same thing. This patch
consolidates the logic into one place.

This was reverted in r318455 because some newly introduced asserts,
which I thought were NFC, were firing. I filed PR35338. For now I've
weakened the asserts.

Testing: check-llvm, check-clang, and a stage2 Rel+Deb build of clang

Differential Revision: https://reviews.llvm.org/D40104

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318498 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[asan] Test ASan instrumentation for shadow scale value of 5
Walter Lee [Fri, 17 Nov 2017 01:15:31 +0000 (01:15 +0000)]
[asan] Test ASan instrumentation for shadow scale value of 5

Add additional RUN clauses to test for -asan-mapping-scale=5 in
selective tests, with special CHECK statements where needed.

Differential Revision: https://reviews.llvm.org/D39775

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318493 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix a bunch more layering of CodeGen headers that are in Target
David Blaikie [Fri, 17 Nov 2017 01:07:10 +0000 (01:07 +0000)]
Fix a bunch more layering of CodeGen headers that are in Target

All these headers already depend on CodeGen headers so moving them into
CodeGen fixes the layering (since CodeGen depends on Target, not the
other way around).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318490 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDon't #include MemoryBuffer.h from Host.h.
Zachary Turner [Fri, 17 Nov 2017 01:00:35 +0000 (01:00 +0000)]
Don't #include MemoryBuffer.h from Host.h.

It turns out this #include isn't used from Host.h anyway,
but by having it it causes circular include dependencies.
This issues only surfaced while I was working on a separate
patch, so I'm submitting this first so that it's independent
of the other, unrelated patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318489 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[PredicateInfo] Add comment about why we require stable sort
Mandeep Singh Grang [Fri, 17 Nov 2017 00:43:24 +0000 (00:43 +0000)]
[PredicateInfo] Add comment about why we require stable sort

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318487 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRe-revert "Refactor debuginfo-tests"
Zachary Turner [Fri, 17 Nov 2017 00:41:18 +0000 (00:41 +0000)]
Re-revert "Refactor debuginfo-tests"

This is still broken because it causes certain tests to be
run twice with slightly different configurations, which is
wrong in some cases.

You can observe this by running:

  ninja -nv check-all | grep debuginfo-tests

And seeing that it passes clang/test and clang/test/debuginfo-tests
to lit, which causes it to run debuginfo-tests twice.  The fix is
going to involve either:

  a) figuring out that we're running in this "deprecated" configuration,
     and then deleting the clang/test/debuginfo-tests path, which should
     cause it to behave identically to before, or:
  b) make lit smart enough that it doesn't descend into a sub-suite if
     that sub-suite already has a lit.cfg file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318486 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoCurrent implementation of Value::replaceUsesExceptBlockAddr() uses UseList
Dmitry Mikulin [Fri, 17 Nov 2017 00:30:24 +0000 (00:30 +0000)]
Current implementation of Value::replaceUsesExceptBlockAddr() uses UseList
iterator to walk the list which keeps changing inside the loop. When the
UseList contains several uses with the same user, we end processing the same
user more than once, which leads to an assert.

With this fix, unique users are saved and processed later to avoid
processing duplicates.

Differential Revision: https://reviews.llvm.org/D39864

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318477 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] 't' asm constraint should accept i32
Yi Kong [Thu, 16 Nov 2017 23:38:17 +0000 (23:38 +0000)]
[ARM] 't' asm constraint should accept i32

't' constraint normally only accepts f32 operands, but for VCVT the
operands can be i32. LLVM is overly restrictive and rejects asm like:

  float foo() {
    float result;
    __asm__ __volatile__(
      "vcvt.f32.s32 %[result], %[arg1]\n"
      : [result]"=t"(result)
      : [arg1]"t"(0x01020304) );
    return result;
  }

Relax the value type for 't' constraint to either f32 or i32.

Differential Revision: https://reviews.llvm.org/D40137

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318472 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Add DAG combine to remove sext i32->i64 from gather/scatter instructions.
Craig Topper [Thu, 16 Nov 2017 23:09:06 +0000 (23:09 +0000)]
[X86] Add DAG combine to remove sext i32->i64 from gather/scatter instructions.

Only do this pre-legalize in case we're using the sign extend to legalize for KNL.

This recovers all of the tests that changed when I stopped SelectionDAGBuilder from deleting sign extends.

There's more work that could be done here particularly to fix the i8->i64 test case that experienced split.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318468 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Add gather test with index sign extended from i8 type.
Craig Topper [Thu, 16 Nov 2017 23:09:03 +0000 (23:09 +0000)]
[X86] Add gather test with index sign extended from i8 type.

Previously SelectionDAGBuilder would remove this sign extend leading to a failure during isel.

The codegen here isn't very nice as we ended up triggering a split.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318467 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Don't remove sign extend of gather/scatter indices during SelectionDAGBuilder.
Craig Topper [Thu, 16 Nov 2017 23:08:57 +0000 (23:08 +0000)]
[X86] Don't remove sign extend of gather/scatter indices during SelectionDAGBuilder.

The sign extend might be from an i16 or i8 type and was inserted by InstCombine to match the pointer width. X86 gather legalization isn't currently detecting this to reinsert a sign extend to make things legal.

It's a bit weird for the SelectionDAGBuilder to do this kind of optimization in the first place. With this removed we can at least lean on InstCombine somewhat to ensure the index is i32 or i64.

I'll work on trying to recover some of the test cases by removing sign extends in the backend when its safe to do so with an understanding of the current legalizer capabilities.

This should fix PR30690.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318466 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Support] Support NetBSD PaX MPROTECT in sys::Memory.
Lang Hames [Thu, 16 Nov 2017 23:04:44 +0000 (23:04 +0000)]
[Support] Support NetBSD PaX MPROTECT in sys::Memory.

Removes AllocateRWX, setWritable and setExecutable from sys::Memory and
standardizes on allocateMappedMemory / protectMappedMemory. The
allocateMappedMemory method is updated to request full permissions for memory
blocks so that they can be marked executable later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318464 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAttempt to fix inscrutible build break...
David Blaikie [Thu, 16 Nov 2017 22:40:02 +0000 (22:40 +0000)]
Attempt to fix inscrutible build break...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318463 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix some undefined beahvior in FileMapping.
Zachary Turner [Thu, 16 Nov 2017 22:39:55 +0000 (22:39 +0000)]
Fix some undefined beahvior in FileMapping.

This was broken when building a 32-bit native toolchain, as
shifting a size_t right by 32 is UB when sizeof(size_t) == 8.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318462 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agollvm-readobj/ARMEHABIPrinter.h: Make this a real/modular header
David Blaikie [Thu, 16 Nov 2017 22:30:36 +0000 (22:30 +0000)]
llvm-readobj/ARMEHABIPrinter.h: Make this a real/modular header

Had several non-inline/strong function definitions that needed to be
marked inline, etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318461 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoConvert the last use of sys::fs::createUniqueFile in bugpoint.
Rafael Espindola [Thu, 16 Nov 2017 21:53:51 +0000 (21:53 +0000)]
Convert the last use of sys::fs::createUniqueFile in bugpoint.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318459 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoConvert another use of createUniqueFile to TempFile::create.
Rafael Espindola [Thu, 16 Nov 2017 21:40:10 +0000 (21:40 +0000)]
Convert another use of createUniqueFile to TempFile::create.

This one requires a new small feature in TempFile: the ability to keep
the temporary file with the temporary name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318458 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake][runtimes] Use cmake_parse_arguments in runtimes functions
Petr Hosek [Thu, 16 Nov 2017 21:28:54 +0000 (21:28 +0000)]
[CMake][runtimes] Use cmake_parse_arguments in runtimes functions

Passing lists to functions in CMake is tricky, any list argument
has to be quoted otherwise it'll be expanded. To avoid this issue,
use cmake_parse_arguments in runtime functions and pass lists using
a keyword argument which eliminates any ambiguity when dealing with
lists.

Differential Revision: https://reviews.llvm.org/D40087

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318457 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert "[SelectionDAG] Consolidate (t|T)ransferDbgValues methods, NFC."
Vedant Kumar [Thu, 16 Nov 2017 21:08:51 +0000 (21:08 +0000)]
Revert "[SelectionDAG] Consolidate (t|T)ransferDbgValues methods, NFC."

This reverts commit r318448. It looks like some of the asserts need to
be weakened.

http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-autoconf/builds/16296

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318455 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[RISCV] Fix 64-bit data layout mismatch between backend and target description
Mandeep Singh Grang [Thu, 16 Nov 2017 20:30:49 +0000 (20:30 +0000)]
[RISCV] Fix 64-bit data layout mismatch between backend and target description

Reviewers: asb

Reviewed By: asb

Subscribers: rbar, johnrusso, simoncook, jordy.potman.lists, llvm-commits

Differential Revision: https://reviews.llvm.org/D40145

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318454 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Pre-truncate gather/scatter indices that have element sizes larger than 64...
Craig Topper [Thu, 16 Nov 2017 20:23:22 +0000 (20:23 +0000)]
[X86] Pre-truncate gather/scatter indices that have element sizes larger than 64-bits before Legalize.

The wider element type will normally cause legalize to try to split and scalarize the gather/scatter, but we can't handle that. Instead, truncate the index early so the gather/scatter node is insulated from the legalization.

This really shouldn't happen in practice since InstCombine will normalize index types to the same size as pointers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318452 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] DAGCombinerInfo is in TargetLowering not X86TargetLowering.
Craig Topper [Thu, 16 Nov 2017 20:23:17 +0000 (20:23 +0000)]
[X86] DAGCombinerInfo is in TargetLowering not X86TargetLowering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318451 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[DAGCombiner] Use cast instead of an unchecked dyn_cast.
Craig Topper [Thu, 16 Nov 2017 20:23:12 +0000 (20:23 +0000)]
[DAGCombiner] Use cast instead of an unchecked dyn_cast.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318450 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agollvm-objdump: Add missing requires for the test in r313921
Justin Bogner [Thu, 16 Nov 2017 20:20:56 +0000 (20:20 +0000)]
llvm-objdump: Add missing requires for the test in r313921

This test needs an arm backend in addition to the arm64 one that's
already checked in the test directory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318449 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SelectionDAG] Consolidate (t|T)ransferDbgValues methods, NFC.
Vedant Kumar [Thu, 16 Nov 2017 19:50:24 +0000 (19:50 +0000)]
[SelectionDAG] Consolidate (t|T)ransferDbgValues methods, NFC.

TransferDbgValues (capital 'T') is wired into ReplaceAllUsesWith, and
transferDbgValues (lowercase 't') is used elsewhere (e.g in Legalize).

Both functions should be doing the exact same thing. This patch
consolidates the logic into one place.

Differential Revision: https://reviews.llvm.org/D40104

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@318448 91177308-0d34-0410-b5e6-96231b3b80d8