Chandler Carruth [Thu, 22 Jan 2015 05:25:13 +0000 (05:25 +0000)]
[PM] Rename InstCombine.h to InstCombineInternal.h in preparation for
creating a non-internal header file for the InstCombine pass.
I thought about calling this InstCombiner.h or in some way more clearly
associating it with the InstCombiner clas that it is primarily defining,
but there are several other utility interfaces defined within this for
InstCombine. If, in the course of refactoring, those end up moving
elsewhere or going away, it might make more sense to make this the
combiner's header alone.
Naturally, this is a bikeshed to a certain degree, so feel free to lobby
for a different shade of paint if this name just doesn't suit you.
Chandler Carruth [Thu, 22 Jan 2015 05:08:12 +0000 (05:08 +0000)]
[canonicalize] Teach InstCombine to canonicalize loads which are only
ever stored to always use a legal integer type if one is available.
Regardless of whether this particular type is good or bad, it ensures we
don't get weird differences in generated code (and resulting
performance) from "equivalent" patterns that happen to end up using
a slightly different type.
After some discussion on llvmdev it seems everyone generally likes this
canonicalization. However, there may be some parts of LLVM that handle
it poorly and need to be fixed. I have at least verified that this
doesn't impede GVN and instcombine's store-to-load forwarding powers in
any obvious cases. Subtle cases are exactly what we need te flush out if
they remain.
Also note that this IR pattern should already be hitting LLVM from Clang
at least because it is exactly the IR which would be produced if you
used memcpy to copy a pointer or floating point between memory instead
of a variable.
ARM: fail less catastrophically on invalid Windows input
Windows supports a restricted set of relocations (compared to ARM ELF). In some
cases, we may end up generating an unsupported relocation. This can occur with
bad input to the assembler in particular (the frontend should never generate
code that cannot be compiled). Generate an error rather than just aborting.
The change in the API is driven by the desire to provide a slightly more helpful
message for debugging purposes.
This code was confusing, since it created a `DIExpressionIterator` from
an invalid start point (although it wasn't wrong: it never actually
iterated). Now that the underlying iterator has `getNumber()`, just use
it directly.
Chris Bieneman [Thu, 22 Jan 2015 01:49:59 +0000 (01:49 +0000)]
Assigning and copying command line option objects shouldn't be allowed.
Summary:
The default copy and assignment operators for these objects probably don't actually do what the clients intend, so they should be deleted.
Places using the assignment operator to set the value of an option should cast to the option's data type first to call into the override for operator=. Places using the copy constructor just need to be changed to not copy (i.e. passing by const reference instead of value).
Sanjoy Das [Thu, 22 Jan 2015 00:48:47 +0000 (00:48 +0000)]
Make ScalarEvolution less aggressive with respect to no-wrap flags.
ScalarEvolution currently lowers a subtraction recurrence to an add
recurrence with the same no-wrap flags as the subtraction. This is
incorrect because `sub nsw X, Y` is not the same as `add nsw X, -Y`
and `sub nuw X, Y` is not the same as `add nuw X, -Y`. This patch
fixes the issue, and adds two test cases demonstrating the bug.
Chandler Carruth [Wed, 21 Jan 2015 23:45:01 +0000 (23:45 +0000)]
[canonicalization] Refactor how we create new stores into a helper
function. This is a bit tidier anyways and will make a subsquent patch
simpler as I want to add another case to this combine.
Added most of the missing integer vector folding patterns for SSE (to SSE42) and AVX1.
The most useful of these are probably the i32/i64 extraction, i8/i16/i32/i64 insertions, zero/sign extension, unsigned saturation subtractions, i64 subtractions and the variable mask blends (pblendvb) - others include CLMUL, SSE42 string comparisons and bit tests.
David Blaikie [Wed, 21 Jan 2015 22:57:29 +0000 (22:57 +0000)]
DebugInfo: Use distinct inlinedAt MDLocations to avoid separate inlined calls being coalesced
When two calls from the same MDLocation are inlined they currently get
treated as one inlined function call (creating difficulty debugging,
duplicate variables, etc).
Clang worked around this by including column information on inline calls
which doesn't address LTO inlining or calls to the same function from
the same line and column (such as through a macro). It also didn't
address ctor and member function calls.
By making the inlinedAt locations distinct, every call site has an
explicitly distinct location that cannot be coalesced with any other
call.
This can produce linearly (2x in the worst case where every call is
inlined and the call instruction has a non-call instruction at the same
location) more debug locations. Any increase beyond that are in cases
where the Clang workaround was insufficient and the new scheme is
creating necessary distinct nodes that were being erroneously coalesced
previously.
After this change to LLVM the incomplete workarounds in Clang. That
should reduce the number of debug locations (in a build without column
info, the default on Darwin, not the default on Linux) by not creating
pseudo-distinct locations for every call to an inline function.
(oh, and I made the inlined-at chain rebuilding iterative instead of
recursive because I was having trouble wrapping my head around it the
way it was - open to discussion on the right design for that function
(including going back to a recursive solution))
Chris Bieneman [Wed, 21 Jan 2015 22:45:52 +0000 (22:45 +0000)]
Adding a new cl::HideUnrelatedOptions API to allow clang to migrate off cl::getRegisteredOptions.
Summary: cl::getRegisteredOptions really exposes some of the innards of how command line parsing is implemented. Exposing new APIs that allow us to disentangle client code from implementation details will allow us to make more extensive changes to command line parsing.
Simon Pilgrim [Wed, 21 Jan 2015 22:44:35 +0000 (22:44 +0000)]
[X86][SSE] Added support for SSE3 lane duplication shuffle instructions
This patch adds shuffle matching for the SSE3 MOVDDUP, MOVSLDUP and MOVSHDUP instructions. The big use of these being that they avoid many single source shuffles from needing to use (pre-AVX) dual source instructions such as SHUFPD/SHUFPS: causing extra moves and preventing load folds.
Adding these instructions uncovered an issue in XFormVExtractWithShuffleIntoLoad which crashed on single operand shuffle instructions (now fixed). It also involved fixing getTargetShuffleMask to correctly identify theses instructions as unary shuffles.
Matthias Braun [Wed, 21 Jan 2015 19:02:30 +0000 (19:02 +0000)]
LiveIntervalAnalysis: Factor out code to update liveness on vreg def removal
This cleans up code and is more in line with the general philosophy of
modifying LiveIntervals through LiveIntervalAnalysis instead of changing
them directly.
This also fixes a case where SplitEditor::removeBackCopies() would miss
the subregister ranges.
Matthias Braun [Wed, 21 Jan 2015 18:50:21 +0000 (18:50 +0000)]
LiveIntervalAnalysis: Factor out code to update liveness on physreg def removal
This cleans up code and is more in line with the general philosophy of
modifying LiveIntervals through LiveIntervalAnalysis instead of changing
them directly.
Matt Arsenault [Wed, 21 Jan 2015 18:18:25 +0000 (18:18 +0000)]
R600/SI: Custom lower fround
This fixes it for SI. It also removes the pattern
used previously for Evergreen for f32. I'm not sure
if the the new R600 output is better or not, but it uses
1 fewer instructions if BFI is available.
Ahmed Bougacha [Wed, 21 Jan 2015 17:07:06 +0000 (17:07 +0000)]
[X86] Declare SSE4.1/AVX2 vector extloads covered by PMOV[SZ]X legal.
Now that we can fully specify extload legality, we can declare them
legal for the PMOVSX/PMOVZX instructions. This for instance enables
a DAGCombine to fire on code such as
(and (<zextload-equivalent> ...), <redundant mask>)
to turn it into:
(zextload ...)
as seen in the testcase changes.
There is one regression, in widen_load-2.ll: we're no longer able
to do store-to-load forwarding with illegal extload memory types.
This will be addressed separately.
Use a smaller pragma unroll threshold to reduce test execution time.
When opt is compiled with AddressSanitizer it takes more than 30 seconds
to unroll the loop in unroll_1M().
Jozef Kolek [Wed, 21 Jan 2015 12:39:30 +0000 (12:39 +0000)]
[mips][microMIPS] MicroMIPS 16-bit unconditional branch instruction B
Implement microMIPS 16-bit unconditional branch instruction B.
Implemented 16-bit microMIPS unconditional instruction has real name B16, and
B is an alias which expands to either B16 or BEQ according to the rules:
b 256 --> b16 256 # R_MICROMIPS_PC10_S1
b 12256 --> beq $zero, $zero, 12256 # R_MICROMIPS_PC16_S1
b label --> beq $zero, $zero, label # R_MICROMIPS_PC16_S1
Chandler Carruth [Wed, 21 Jan 2015 11:38:17 +0000 (11:38 +0000)]
[PM] Refactor the InstCombiner interface to use an external worklist.
Because in its primary function pass the combiner is run repeatedly over
the same function until doing so produces no changes, it is essentially
to not re-allocate the worklist. However, as a utility, the more common
pattern would be to put a limited set of instructions in the worklist
rather than the entire function body. That is also the more likely
pattern when used by the new pass manager.
The result is a very light weight combiner that does the visiting with
a separable worklist. This can then be wrapped up in a helper function
for users that want a combiner utility, or as I have here it can be
wrapped up in a pass which manages the iterations used when combining an
entire function's instructions.
Hopefully this removes some of the worst of the interface warts that
became apparant with the last patch here. However, there is clearly more
work. I've again left some FIXMEs for the most egregious. The ones that
stick out to me are the exposure of the worklist and IR builder as
public members, and the use of pointers rather than references. However,
fixing these is likely to be much more mechanical and less interesting
so I didn't want to touch them in this patch.
Chandler Carruth [Wed, 21 Jan 2015 11:23:40 +0000 (11:23 +0000)]
[PM] Simplify (ha! ha!) the way that instcombine calls the
SimplifyLibCalls utility by sinking it into the specific call part of
the combiner.
This will avoid us needing to do any contortions to build this object in
a subsequent refactoring I'm doing and seems generally better factored.
We don't need this utility everywhere and it carries no interesting
state so we might as well build it on demand.
Vladimir Medic [Wed, 21 Jan 2015 10:47:36 +0000 (10:47 +0000)]
[Mips][Disassembler]When disassembler meets load/store from coprocessor 2 instructions for mips r6 it crashes as the access to operands array is out of range. This patch adds dedicated decoder method that properly handles decoding of these instructions.
Chandler Carruth [Wed, 21 Jan 2015 02:11:59 +0000 (02:11 +0000)]
[PM] Replace an abuse of inheritance to override a single function with
a more direct approach: a type-erased glorified function pointer. Now we
can pass a function pointer into this for the easy case and we can even
pass a lambda into it in the interesting case in the instruction
combiner.
I'll be using this shortly to simplify the interfaces to InstCombiner,
but this helps pave the way and seems like a better design for the
libcall simplifier utility.
Simon Pilgrim [Wed, 21 Jan 2015 00:02:13 +0000 (00:02 +0000)]
[X86][AVX] Simplified diff between AVX1 and SSE42 fp stack folding tests. NFC.
Changed the AVX1 tests register spill tail call to return a xmm like the SSE42 version - makes doing diffs between them a lot easier without affecting the spills themselves.
The SSE42 version of the AVX1 float stack folding tests will be added shortly, this renames the AVX1 file so that the files will be near each other in a directory listing to help ensure they are kept in sync.
Chandler Carruth [Tue, 20 Jan 2015 22:44:35 +0000 (22:44 +0000)]
[PM] Separate the InstCombiner from its pass.
This creates a small internal pass which runs the InstCombiner over
a function. This is the hard part of porting InstCombine to the new pass
manager, as at this point none of the code in InstCombine has access to
a Pass object any longer.
The resulting interface for the InstCombiner is pretty terrible. I'm not
planning on leaving it that way. The key thing missing is that we need
to separate the worklist from the combiner a touch more. Once that's
done, it should be possible for *any* part of LLVM to just create
a worklist with instructions, populate it, and then combine it until
empty. The pass will just be the (obvious and important) special case of
doing that for an entire function body.
For now, this is the first increment of factoring to make all of this
work.
Chandler Carruth [Tue, 20 Jan 2015 21:10:35 +0000 (21:10 +0000)]
[PM] Reformat this code with clang-format so that subsequent changes
don't get muddied up by formatting changes.
Some of these don't really seem like improvements to me, but they also
don't seem any worse and I care much more about not formatting them
manually than I do about the particular formatting. =]
Daniel Jasper [Tue, 20 Jan 2015 19:43:33 +0000 (19:43 +0000)]
Prevent binary-tree deterioration in sparse switch statements.
This addresses part of llvm.org/PR22262. Specifically, it prevents
considering the densities of sub-ranges that have fewer than
TLI.getMinimumJumpTableEntries() elements. Those densities won't help
jump tables.
This is not a complete solution but works around the most pressing
issue.
[GC] Verify-pass void vararg functions in gc.statepoint
With the appropriate Verifier changes, exactracting the result out of a
statepoint wrapping a vararg function crashes. However, a void vararg
function works fine: commit this first step.
Adrian Prantl [Tue, 20 Jan 2015 19:42:22 +0000 (19:42 +0000)]
Reapply: Teach SROA how to update debug info for fragmented variables.
This reapplies r225379.
ChangeLog:
- The assertion that this commit previously ran into about the inability
to handle indirect variables has since been removed and the backend
can handle this now.
- Testcases were upgrade to the new MDLocation format.
- Instead of keeping a DebugDeclares map, we now use
llvm::FindAllocaDbgDeclare().
Original commit message follows.
Debug info: Teach SROA how to update debug info for fragmented variables.
This allows us to generate debug info for extremely advanced code such as
Manman Ren [Tue, 20 Jan 2015 19:24:59 +0000 (19:24 +0000)]
[llvm link] Destroy ConstantArrays in LLVMContext if they are not used.
ConstantArrays constructed during linking can cause quadratic memory
explosion. An example is the ConstantArrays constructed when linking in
GlobalVariables with appending linkage.
Releasing all unused constants can cause a 20% LTO compile-time
slowdown for a large application. So this commit releases unused ConstantArrays
only.
rdar://19040716. It reduces memory footprint from 20+G to 6+G.
Tom Stellard [Tue, 20 Jan 2015 17:49:47 +0000 (17:49 +0000)]
R600/SI: Use external symbols for scratch buffer
We were passing the scratch buffer address to the shaders via user sgprs,
but now we use external symbols and have the driver patch the shader
using reloc information.
Tom Stellard [Tue, 20 Jan 2015 17:49:43 +0000 (17:49 +0000)]
R600/SI: Don't store scratch buffer frame index in MUBUF offset field
We don't have a good way of legalizing this if the frame index offset
is more than the 12-bits, which is size of MUBUF's offset field, so
now we store the frame index in the vaddr field.
Jozef Kolek [Tue, 20 Jan 2015 16:45:27 +0000 (16:45 +0000)]
[mips][microMIPS] MicroMIPS 16-bit unconditional branch instruction B
Implement microMIPS 16-bit unconditional branch instruction B.
Implemented 16-bit microMIPS unconditional instruction has real name B16, and
B is an alias which expands to either B16 or BEQ according to the rules:
b 256 --> b16 256 # R_MICROMIPS_PC10_S1
b 12256 --> beq $zero, $zero, 12256 # R_MICROMIPS_PC16_S1
b label --> beq $zero, $zero, label # R_MICROMIPS_PC16_S1
Evgeniy Stepanov [Tue, 20 Jan 2015 15:21:35 +0000 (15:21 +0000)]
[msan] Optimize -msan-check-constant-shadow.
The new code does not create new basic blocks in the case when shadow is a
compile-time constant; it generates either an unconditional __msan_warning
call or nothing instead.
Chandler Carruth [Tue, 20 Jan 2015 10:58:50 +0000 (10:58 +0000)]
[PM] Port LoopInfo to the new pass manager, adding both a LoopAnalysis
pass and a LoopPrinterPass with the expected associated wiring.
I've added a RUN line to the only test case (!!!) we have that actually
prints loops. Everything seems to be working.
This is somewhat exciting as this is the first analysis using another
analysis to go in for the new pass manager. =D I also believe it is the
last analysis necessary for porting instcombine, but of course I may yet
discover more.
projects / llvm / log