Oliver Stannard [Mon, 22 Jul 2019 08:44:36 +0000 (08:44 +0000)]
[IPRA][ARM] Make use of the "returned" parameter attribute
ARM has code to recognise uses of the "returned" function parameter
attribute which guarantee that the value passed to the function in r0
will be returned in r0 unmodified. IPRA replaces the regmask on call
instructions, so needs to be told about this to avoid reverting the
optimisation.
Jay Foad [Mon, 22 Jul 2019 07:19:44 +0000 (07:19 +0000)]
[AMDGPU] Save some work when an atomic op has no uses
Summary:
In the atomic optimizer, save doing a bunch of work and generating a
bunch of dead IR in the fairly common case where the result of an
atomic op (i.e. the value that was in memory before the atomic op was
performed) is not used. NFC.
[Loop Peeling] Fix the handling of branch weights of peeled off branches.
Current algorithm to update branch weights of latch block and its copies is
based on the assumption that number of peeling iterations is approximately equal
to trip count.
However it is not correct. According to profitability check in one case we can decide to peel
in case it helps to reduce the number of phi nodes. In this case the number of peeled iteration
can be less then estimated trip count.
This patch introduces another way to set the branch weights to peeled of branches.
Let F is a weight of the edge from latch to header.
Let E is a weight of the edge from latch to exit.
F/(F+E) is a probability to go to loop and E/(F+E) is a probability to go to exit.
Then, Estimated TripCount = F / E.
For I-th (counting from 0) peeled off iteration we set the the weights for
the peeled latch as (TC - I, 1). It gives us reasonable distribution,
The probability to go to exit 1/(TC-I) increases. At the same time
the estimated trip count of remaining loop reduces by I.
As a result after peeling off N iteration the weights will be
(F - N * E, E) and trip count of loop becomes
F / E - N or TC - N.
The idea is taken from the review of the patch D63918 proposed by Philip.
[InstCombine] Add foldAndOfICmps test cases inspired by PR42691.
icmp ne %x, INT_MIN can be treated similarly to icmp sgt %x, INT_MIN.
icmp ne %x, INT_MAX can be treated similarly to icmp slt %x, INT_MAX.
icmp ne %x, UINT_MAX can be treated similarly to icmp ult %x, UINT_MAX.
We already treat icmp ne %x, 0 similarly to icmp ugt %x, 0
[InstCombine] Remove insertRangeTest code that handles the equality case.
For equality, the function called getTrue/getFalse with the VT
of the comparison input. But getTrue/getFalse need the boolean VT.
So if this code ever executed, it would assert.
I believe these cases are removed by InstSimplify so we don't get here.
So this patch just fixes up an assert to exclude the equality
possibility and removes the broken code.
[InstCombine] Don't use AddOne/SubOne to see if two APInts are 1 apart. Use APInt operations instead. NFCI
AddOne/SubOne create new Constant objects. That seems heavy for
comparing ConstantInts which wrap APInts. Just do the math on
on the APInts and compare them.
Roman Lebedev [Sat, 20 Jul 2019 16:33:15 +0000 (16:33 +0000)]
[Codegen][SelectionDAG] X u% C == 0 fold: non-splat vector improvements
Summary:
Four things here:
1. Generalize the fold to handle non-splat divisors. Reasonably trivial.
2. Unban power-of-two divisors. I don't see any reason why they should
be illegal.
* There is no ban in Hacker's Delight
* I think the ban came from the same bug that caused the miscompile
in the base patch - in `floor((2^W - 1) / D)` we were dividing by
`D0` instead of `D`, and we **were** ensuring that `D0` is not `1`,
which made sense.
3. Unban `1` divisors. I no longer believe Hacker's Delight actually says
that the fold is invalid for `D = 0`. Further considerations:
* We know that
* `(X u% 1) == 0` can be constant-folded to `1`,
* `(X u% 1) != 0` can be constant-folded to `0`,
* Also, we know that
* `X u<= -1` can be constant-folded to `1`,
* `X u> -1` can be constant-folded to `0`,
* https://godbolt.org/z/7jnZJX https://rise4fun.com/Alive/oF6p
* We know will end up with the following:
`(setule/setugt (rotr (mul N, P), K), Q)`
* Therefore, for given new DAG nodes and comparison predicates
(`ule`/`ugt`), we will still produce the correct answer if:
`Q` is a all-ones constant; and both `P` and `K` are *anything*
other than `undef`.
* The fold will indeed produce `Q = all-ones`.
4. Try to re-splat the `P` and `K` vectors - we don't care about
their values for the lanes where divisor was `1`.
Simon Pilgrim [Sat, 20 Jul 2019 15:20:11 +0000 (15:20 +0000)]
[X86][SSE] Use PSADBW to improve vXi8 sum reduction (PR42674)
As detailed on PR42674, we can reduce a vXi8 down until we have the final <8 x i8>, and then use PSADBW with zero, to sum those values. We then extract the bottom i8, discarding any overflow from the upper bits of the i16 result.
[GlobalISel][AArch64] Contract trivial same-size cross-bank copies into G_STOREs
Sometimes, you can end up with cross-bank copies between same-sized GPRs and
FPRs, which feed into G_STOREs. When these copies feed only into stores, they
aren't necessary; we can just store using the original register bank.
This provides some minor code size savings for some floating point SPEC
benchmarks. (Around 0.2% for 453.povray and 450.soplex)
This issue doesn't seem to show up due to regbankselect or anything similar. So,
this patch introduces an early select function, `contractCrossBankCopyIntoStore`
which performs the contraction when possible. The selector then continues
normally and selects the correct store opcode, eliminating needless copies
along the way.
Daniel Sanders [Fri, 19 Jul 2019 22:46:47 +0000 (22:46 +0000)]
Re-commit: r366610 and r366612: Expand pseudo-components before embedding in llvm-config
There were two main problems:
* The 'nativecodegen' pseudo-component was unconditionally adding
${native_tgt}CodeGen even though it conditionally added ${native_tgt}Info and
${native_tgt}Desc. This has been fixed by making ${native_tgt}CodeGen
conditional too
* The 'all' pseudo-component was causing library names like LLVMLLVMDemangle as
the expansion was to a library name and not a component. There doesn't seem to
be a list of available components anywhere so this has been fixed by moving the
expansion of 'all' back where it was before. This manifested in different ways
on different builders but it was the same root cause
Daniel Sanders [Fri, 19 Jul 2019 21:11:05 +0000 (21:11 +0000)]
Revert r366610 and r366612: Expand pseudo-components before embedding in llvm-config
Some targets are missing LLVMDemangle, one is adding the LLVM prefix twice, and two
are hitting the very error this patch fixes for my target. Reverting while I work
through the reports.
Daniel Sanders [Fri, 19 Jul 2019 20:58:11 +0000 (20:58 +0000)]
Fix a latent bug discovered by r366610: nativecodegen includes X86CodeGen when X86 is not compiled
I believe this to have been a latent bug as the same expansion checks for the
existence of ${native_tgt}Info and ${native_tgt}Desc and only adds them if
they were compiled but unconditionally adds ${native_tgt}CodeGen.
This should fix llvm-clang-x86_64-win-fast which builds ARM only on an X86 host and similar builders.
Daniel Sanders [Fri, 19 Jul 2019 20:38:05 +0000 (20:38 +0000)]
Expand pseudo-components before embedding in llvm-config
Summary:
If you use pseudo-targets like AllTargetsCodeGens in LLVM_DYLIB_COMPONENTS
then a test will fail because `./bin/llvm-config --shared-mode` can't
handle these targets. We can fix this by expanding them before embedding
the string into llvm-config
Matt Arsenault [Fri, 19 Jul 2019 19:32:00 +0000 (19:32 +0000)]
LiveIntervals: Fix handleMove asserting on BUNDLE
The top-level BUNDLE instruction should behave as an ordinary
instruction. It is supposed to have all relevant registers as implicit
operands. Moving it should work as any other instruction. I believe
the assert intended to avoid moving instructions inside bundles.
This change reverts most of the previous register name generation.
The real problem is that RegisterTuple does not generate asm names.
Added optional operand to RegisterTuple. This way we can simplify
register name access and dramatically reduce the size of static
tables for the backend.
Louis Dionne [Fri, 19 Jul 2019 17:02:42 +0000 (17:02 +0000)]
[libc++] Integrate the PSTL into libc++
Summary:
This commit allows specifying LIBCXX_ENABLE_PARALLEL_ALGORITHMS when
configuring libc++ in CMake. When that option is enabled, libc++ will
assume that the PSTL can be found somewhere on the CMake module path,
and it will provide the C++17 parallel algorithms based on the PSTL
(that is assumed to be available).
The commit also adds support for running the PSTL tests as part of
the libc++ test suite.
Matt Arsenault [Fri, 19 Jul 2019 14:15:18 +0000 (14:15 +0000)]
AMDGPU/GlobalISel: Rewrite lowerFormalArguments
This should now handle everything except structs passed as multiple
registers.
I think most of the packing logic should be handled by
handleAssignments, but I'm unclear on what the contract is for
multiple registers. This is copying how x86 handles this.
This does change the behavior of the test_sgpr_alignment0 amdgpu_vs
test. I don't think shader arguments should try to follow the
alignment, and registers need to be repacked. I also don't think it
matters, since I think the pointers are packed to the beginning of the
argument list anyway.
Matt Arsenault [Fri, 19 Jul 2019 13:36:46 +0000 (13:36 +0000)]
DAG: Handle dbg_value for arguments split into multiple subregs
This was handled previously for arguments split due to not fitting in
an MVT. This was dropping the register for argument registers split
due to TLI::getRegisterTypeForCallingConv.
Kai Luo [Fri, 19 Jul 2019 12:58:16 +0000 (12:58 +0000)]
[MachineCSE][MachinePRE] Avoid hoisting code from code regions into hot BBs.
Summary:
Current PRE hoists common computations into
CMBB = DT->findNearestCommonDominator(MBB, MBB1).
However, if CMBB is in a hot loop body, we might get performance
degradation.
Roman Lebedev [Fri, 19 Jul 2019 11:29:04 +0000 (11:29 +0000)]
[NFC][InstCombine] Redundant masking before left-shift: tests with assume
If the legality check is `(shiftNbits-maskNbits) s>= 0`,
then we can simplify it to `shiftNbits u>= maskNbits`,
which is easier to check for.
However, currently switching the `dropRedundantMaskingOfLeftShiftInput()`
to `SimplifyICmpInst()` does not catch these cases and regresses
currently-handled cases, so i'll leave it as is for now.
Oliver Stannard [Fri, 19 Jul 2019 10:37:37 +0000 (10:37 +0000)]
Don't update NoTrappingFPMath and FPDenormalMode in resetTargetOptions
We'd like to remove this whole function, because these are properties of
functions, not the target as a whole. These two are easy to remove
because they are only used for emitting ARM build attributes, which
expects them to represent the defaults for the whole module, not just
the last function generated.
This is needed to get correct build attributes when using IPRA on ARM,
because IPRA causes resetTargetOptions to get called before
ARMAsmPrinter::emitAttributes.
[ARM] Add <saturate> operand to SQRSHRL and UQRSHLL
Summary:
According to the new Armv8-M specification
https://static.docs.arm.com/ddi0553/bh/DDI0553B_h_armv8m_arm.pdf the
instructions SQRSHRL and UQRSHLL now have an additional immediate
operand <saturate>. The new assembly syntax is:
where <saturate> can be either 64 (the existing behavior) or 48, in
that case the result is saturated to 48 bits.
The new operand is encoded as follows:
#64 Encoded as sat = 0
#48 Encoded as sat = 1
sat is bit 7 of the instruction bit pattern.
This patch adds a new assembler operand class MveSaturateOperand which
implements parsing and encoding. Decoding is implemented in
DecodeMVEOverlappingLongShift.
[sanitizers] Use covering ObjectFormatType switches
Summary:
This patch removes the `default` case from some switches on
`llvm::Triple::ObjectFormatType`, and cases for the missing enumerators
(`UnknownObjectFormat`, `Wasm`, and `XCOFF`) are then added.
For `UnknownObjectFormat`, the effect of the action for the `default`
case is maintained; otherwise, where `llvm_unreachable` is called,
`report_fatal_error` is used instead.
Where the `default` case returns a default value, `report_fatal_error`
is used for XCOFF as a placeholder. For `Wasm`, the effect of the action
for the `default` case in maintained.
The code is structured to avoid strongly implying that the `Wasm` case
is present for any reason other than to make the switch cover all
`ObjectFormatType` enumerator values.
Jay Foad [Fri, 19 Jul 2019 08:40:37 +0000 (08:40 +0000)]
[AMDGPU] Simplify the exclusive scan used for optimized atomics
Summary:
Change the scan algorithm to use only power-of-two shifts (1, 2, 4, 8,
16, 32) instead of starting off shifting by 1, 2 and 3 and then doing
a 3-way ADD, because:
1. It simplifies the compiler a little.
2. It minimizes vgpr pressure because each instruction is now of the
form vn = vn + vn << c.
3. It is more friendly to the DPP combiner, which currently can't
combine into an ADD3 instruction.
Because of #2 and #3 the end result is improved from this:
Roman Lebedev [Fri, 19 Jul 2019 08:26:58 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [5/5] (PR42563)
Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.
There are many variants to this pattern:
f. `((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
f. `(ShiftShAmt-MaskShAmt) s>= 0` (i.e. `ShiftShAmt u>= MaskShAmt`)
Normally, the inner pattern is sign-extend,
but for our purposes it's no different to other patterns:
alive proofs:
f: https://rise4fun.com/Alive/7U3
For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.
Roman Lebedev [Fri, 19 Jul 2019 08:26:47 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [4/5] (PR42563)
Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.
There are many variants to this pattern:
e. `((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
e. `(ShiftShAmt-MaskShAmt) s>= 0` (i.e. `ShiftShAmt u>= MaskShAmt`)
alive proofs:
e: https://rise4fun.com/Alive/0FT
For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.
Roman Lebedev [Fri, 19 Jul 2019 08:26:37 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [3/5] (PR42563)
Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.
There are many variants to this pattern:
d. `(x & ((-1 << MaskShAmt) >> MaskShAmt)) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
d. `(ShiftShAmt-MaskShAmt) s>= 0` (i.e. `ShiftShAmt u>= MaskShAmt`)
alive proofs:
d: https://rise4fun.com/Alive/I5Y
For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.
Roman Lebedev [Fri, 19 Jul 2019 08:26:25 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [2/5] (PR42563)
Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.
There are many variants to this pattern:
c. `(x & (-1 >> MaskShAmt)) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
c. `(ShiftShAmt-MaskShAmt) s>= 0` (i.e. `ShiftShAmt u>= MaskShAmt`)
alive proofs:
c: https://rise4fun.com/Alive/RgJh
For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.
Roman Lebedev [Fri, 19 Jul 2019 08:26:13 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [1/5] (PR42563)
Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.
There are many variants to this pattern:
b. `(x & (~(-1 << maskNbits))) << shiftNbits`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
b. `(MaskShAmt+ShiftShAmt) u>= bitwidth(x)`
alive proof:
b: https://rise4fun.com/Alive/y8M
For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.
Roman Lebedev [Fri, 19 Jul 2019 08:25:43 +0000 (08:25 +0000)]
[InstCombine] Dropping redundant masking before left-shift [0/5] (PR42563)
Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.
There are many variants to this pattern:
a. `(x & ((1 << MaskShAmt) - 1)) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
a. `(MaskShAmt+ShiftShAmt) u>= bitwidth(x)`
alive proof:
a: https://rise4fun.com/Alive/wi9
Indeed, not all of these patterns are canonical.
But since this fold will only produce a single instruction
i'm really interested in handling even uncanonical patterns,
since i have this general kind of pattern in hotpaths,
and it is not totally outlandish for bit-twiddling code.
For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.
Hsiangkai Wang [Fri, 19 Jul 2019 06:10:36 +0000 (06:10 +0000)]
[DebugInfo] Some fields do not need relocations even relax is enabled.
In debug frame information, some fields, e.g., Length in CIE/FDE and
Offset in FDE are attributes to describe the structure of CIE/FDE. They
are not related to the relaxed code. However, these attributes are
symbol differences. So, in current design, these attributes will be
filled as zero and LLVM generates relocations for them.
We only need to generate relocations for symbols in executable sections.
So, if the symbols are not located in executable sections, we still
evaluate their values under relaxation.
Alex Brachet [Fri, 19 Jul 2019 02:31:21 +0000 (02:31 +0000)]
[test] [llvm-objcopy] Fix broken test case
Summary: The test case added in D62718 did not work unless the user was root because write bits were not set for the output file. This change uses only permissions with user write (0200) to ensure tests pass regardless of the users permissions.
Bill Wendling [Fri, 19 Jul 2019 01:10:28 +0000 (01:10 +0000)]
Use the MachineBasicBlock symbol for a callbr target
Summary:
Inline asm doesn't use labels when compiled as an object file. Therefore, we
shouldn't create one for the (potential) callbr destination. Instead, use the
symbol for the MachineBasicBlock.
[GlobalISel] Translate calls to memcpy et al to G_INTRINSIC_W_SIDE_EFFECTs and legalize later.
I plan on adding memcpy optimizations in the GlobalISel pipeline, but we can't
do that unless we delay lowering to actual function calls. This patch changes
the translator to generate G_INTRINSIC_W_SIDE_EFFECTS for these functions, and
then have each target specify that using the new custom legalizer for intrinsics
hook that they want it expanded it a libcall.
[cmake] Convert the NATIVE llvm build process to be project agnostic
lldb recently added a tablegen tool. In order to properly cross compile
lldb standalone there needs to be a mechanism to generate the native
lldb build, analgous to what's done for the NATIVE llvm build. Thus,
we can simply modify this setup to allow for any project to be used.
This reapplies r366142 with a fix for the failing Windows test.
Original commit message:
Creates universal binary output file from input files. Currently uses
hard coded value for alignment. Want to get the create functionality
approved before implementing the alignment function.
CodeGen: Allow !associated metadata to point to aliases.
This is a small extension of !associated, mostly useful for the implementation
convenience of instrumentation passes that RAUW globals with aliases, such
as LowerTypeTests.
[cmake] Only run llvm-codesign if targetting apple on an apple host
Summary:
Other platforms don't have the capability to perform llvm_codesign
step. If LLVM_CODESIGNING_IDENTITY is set then this chunk of code would
attempt to codesign if the target was Apple. But when cross compiling
to Darwin from Linux, for example, this step would fail. So test if the
host is Apple as well.
projects / llvm / log