[X86][LLVM] Converting __mm{|256|512}_movm_epi{8|16|32|64} LLVMIR call into generic intrinsics.
This patch is a part one of two reviews, one for the clang and the other for LLVM.
The patch deletes the back-end intrinsics and adds support for them in the auto upgrade.
Daniel Sanders [Tue, 4 Apr 2017 13:25:23 +0000 (13:25 +0000)]
[tablegen][globalisel] Add support for nested instruction matching.
Summary:
Lift the restrictions that prevented the tree walking introduced in the
previous change and add support for patterns like:
(G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3
Also adds support for G_SEXT and G_ZEXT to support these cases.
One particular aspect of this that I should draw attention to is that I've
tried to be overly conservative in determining the safety of matches that
involve non-adjacent instructions and multiple basic blocks. This is intended
to be used as a cheap initial check and we may add a more expensive check in
the future. The current rules are:
* Reject if any instruction may load/store (we'd need to check for intervening
memory operations.
* Reject if any instruction has implicit operands.
* Reject if any instruction has unmodelled side-effects.
See isObviouslySafeToFold().
PSADBW pattern currently supports the 32 bit IR pattern and only GLT (greather than) comparison.
The patch extends the pattern to catch also 64 bit IR pattern and includes all other comparison types (not only GLT).
Summary:
The TypeTableBuilder provides stable storage for type records. We don't
need to copy all of the bytes into a flat vector before adding it to the
TpiStreamBuilder.
This makes addTypeRecord take an ArrayRef<uint8_t> and a hash code to go
with it, which seems like a simplification.
[codeview] Cope with unsorted streams in type merging
Summary:
MASM can produce type streams that are not topologically sorted. It can
even produce type streams with circular references, but those are not
common in practice.
Summary:
Add a hook for simplification of shufflevector's with the following rules:
- Constant folding - NFC, as it was already being done by the default handler.
- If only one of the operands is constant, constant fold the shuffle if the
mask does not select elements from the variable operand - to show the hook is firing and affecting the test-cases.
Simon Pilgrim [Mon, 3 Apr 2017 21:06:51 +0000 (21:06 +0000)]
[X86][SSE]] Lower BUILD_VECTOR with repeated elts as BUILD_VECTOR + VECTOR_SHUFFLE
It can be costly to transfer from the gprs to the xmm registers and can prevent loads merging.
This patch splits vXi16/vXi32/vXi64 BUILD_VECTORS that use the same operand in multiple elements into a BUILD_VECTOR with only a single insertion of each of those elements and then performs an unary shuffle to duplicate the values.
There are a couple of minor regressions this patch unearths due to some missing MOVDDUP/BROADCAST folds that I will address in a future patch.
Note: Now that vector shuffle lowering and combining is pretty good we should be reusing that instead of duplicating so much in LowerBUILD_VECTOR - this is the first of several patches to address this.
[InstCombine] Remove canonicalization for (X & C1) | C2 --> (X | C2) & (C1|C2) when C1 & C2 have common bits.
It turns out that SimplifyDemandedInstructionBits will get called earlier and remove bits from C1 first. Effectively doing (X & (C1&C2)) | C2. So by the time it got to this check there could be no common bits.
I think the DAGCombiner has the same check but its check can be executed because it handles demanded bits later. I'll look at it next.
x86 interrupt calling convention: re-align stack pointer on 64-bit if an error code was pushed
The x86_64 ABI requires that the stack is 16 byte aligned on function calls. Thus, the 8-byte error code, which is pushed by the CPU for certain exceptions, leads to a misaligned stack. This results in bugs such as Bug 26413, where misaligned movaps instructions are generated.
This commit fixes the misalignment by adjusting the stack pointer in these cases. The adjustment is done at the beginning of the prologue generation by subtracting another 8 bytes from the stack pointer. These additional bytes are popped again in the function epilogue.
Summary:
Move the aarch64-type-promotion pass within the existing type promotion framework in CGP.
This change also support forking sexts when a new sext is required for promotion.
Note that change is based on D27853 and I am submitting this out early to provide a better idea on D27853.
[Support] Make printAllJSONValues public, for custom output.
Summary:
This changes the static method TimerGroup::printAllJSONValues from private to
public, to match the static method TimerGroup::printAll. When trying to drive
the reporting machinery by hand, the existing API is _almost_ flexible enough,
but this entrypoint is required to intermix printing timers with other
non-timer output.
The underlying motive here is a Swift change to consolidate the collection of
timers, LLVM statistics and other (non-assert-dependent) counters into JSON
files, which requires a bit of manual intervention in LLVM's stat and timer
output routines. See https://github.com/apple/swift/pull/8477 for details.
Bitcode: Remove reader support for MODULE_CODE_PURGEVALS.
Support for writing this module code was removed in r73220, which was well
before the LLVM 3.0 release, so we do not need to be able to understand it
for backwards compatibility.
Revert "[DAGCombine] A shuffle of a splat is always the splat itself"
This reverts commit r299047 which is incorrect because the
simplification may result in incorrect propogation of undefs to users of
the folded shuffle.
[InstCombine] fix formatting for foldLogOpOfMaskedICmps and related bits; NFCI
1. Improve enum, function, and variable names.
2. Improve comments.
3. Fix variable capitalization.
4. Run clang-format.
As an existing code comment suggests, this should work with vector types / splat constants too,
so making this look right first will reduce the diffs needed for that change.
[APInt] Move isMask and isShiftedMask out of APIntOps and into the APInt class. Implement them without memory allocation for multiword
This moves the isMask and isShiftedMask functions to be class methods. They now use the MathExtras.h function for single word size and leading/trailing zeros/ones or countPopulation for the multiword size. The previous implementation made multiple temorary memory allocations to do the bitwise arithmetic operations to match the MathExtras.h implementation.
- we are now using immediate AsmOperands so that the range check functions are
tablegen'ed.
- Big bonus is that error messages become much more accurate, i.e. instead of a
useless "invalid operand" error message it will not say that the immediate
operand must in range [x,y], which is why regression tests needed updating.
More tablegen operand descriptions could probably benefit from using
immediateAsmOperand, but this is a first good step to get rid of most of the
nearly identical range check functions. I will address the remaining immediate
operands in next clean ups.
[APInt] Add a public typedef for the internal type of APInt use it instead of integerPart. Make APINT_BITS_PER_WORD and APINT_WORD_SIZE public.
This patch is one step to attempt to unify the main APInt interface and the tc functions used by APFloat.
This patch adds a WordType to APInt and uses that in all the tc functions. I've added temporary typedefs to APFloat to alias it to integerPart to keep the patch size down. I'll work on removing that in a future patch.
In future patches I hope to reuse the tc functions to implement some of the main APInt functionality.
I may remove APINT_ from BITS_PER_WORD and WORD_SIZE constants so that we don't have the repetitive APInt::APINT_ externally.
[InstCombine] Make InstCombiner::OptAndOp take a BinaryOperator instead of an Instruction.
The callers have already performed the necessary cast before calling. This allows us to remove a comment that says the instruction must be a BinaryOperator and make it explicit in the argument type.
Had to add a default case to the switch because BinaryOperator::getOpcode() returns a BinaryOps enum.
[InstCombine] Remove redundant combine from visitAnd
As far as I can tell this combine is fully handled by SimplifyDemandedInstructionBits.
I was only looking at this because it is the only user of APIntOps::isShiftedMask which is itself broken. As demonstrated by r299187. I was going to fix isShiftedMask and needed to make sure we had coverage for the new cases it would expose to this combine. But looks like we can nuke it instead.
Daniel Berlin [Sun, 2 Apr 2017 13:23:44 +0000 (13:23 +0000)]
NewGVN: Handle coercion of constant stores, loads, memory insts.
Summary:
Depends on D30928.
This adds support for coercion of stores and memory instructions that do not require insertion to process.
Another few tests down.
I added the relevant tests from rle.ll
[BypassSlowDivision] Do not bypass division of hash-like values
Disable bypassing if one of the operands looks like a hash value. Slow
division often occurs in hashtable implementations and fast division is
never taken there because a hash value is extremely unlikely to have
enough upper bits set to zero.
A value is considered to be hash-like if it is produced by
1) XOR operation
2) Multiplication by a constant wider than the shorter type
3) PHI node with all incoming values being hash-like
[APInt] Simplify some code by using operator+=(uint64_t) instead of doing a more complex assignment into a temporary APInt just to use the APInt operator+=.
Daniel Berlin [Sun, 2 Apr 2017 05:09:15 +0000 (05:09 +0000)]
MemorySSA: Add support for caching clobbering access in stores
Summary:
This enables us to cache the clobbering access for stores, despite the
fact that we can't rewrite the use-def chains themselves.
Early testing shows that, after this change, for larger testcases, it will be a significant net positive (memory and time) to remove the walker caching.
[APInt] Allow GreatestCommonDivisor to take rvalue inputs efficiently. Use moves instead of copies in the loop.
Summary:
GreatestComonDivisor currently makes a copy of both its inputs. Then in the loop we do one move and two copies, plus any allocation the urem call does.
This patch changes it to take its inputs by value so that we can do a move of any rvalue inputs instead of copying. Then in the loop we do 3 move assignments and no copies. This way the only possible allocations we have in the loop is from the urem call.
[DAGCombiner] enable vector transforms for any/all {sign} bits set/clear
The code already allowed vector types in via "isInteger" (which might want
a more specific name), so use splat-friendly constant predicates to match
those types.
Daniel Berlin [Sat, 1 Apr 2017 09:44:33 +0000 (09:44 +0000)]
NewGVN: Don't try to kill off the stored value of stores when
processing the congruence class of the store.
Because we use the stored value of a store as the def, it isn't dead
just because it appears as a def when it comes from a store.
Note: I have not hit any cases with the memory code as it is where
this breaks anything, just because of what memory congruences we
actually allow. In a followup that improves memory congruence,
this bug actually breaks real stuff (but the verifier catches it).
[DAGCombiner] Fix fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask) to explicitly ensure that only one of the inputs of each shuffle is a zero vector.
This can only happen when we have a mix of zero and undef elements and the two vectors have a different arrangement of zeros/undefs. The shuffle should eventually be constant folded to all zeros.
[RegBankSelect] Support REG_SEQUENCE for generic mapping
REG_SEQUENCE falls into the same category as COPY for operands mapping:
- They don't have MCInstrDesc with register constraints
- The input variable could use whatever register classes
- It is possible to have register class already assigned to the operands
In particular, given REG_SEQUENCE are always target specific because of
the subreg indices. Those indices must apply to the register class of
the definition of the REG_SEQUENCE and therefore, the target must set a
register class to that definition. As a result, the generic code can
always use that register class to derive a valid mapping for a
REG_SEQUENCE.
Feature generic option to setup start/stop-after/before
This patch refactors the code used in llc such that all the users of the
addPassesToEmitFile API have access to a homogeneous way of handling
start/stop-after/before options right out of the box.
Previously each user would have needed to duplicate this logic and set
up its own options.
Craig Topper [Fri, 31 Mar 2017 22:23:42 +0000 (22:23 +0000)]
[APInt] Fix bugs in isShiftedMask to match behavior of the similar function in MathExtras.h
This removes a parameter from the routine that was responsible for a lot of the issue. It was a bit count that had to be set to the BitWidth of the APInt and would get passed to getLowBitsSet. This guaranteed the call to getLowBitsSet would create an all ones value. This was then compared to (V | (V-1)). So the only shifted masks we detected had to have the MSB set.
The one in tree user is a transform in InstCombine that never fires due to earlier transforms covering the case better. I've submitted a patch to remove it completely, but for now I've just adapted it to the new interface for isShiftedMask.
projects / llvm / log