Currently, we sometimes miscompile this vector pattern:
(c ? -v : v)
We lower it to (because "c" is <4 x i1>, lowered as a vector mask):
(~c & v) | (c & -v)
When we have SSSE3, we incorrectly lower that to PSIGN, which does:
(c < 0 ? -v : c > 0 ? v : 0)
in other words, when c is either all-ones or all-zero:
(c ? -v : 0)
While this is an old bug, it rarely triggers because the PSIGN combine
is too sensitive to operand order. This will be improved separately.
Note that the PSIGN tests are also incorrect. Consider:
%b.lobit = ashr <4 x i32> %b, <i32 31, i32 31, i32 31, i32 31>
%sub = sub nsw <4 x i32> zeroinitializer, %a
%0 = xor <4 x i32> %b.lobit, <i32 -1, i32 -1, i32 -1, i32 -1>
%1 = and <4 x i32> %a, %0
%2 = and <4 x i32> %b.lobit, %sub
%cond = or <4 x i32> %1, %2
ret <4 x i32> %cond
if %b is zero:
%b.lobit = <4 x i32> zeroinitializer
%sub = sub nsw <4 x i32> zeroinitializer, %a
%0 = <4 x i32> <i32 -1, i32 -1, i32 -1, i32 -1>
%1 = <4 x i32> %a
%2 = <4 x i32> zeroinitializer
%cond = or <4 x i32> %a, zeroinitializer
ret <4 x i32> %a
whereas we currently generate:
psignd %xmm1, %xmm0
retq
which returns 0, as %xmm1 is 0.
Instead, use a pure logic sequence, as described in:
https://graphics.stanford.edu/~seander/bithacks.html#ConditionalNegate
Fixes PR26110.
Differential Revision: http://reviews.llvm.org/D17181
llvm-svn: 261023
We're going to stop generating PSIGN, so calling a test "psign"
isn't ideal. Instead, call these tests what they really are:
variable blends using logic.
Also add a test to exhibit a case we're currently missing in
the PSIGN combine.
llvm-svn: 261022
The register stackifier currently checks for intervening stores (and
loads that may alias them) but doesn't account for the fact that the
instruction being moved may affect intervening loads.
Differential Revision: http://reviews.llvm.org/D17298
llvm-svn: 261014
Summary:
This change will add a pass to remove unnecessary zero copies in target blocks
of cbz/cbnz instructions. E.g., the copy instruction in the code below can be
removed because the cbz jumps to BB1 when x0 is zero :
BB0:
cbz x0, .BB1
BB1:
mov x0, xzr
Jun
Reviewers: gberry, jmolloy, HaoLiu, MatzeB, mcrosier
Subscribers: mcrosier, mssimpso, haicheng, bmakam, llvm-commits, aemerson, rengolin
Differential Revision: http://reviews.llvm.org/D16203
llvm-svn: 261004
CopyToReg nodes don't support FrameIndex operands. Other targets select
the FI to some LEA-like instruction, but since we don't have that, we
need to insert some kind of instruction that can take an FI operand and
produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy
copy_local between Op and its FI operand. This results in a redundant
copy which we should optimize away later (maybe in the post-FI-lowering
peephole pass).
Differential Revision: http://reviews.llvm.org/D17213
llvm-svn: 260987
WebAssembly doesn't require full RPO; topological sorting is sufficient and
can preserve more of the MachineBlockPlacement ordering. Unfortunately, this
still depends a lot on heuristics, because while we use the
MachineBlockPlacement ordering as a guide, we can't use it in places where
it isn't topologically ordered. This area will require further attention.
llvm-svn: 260978
This avoids some complications updating LiveIntervals to be aware of the new
register lifetimes, because we can just compute new intervals from scratch
rather than describe how the old ones have been changed.
llvm-svn: 260971
If KMOVB not supported (require AVX512DQ) only KMOVW can be used so store size should be 2 bytes.
Differential Revision: http://reviews.llvm.org/D17138
llvm-svn: 260878
This patch attempts to represent a shuffle as a repeating shuffle (recognisable by is128BitLaneRepeatedShuffleMask) with the source input(s) in their original lanes, followed by a single permutation of the 128-bit lanes to their final destinations.
On AVX2 we can additionally attempt to match using 64-bit sub-lane permutation. AVX2 can also now match a similar 'broadcasted' repeating shuffle.
This patch has several benefits:
* Avoids prematurely matching with lowerVectorShuffleByMerging128BitLanes which can require both inputs to have their input lanes permuted before shuffling.
* Can replace PERMPS/PERMD instructions - although these are useful for cross-lane unary shuffling, they require their shuffle mask to be pre-loaded (and increase register pressure).
* Matching the repeating shuffle makes use of a lot of existing shuffle lowering.
There is an outstanding minor AVX1 regression (combine_unneeded_subvector1 in vector-shuffle-combining.ll) of a previously 128-bit shuffle + subvector splat being converted to a subvector splat + (2 instruction) 256-bit shuffle, I intend to fix this in a followup patch for review.
Differential Revision: http://reviews.llvm.org/D16537
llvm-svn: 260834
As shown in:
https://llvm.org/bugs/show_bug.cgi?id=23203
...we currently die because lowering believes that mfence is allowed without SSE2 on x86-64,
but the instruction def doesn't know that.
I don't know if allowing mfence without SSE is right, but if not, at least now it's consistently wrong. :)
Differential Revision: http://reviews.llvm.org/D17219
llvm-svn: 260828
Tests for the new scalarize all private access options will be
included with a future commit.
The only functional change is to make the split/scalarize behavior
for private access of > 4 element vectors to be consistent
with the flat/global handling. This makes the spilling worse
in the two changed tests.
llvm-svn: 260804
Summary:
This patch skips DAG combine of fp_round (fp_round x) if it results in
an fp_round from f80 to f16.
fp_round from f80 to f16 always generates an expensive (and as yet,
unimplemented) libcall to __truncxfhf2. This prevents selection of
native f16 conversion instructions from f32 or f64. Moreover, the first
(value-preserving) fp_round from f80 to either f32 or f64 may become a
NOP in platforms like x86.
Reviewers: ab
Subscribers: srhines, llvm-commits
Differential Revision: http://reviews.llvm.org/D17221
llvm-svn: 260769
Replace spills to memory with spills to registers, if possible. This
applies mostly to predicate registers (both scalar and vector), since
they are very limited in number. A spill of a predicate register may
happen even if there is a general-purpose register available. In cases
like this the stack spill/reload may be eliminated completely.
This optimization will consider all stack objects, regardless of where
they came from and try to match the live range of the stack slot with
a dead range of a register from an appropriate register class.
llvm-svn: 260758
Rewrite the code to handle all pseudo-instructions in a single pass.
This temporarily reverts spill slot optimization that used general-
purpose registers to hold values of spilled predicate registers.
llvm-svn: 260696
Historically, AMD internal sp3 assembler has flat_store* addr, data
format. To match existing code and to enable reuse, change LLVM
definitions to match. Also update MC and CodeGen tests.
Differential Revision: http://reviews.llvm.org/D16927
Patch by: Nikolay Haustov
llvm-svn: 260694
Summary:
It is possible that the loop condition can be a boolean constant (infinite loop,
for example). So we sould handle constant condition in annotating a loop. This
patch adds this functionality to support annotating constant condition.
Reviewers: tstellarAMD, arsenm
Subscribers: llvm-commits, arsenm
Differential Revision: http://reviews.llvm.org/D15093
llvm-svn: 260692
We can generate the actual instructions from the intrinsics without the
need for pseudo-instructions. Also, since the intrinsics have a side-
effect in a form of a store, attempt to optimize away loads from the
store location.
llvm-svn: 260690
Summary:
Before this change, callee-save registers would be rounded up to even
pairs of GPRs and FPRs. This change eliminates these extra padding
load/stores, though it does keep the stack allocation the same size
unless both the GPR and FPR sets have an odd size, in which case one
full pair stack slot (16 bytes) is saved.
This optimization cannot currently be done for MachO targets since they
rely on a fast-path .debug_frame equivalent that can only encode
callee-save registers as pairs.
Reviewers: t.p.northover, rengolin, mcrosier, jmolloy
Subscribers: aemerson, rengolin, mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D17000
llvm-svn: 260689
The DataLayout can calculate alignment of vectors based on the alignment
of the element type and the number of elements. In fact, it is the product
of these two values. The problem is that for vectors of N x i1, this will
return the alignment of N bytes, since the alignment of i1 is 8 bits. The
vector types of vNi1 should be aligned to N bits instead. Provide explicit
alignment for HVX vectors to avoid such complications.
llvm-svn: 260678
This was hardcoded to the static private size, but this
would be missing the offset and additional size for someday
when we have dynamic sizing.
Also stops always initializing flat_scratch even when unused.
In the future we should stop emitting this unless flat instructions
are used to access private memory. For example this will initialize
it almost always on VI because flat is used for global access.
llvm-svn: 260658
Introduce a subtarget feature for this, and leave the default with
the current behavior which assumes up to 16-byte loads/stores can
be used. The field also seems to have the ability to be set to 2 bytes,
but I'm not sure what that would be used for.
llvm-svn: 260651
Summary:
It's possible to have resource descriptors and samplers stored in
VGPRs, either by a VMEM instruction or in the case of samplers,
floating-point calculations. When this happens, we need to use
v_readfirstlane to copy these values back to sgprs.
Reviewers: mareko, arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D17102
llvm-svn: 260599
The code change is simple enough: instead of attaching an anonymous SDLoc to splatted
vector constants, use the scalar constant's existing SDLoc since that is what is passed
into getConstant() as a param. But this changes instruction scheduling, so I'll explain
why that happens.
The motivation for this patch starts near:
http://reviews.llvm.org/rL258833
...x86's getZeroVector() could be similarly cleaned up and I thought it would be 'NFC'.
But when I made that change locally, several x86 codegen tests wiggled.
It turns out that the lack of SDLoc consistency in getConstant() changes the way
ScheduleDAGRRList behaves. This is because the SDLoc contains 'IROrder' and some DAG
scheduler algorithms use IROrder for tie-breaking.
Differential Revision: http://reviews.llvm.org/D16972
llvm-svn: 260582
This patch allows the mixing of scaled and unscaled load/stores to form
load/store pairs.
This is a reapplication of r259812, which had an incorrect assert. The
test_stur_str_no_assert() test is a reduced version of the issue hit in
the AArch64 self-host.
PR24465
llvm-svn: 260523
If the two operands to an instruction were both
subregisters of the same super register, it would incorrectly
think this counted as the same constant bus use.
This fixes the verifier error in fmin_legacy.ll which
was missing -verify-machineinstrs.
llvm-svn: 260495
Summary:
This fixes a crash where subsequent spills would be unable to scavenge
a register. In particular, it fixes a crash in piglit's
spec@glsl-1.50@execution@geometry@max-input-components (the test still
has a shader that fails to compile because of too many SGPR spills, but
at least it doesn't crash any more).
This is a candidate for the release branch.
Reviewers: arsenm, tstellarAMD
Subscribers: qcolombet, arsenm
Differential Revision: http://reviews.llvm.org/D16558
llvm-svn: 260427
Instead of passing varargs directly on the user stack, allocate a buffer in
the caller's stack frame and pass a pointer to it. This simplifies the C
ABI (e.g. non-C callers of C functions do not need to use C's user stack if
they have their own mechanism) and allows further optimizations in the future
(e.g. fewer functions may need to use the stack).
Differential Revision: http://reviews.llvm.org/D17048
llvm-svn: 260421
Summary:
Fix case where a pre-inc/dec load/store would not be formed if the
add/sub that forms the inc/dec part of the operation was the first
instruction in the block being examined.
Reviewers: mcrosier, jmolloy, t.p.northover, junbuml
Subscribers: aemerson, rengolin, mcrosier, llvm-commits
Differential Revision: http://reviews.llvm.org/D16785
llvm-svn: 260275
On AVX2 target we are poorly legalizing SIGN_EXTEND ops for which the input's legalized type doesn't have the same number of elements as the destination, resulting in an ANY_EXTEND followed by a SIGN_EXTEND_INREG.
This patch uses the existing SIGN_EXTEND -> SIGN_EXTEND_VECTOR_INREG combine to extend the input to the size of the result and using SIGN_EXTEND_VECTOR_INREG instead.
Differential Revision: http://reviews.llvm.org/D16994
llvm-svn: 260210
As discussed on PR26491, this patch adds support for lowering v4f32 shuffles to the MOVLHPS/MOVHLPS instructions. It also adds support for memory folding with their MOVLPS/MOVHPS load equivalents.
This first patch only really helps SSE1 targets as SSE2+ targets will widen the shuffle mask and use v2f64 equivalents (although they still combine to MOVLHPS/MOVHLPS for v2f64 splats). This will have to be addressed in a future patch, most likely when we add support for binary target shuffle combines.
Differential Revision: http://reviews.llvm.org/D16956
llvm-svn: 260168
Another opportunity to reduce masked stores: in D16691, we decided not to attempt the 'one mask element is set'
transform in InstCombine, but this should be a win for any AVX machine.
Code comments note that this transform could be extended for other targets / cases.
Differential Revision: http://reviews.llvm.org/D16828
llvm-svn: 260145
This matches GCC and MSVC's behaviour, and saves on code size.
We were already not extending i1 return values on x86_64 after r127766. This
takes that patch further by applying it to x86 target as well, and also for i8
and i16.
The ABI docs have been unclear about the required behaviour here. The new i386
psABI [1] clearly states (Table 2.4, page 14) that i1, i8, and i16 return
vales do not need to be extended beyond 8 bits. The x86_64 ABI doc is being
updated to say the same [2].
Differential Revision: http://reviews.llvm.org/D16907
[1]. https://01.org/sites/default/files/file_attach/intel386-psabi-1.0.pdf
[2]. https://groups.google.com/d/msg/x86-64-abi/E8O33onbnGQ/_RFWw_ixDQAJ
llvm-svn: 260133
If a range has a lower bound of 0, add an AssertZext from the
nearest floor power of two.
This allows operations with some workitem intrinsics with known
maximum ranges to use fast 24-bit multiplies.
llvm-svn: 260109
The combineX86ShufflesRecursively only supports unary shuffles, but was missing the opportunity to combine binary shuffles with a zero / undef second input.
This patch resolves target shuffle inputs, converting the shuffle mask elements to SM_SentinelUndef/SM_SentinelZero where possible. It then resolves the updated mask to check if we have created a faux unary shuffle.
Additionally, we now attempt to recursively call combineX86ShufflesRecursively for all input operands (we used to just recurse for unary integer shuffles and unary unpacks) - it safely returns early if its not a target shuffle.
Differential Revision: http://reviews.llvm.org/D16683
llvm-svn: 260063
The current situation isn't great, because the amount of padding
requires is determined by the inverse order of the first encountered
use. We should eventually somehow sort these to minimize wasted space.
Another problem is the alignment of kernel arguments isn't
respected. The group_segment_alignment is always emitted as
the default 16, and typed arguments with higher alignments
or an explicitly set alignment are also ignored.
llvm-svn: 259912
LiveRangeEdit::eliminateDeadDef is used to remove dead define instructions
after rematerialization. To remove a VNI for a vreg from its LiveInterval,
LiveIntervals::removeVRegDefAt is used. However, after non-PHI VNIs are all
removed, PHI VNI are still left in the LiveInterval. Such unused vregs will
be kept in RegsToSpill[] at the end of InlineSpiller::reMaterializeAll and
spiller will allocate stackslot for them.
The fix is to get rid of unused reg by checking whether it has non-dbg
reference instead of whether it has non-empty interval.
llvm-svn: 259895
Using the load immediate only when the immediate (whether signed or unsigned)
can fit in a 16-bit signed field. Namely, from -32768 to 32767 for signed and
0 to 65535 for unsigned. This patch also ensures that we sign-extend under the
right conditions.
llvm-svn: 259840
Choose between MOVD/MOVSS and MOVQ/MOVSD depending on the target vector type.
This has a lot fewer test changes than trying to add this to X86InstrInfo::setExecutionDomain.....
llvm-svn: 259816
This patch allows the mixing of scaled and unscaled load/stores to form
load/store pairs.
PR24465
http://reviews.llvm.org/D12116
Many thanks to Ahmed and Michael for fixes and code review.
This is a reapplication of r246769 and r259790. The tramp3d failure was caused
by an incorrect refactoring in the patch. Specifically, we weren't always
properly clearing the SExtIdx flag.
llvm-svn: 259812
During instruction selection, the AArch64 backend can recognise the
following pattern and generate an [U|S]MADDL instruction, i.e. a
multiply of two 32-bit operands with a 64-bit result:
(mul (sext i32), (sext i32))
However, when one of the operands is constant, the sign extension
gets folded into the constant in SelectionDAG::getNode(). This means
that the instruction selection sees this:
(mul (sext i32), i64)
...which doesn't match the pattern. Sign-extension and 64-bit
multiply instructions are generated, which are slower than one 32-bit
multiply.
Add a pattern to match this and generate the correct instruction, for
both signed and unsigned multiplies.
Patch by Chris Diamand!
llvm-svn: 259800
Fix the lit bug that enabled this "feature" (empty triple is substring
of all possible target triples) and change the two outliers to use the
documented * syntax.
llvm-svn: 259799
This patch adds support for consecutive (load/undef elements) 32-bit loads, followed by trailing undef/zero elements to be combined to a single MOVD load.
Differential Revision: http://reviews.llvm.org/D16729
llvm-svn: 259796
This patch implements softening of long double type (ppcf128) on ppc32
architecture and enables operations for this type for soft float.
Patch by Strahinja Petrovic.
Differential Revision: http://reviews.llvm.org/D15811
llvm-svn: 259791
This patch allows the mixing of scaled and unscaled load/stores to form
load/store pairs.
PR24465
http://reviews.llvm.org/D12116
Many thanks to Ahmed and Michael for fixes and code review.
This is a reapplication of r246769, which was reverted in r246782 due to a
test-suite failure. I'm unable to reproduce the issue at this time.
llvm-svn: 259790
Current SCEV expansion will expand SCEV as a sequence of operations
and doesn't utilize the value already existed. This will introduce
redundent computation which may not be cleaned up throughly by
following optimizations.
This patch introduces an ExprValueMap which is a map from SCEV to the
set of equal values with the same SCEV. When a SCEV is expanded, the
set of values is checked and reused whenever possible before generating
a sequence of operations.
The original commit triggered regressions in Polly tests. The regressions
exposed two problems which have been fixed in current version.
1. Polly will generate a new function based on the old one. To generate an
instruction for the new function, it builds SCEV for the old instruction,
applies some tranformation on the SCEV generated, then expands the transformed
SCEV and insert the expanded value into new function. Because SCEV expansion
may reuse value cached in ExprValueMap, the value in old function may be
inserted into new function, which is wrong.
In SCEVExpander::expand, there is a logic to check the cached value to
be used should dominate the insertion point. However, for the above
case, the check always passes. That is because the insertion point is
in a new function, which is unreachable from the old function. However
for unreachable node, DominatorTreeBase::dominates thinks it will be
dominated by any other node.
The fix is to simply add a check that the cached value to be used in
expansion should be in the same function as the insertion point instruction.
2. When the SCEV is of scConstant type, expanding it directly is cheaper than
reusing a normal value cached. Although in the cached value set in ExprValueMap,
there is a Constant type value, but it is not easy to find it out -- the cached
Value set is not sorted according to the potential cost. Existing reuse logic
in SCEVExpander::expand simply chooses the first legal element from the cached
value set.
The fix is that when the SCEV is of scConstant type, don't try the reuse
logic. simply expand it.
Differential Revision: http://reviews.llvm.org/D12090
llvm-svn: 259736
This patch consists of two parts: a performance fix in DAGCombiner.cpp
and a correctness fix in SelectionDAG.cpp.
The test case tests the bug that's uncovered by the performance fix, and
fixed by the correctness fix.
The performance fix keeps the containers required by the
hasPredecessorHelper (which is a lazy DFS) and reuse them. Since
hasPredecessorHelper is called in a loop, the overall efficiency reduced
from O(n^2) to O(n), where n is the number of SDNodes.
The correctness fix keeps iterating the neighbor list even if it's time
to early return. It will return after finishing adding all neighbors to
Worklist, so that no neighbors are discarded due to the original early
return.
llvm-svn: 259691
Add support for TLS access for Windows on ARM. This generates a similar access
to MSVC for ARM.
The changes to the tablegen data is needed to support loading an external symbol
global that is not for a call. The adjustments to the DAG to DAG transforms are
needed to preserve the 32-bit move.
llvm-svn: 259676
Recommited, after some fixing with test cases.
Updated test cases:
test/CodeGen/AArch64/arm64-misched-memdep-bug.ll
test/CodeGen/AArch64/tailcall_misched_graph.ll
Temporarily disabled test cases:
test/CodeGen/AMDGPU/split-vector-memoperand-offsets.ll
test/CodeGen/PowerPC/ppc64-fastcc.ll (partially updated)
test/CodeGen/PowerPC/vsx-fma-m.ll
test/CodeGen/PowerPC/vsx-fma-sp.ll
http://reviews.llvm.org/D8705
Reviewers: Hal Finkel, Andy Trick.
llvm-svn: 259673
Current SCEV expansion will expand SCEV as a sequence of operations
and doesn't utilize the value already existed. This will introduce
redundent computation which may not be cleaned up throughly by
following optimizations.
This patch introduces an ExprValueMap which is a map from SCEV to the
set of equal values with the same SCEV. When a SCEV is expanded, the
set of values is checked and reused whenever possible before generating
a sequence of operations.
Differential Revision: http://reviews.llvm.org/D12090
llvm-svn: 259662
The GNU toolchain emits __aeabi_divmod for soft-divide on ARM cores
which happens to be a lot faster than __divsi3/__modsi3 when the core
has hardware divide instructions. Do the same here.
Fixes PR26450.
llvm-svn: 259657
MIPS ABI states that .sbss and .sdata sections must have SHF_MIPS_GPREL
flag. See Figure 4–7 on page 69 in the following document:
ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf.
Differential Revision: http://reviews.llvm.org/D15740
llvm-svn: 259641
Follow up to D16217 and D16729
This change uncovered an odd pattern where VZEXT_LOAD v4i64 was being lowered to a load of the lower v2i64 (so the 2nd i64 destination element wasn't being zeroed), I can't find any use/reason for this and have removed the pattern and replaced it so only the 1st i64 element is loaded and the upper bits all zeroed. This matches the description for X86ISD::VZEXT_LOAD
Differential Revision: http://reviews.llvm.org/D16768
llvm-svn: 259635
The purpose of PPCVSXFMAMutate is to elide copies by changing FMA forms
on PPC.
%vreg6<def> = COPY %vreg96
%vreg6<def,tied1> = XSMADDASP %vreg6<tied0>, %vreg5<kill>, %vreg7
;v6 = v6 + v5 * v7
is replaced by
%vreg5<def,tied1> = XSMADDMSP %vreg5<tied0>, %vreg7, %vreg96
;v5 = v5 * v7 + v96
This was broken in the case where the target register was also used as a
multiplicand. Fix this case by checking for it and replacing both uses
with the copied register.
%vreg6<def> = COPY %vreg96
%vreg6<def,tied1> = XSMADDASP %vreg6<tied0>, %vreg5<kill>, %vreg6
;v6 = v6 + v5 * v6
is replaced by
%vreg5<def,tied1> = XSMADDMSP %vreg5<tied0>, %vreg96, %vreg96
;v5 = v5 * v96 + v96
llvm-svn: 259617
If we can't assume the pointer value isn't within the bounds
of the object, it seems risky to try to replace the pointer
calculations.
llvm-svn: 259573
When promoting allocas to LDS, we know we are indexing
into a specific area just created, and the calculation
will also never overflow.
Also emit some of the muls as nsw nuw, because instcombine
infers this already from the range metadata. I think
putting this on the other adds and muls might be OK too,
but I'm not 100% sure.
llvm-svn: 259545
Re-commit of r258951 after fixing layering violation.
The BPF and WebAssembly backends had identical code for emitting errors
for unsupported features, and AMDGPU had very similar code. This merges
them all into one DiagnosticInfo subclass, that can be used by any
backend.
There should be minimal functional changes here, but some AMDGPU tests
have been updated for the new format of errors (it used a slightly
different format to BPF and WebAssembly). The AMDGPU error messages will
now benefit from having precise source locations when debug info is
available.
llvm-svn: 259498
description and changed the regression test accordingly.
The default configuration of a Cortex-R7 is to implement the
VFPv3-D16 architecture and the feature line as it was is too
restrictive.
llvm-svn: 259480
When rematerializing a computation by replacing the copy, use the copy's
location. The location of the copy is more representative of the
original program.
This partially fixes PR10003.
llvm-svn: 259469
Fix a crash in `getMemOpBaseRegImmOfs` that happens if the base of
`MemOp` is a frame index memory operand. The fix is to have
`getMemOpBaseRegImmOfs` bail out in such cases. We can possibly be more
clever here, if needed.
llvm-svn: 259456
Officially, we don't acknowledge non-default configurations of MXCSR,
as getting there would require usage of the FENV_ACCESS pragma (at
least insofar as rounding mode is concerned).
We don't support the pragma, so we can assume that the default
rounding mode - round to nearest, ties to even - is always used.
However, it's inconsistent with the rest of the instruction set,
where MXCSR is always effective (unless otherwise specified).
Also, it's an unnecessary obstacle to the few brave souls that use
fenv.h with LLVM.
Avoid the hard-coded rounding mode for fp_to_f16; use MXCSR instead.
llvm-svn: 259448
Summary:
This is an extension to the existing implementation of r242436 which
restricts to only select inputs. This version fixes missed opportunities
in pr26084 by attempting to lower conditional compare sequences of
and/or trees with setcc leafs. This will additionaly handle the case
when a tree with select input is not a conjunction-disjunction tree
but some of the sub trees are conjunction-disjunction trees.
Reviewers: jmolloy, t.p.northover, mcrosier, MatzeB
Subscribers: mcrosier, llvm-commits, junbuml, haicheng, mssimpso, gberry
Differential Revision: http://reviews.llvm.org/D16291
llvm-svn: 259387
We've found another bug in the code generation logic conditions for a
certain class of always-false conditions, those of the form
if ((a & 1) < 0)
These only reach the back end when compiling without optimization.
The bug was introduced by the choice of using TEST UNDER MASK
to implement a check for
if ((a & MASK) < VAL)
as
if ((a & MASK) == 0)
where VAL is less than the the lowest bit of MASK. This is correct
in all cases except for VAL == 0, in which case the original
condition is always false, but the replacement isn't.
Fixed by excluding that particular case.
llvm-svn: 259381
Remove the old select.ll file and use select-int.ll, select-flt.ll,
select-dbl.ll for testing selects on integers, floats & doubles respectivelly.
llvm-svn: 259361
Minor patch to trace back through target shuffles to the source of the inserted element in a (V)INSERTPS shuffle.
Differential Revision: http://reviews.llvm.org/D16652
llvm-svn: 259343
Those commits created an artificial edge from a cleanup to a synthesized
catchswitch in order to get the MSVC personality routine to execute
cleanups which don't cleanupret and are not wrapped by a catchswitch.
This worked well enough but is not a complete solution in situations
where there the cleanup infinite loops.
However, the real deal breaker behind this approach comes about from a
degenerate case where the cleanup is post-dominated by unreachable *and*
throws an exception. This ends poorly because the catchswitch will
inadvertently catch the exception.
Because of this we should go back to our previous behavior of not
executing certain cleanups (identical behavior with the Itanium ABI
implementation in clang, GCC and ICC).
N.B. I think this could be salvaged by making the catchpad rethrow the
exception and properly transforming throwing calls in the cleanup into
invokes.
llvm-svn: 259338
Previously the code assumed all uses of FI on loads and stores were as
addresses. This checks whether the use is the address or a value and
handles the latter case as it does for non-memory instructions.
llvm-svn: 259306
The previous code was incorrect (can't getReg a frameindex). We could instead optimize it to reduce tree height, but I'm not sure that's worthwhile yet because we then try to eliminate the frameindex.
This patch also fixes frame index elimination for operations which may load or store: it used to assume the base was operand 2 and immediate offset operand 1. That's not true for stores, where they're 4 and 3.
llvm-svn: 259305
The AMDGPUPromoteAlloca pass was emitting the read.local.size
calls, which with HSA was incorrectly selected to reading from
the offset mesa uses off of the kernarg pointer.
Error on intrinsics which aren't supported by HSA, and start
emitting the correct IR to read the workgroup size
out of the dispatch pointer.
Also initialize the pass so it can be tested with opt, and
start moving towards not depending on the subtarget as an
argument.
Start emitting errors for the intrinsics not handled with HSA.
llvm-svn: 259297
Only the dispatch.ptr intrinsic is supposed to be used now to get
the workgroup size, and the read.local.size intrinsics do not
work correctly.
llvm-svn: 259296
Refine the test for whether an instruction is in an expression tree so that
it detects when one tree ends and another begins, so we can place a block
at that point, rather than continuing to find the first instruction not in
a tree at all.
llvm-svn: 259294
The basic optimisation was to convert (mul $LHS, $complex_constant) into
roughly "(shl (mul $LHS, $simple_constant), $simple_amt)" when it was expected
to be cheaper. The original logic checks that the mul only has one use (since
we're mangling $complex_constant), but when used in even more complex
addressing modes there may be an outer addition that can pick up the wrong
value too.
I *think* the ARM addressing-mode problem is actually unreachable at the
moment, but that depends on complex assessments of the profitability of
pre-increment addressing modes so I've put a real check in there instead of an
assertion.
llvm-svn: 259228
Add support for frame pointer use in prolog/epilog.
Supports dynamic allocas but not yet over-aligned locals.
Target-independend CG generates SP updates, but we still need to write
back the SP value to memory when necessary.
llvm-svn: 259220
The trap instruction is emitted as a data-in-text rather
than an instruction. This patch uses the .inst directive
for emitting trap.
Differential Revision: http://reviews.llvm.org/D16684
llvm-svn: 259182
check that the sign extended constant fits into 16-bits if we want a
zero extended value, otherwise go ahead and put it together piecemeal.
Fixes PR26356.
llvm-svn: 259177
Since we only have pair - not single - nontemporal store instructions,
we have to extract the high part into a separate register to be able
to use them.
When the initial nontemporal codegen support was added, I wrote the
extract using the nonsensical UBFX [0,32[.
Use the correct LSR form instead.
llvm-svn: 259134
While legalizing a 64-bit shift left by 1, the following occurs:
We split the shift operand in half: a high half and a low half.
We then create an ADDC with the low half and a ADDE with the high half +
the carry bit from the ADDC.
This is problematic if X is any_ext'd because the high half computation
is now undef + undef + carry bit and there is no way to ensure that the
two undef values had the same bitwise representation. This results in
the lowest bit in the high half turning into garbage.
Instead, do not try to turn shifts into arithmetic during type
legalization.
This fixes PR26350.
llvm-svn: 259065
Enable truncate 128/256bit packed byte/word with AVX512BW but without AVX512VL, use 512bit instructions.
Differential Revision: http://reviews.llvm.org/D16531
llvm-svn: 259044
Author: milena.vujosevic.janicic
Reviewers: dsanders
FastIsel is not supported for microMIPS, thus it needs to be disabled.
Test micromips-zero-mat-uses.ll is deleted since the tested sequence of instructions is not generated for microMIPS without FastISel.
Differential Revision: http://reviews.llvm.org/D15892
llvm-svn: 259039
Re-commit of r258951 after fixing layering violation.
The related LLVM patch adds a backend diagnostic type for reporting
unsupported features, this adds a printer for them to clang.
In the case where debug location information is not available, I've
changed the printer to report the location as the first line of the
function, rather than the closing brace, as the latter does not give the
user any information. This also affects optimisation remarks.
Differential Revision: http://reviews.llvm.org/D16590
llvm-svn: 259035
move ptestm{q|d} intrinsics from patterns form (in td file) to the intrinsics table
Differential Revision: http://reviews.llvm.org/D16633
llvm-svn: 259029
Summary:
findBetterNeighborChains does not handle volatile or indexed stores.
However, it did not check when adding stores to ChainedStores.
Reviewers: arsenm
Differential Revision: http://reviews.llvm.org/D16463
llvm-svn: 259024
This patch revamps the RegStackifier pass with a new tree traversal mechanism,
enabling three major new features:
- Stackification of values with multiple uses, using the result value of set_local
- More aggressive stackification of instructions with side effects
- Reordering operands in commutative instructions to enable more stackification.
llvm-svn: 259009
Summary:
Just does the simple allocation of a stack object and passes
a pointer to the callee.
Differential Revision: http://reviews.llvm.org/D16610
llvm-svn: 258989
Various bits we want to use the new ABI actually compile with "-arch armv7k
-miphoneos-version-min=9.0". Not ideal, but also not ridiculous given how
slices work.
llvm-svn: 258975
The BPF and WebAssembly backends had identical code for emitting errors
for unsupported features, and AMDGPU had very similar code. This merges
them all into one DiagnosticInfo subclass, that can be used by any
backend.
There should be minimal functional changes here, but some AMDGPU tests
have been updated for the new format of errors (it used a slightly
different format to BPF and WebAssembly). The AMDGPU error messages will
now benefit from having precise source locations when debug info is
available.
The implementation of DiagnosticInfoUnsupported::print must be in
lib/Codegen rather than in the existing file in lib/IR/ to avoid
introducing a dependency from IR to CodeGen.
Differential Revision: http://reviews.llvm.org/D16590
llvm-svn: 258951
Summary:
This is a candidate for stable, along with all patches that add the "stoney"
processor.
Reviewers: tstellarAMD
Subscribers: arsenm
Differential Revision: http://reviews.llvm.org/D16485
llvm-svn: 258922
When no device name is specified, default to kaveri
for HSA since SI is not supported and it woud fail.
Default to "tahiti" instead of "SI" since these are
effectively the same, and tahiti is an actual device.
Move default device handling to the TargetMachine
rather than the AMDGPUSubtarget. The module ISA version
is computed from the device name provided with the target
machine, so the attributes printed by the AsmPrinter were
inconsistent with those computed in the subtarget.
Also remove DevName field from subtarget since it's redundant
with getCPU() in the superclass.
llvm-svn: 258901
Summary:
This patch is provided in preparation for removing autoconf on 1/26. The proposal to remove autoconf on 1/26 was discussed on the llvm-dev thread here: http://lists.llvm.org/pipermail/llvm-dev/2016-January/093875.html
"I felt a great disturbance in the [build system], as if millions of [makefiles] suddenly cried out in terror and were suddenly silenced. I fear something [amazing] has happened."
- Obi Wan Kenobi
Reviewers: chandlerc, grosbach, bob.wilson, tstellarAMD, echristo, whitequark
Subscribers: chfast, simoncook, emaste, jholewinski, tberghammer, jfb, danalbert, srhines, arsenm, dschuff, jyknight, dsanders, joker.eph, llvm-commits
Differential Revision: http://reviews.llvm.org/D16471
llvm-svn: 258861
r258781 optimized memcpy/memmove/memcpy so the intrinsic call can return its first argument, but missed the frame index case. Teach it to ignore that case so C code doesn't assert out in these cases.
llvm-svn: 258851
Currently, AnalyzeBranch() fails non-equality comparison between floating points
on X86 (see https://llvm.org/bugs/show_bug.cgi?id=23875). This is because this
function can modify the branch by reversing the conditional jump and removing
unconditional jump if there is a proper fall-through. However, in the case of
non-equality comparison between floating points, this can turn the branch
"unanalyzable". Consider the following case:
jne.BB1
jp.BB1
jmp.BB2
.BB1:
...
.BB2:
...
AnalyzeBranch() will reverse "jp .BB1" to "jnp .BB2" and then "jmp .BB2" will be
removed:
jne.BB1
jnp.BB2
.BB1:
...
.BB2:
...
However, AnalyzeBranch() cannot analyze this branch anymore as there are two
conditional jumps with different targets. This may disable some optimizations
like block-placement: in this case the fall-through behavior is enforced even if
the fall-through block is very cold, which is suboptimal.
Actually this optimization is also done in block-placement pass, which means we
can remove this optimization from AnalyzeBranch(). However, currently
X86::COND_NE_OR_P and X86::COND_NP_OR_E are not reversible: there is no defined
negation conditions for them.
In order to reverse them, this patch defines two new CondCode X86::COND_E_AND_NP
and X86::COND_P_AND_NE. It also defines how to synthesize instructions for them.
Here only the second conditional jump is reversed. This is valid as we only need
them to do this "unconditional jump removal" optimization.
Differential Revision: http://reviews.llvm.org/D11393
llvm-svn: 258847
This patch adds support for trailing zero elements to VZEXT_LOAD loads (and checks that no zero elts occur within the consecutive load).
It also generalizes the 64-bit VZEXT_LOAD load matching to work for loads other than 2x32-bit loads.
After this patch it will also be easier to add support for other basic load patterns like 32-bit VZEXT_LOAD loads, PMOVZX and subvector load insertion.
Differential Revision: http://reviews.llvm.org/D16217
llvm-svn: 258798
I did my best to try to update all the uses in tests that
just happened to use the old ones to the newer intrinsics.
I'm not sure I got all of the immediate operand conversions
correct, since the value seems to have been ignored by the
old pattern but I don't think it really matters.
llvm-svn: 258787
Some of the special intrinsics now that now correspond to a instruction
also have special setting of some registers, e.g. llvm.SI.sendmsg sets
m0 as well as use s_sendmsg. Using these explicit register intrinsics
may be a better option.
Reading the exec mask and others may be useful for debugging. For this
I'm not sure this is entirely correct because we would want this to
be convergent, although it's possible this is already treated
sufficently conservatively.
llvm-svn: 258785
These calls return their first argument, but because LLVM uses an intrinsic
with a void return type, they can't use the returned attribute. Generalize
the store results pass to optimize these calls too.
llvm-svn: 258781
For historic reasons, the behavior of .align differs between targets.
Fortunately, there are alternatives, .p2align and .balign, which make the
interpretation of the parameter explicit, and which behave consistently across
targets.
This patch teaches MC to use .p2align instead of .align, so that people reading
code for multiple architectures don't have to remember which way each platform
does its .align directive.
Differential Revision: http://reviews.llvm.org/D16549
llvm-svn: 258750
There's a special case in EmitLoweredSelect() that produces an improved
lowering for cmov(cmov) patterns. However this special lowering is
currently broken if the inner cmov has multiple users so this patch
stops using it in this case.
If you wonder why this wasn't fixed by continuing to use the special
lowering and inserting a 2nd PHI for the inner cmov: I believe this
would incur additional copies/register pressure so the special lowering
does not improve upon the normal one anymore in this case.
This fixes http://llvm.org/PR26256 (= rdar://24329747)
llvm-svn: 258729
Instructions can be DCE'd after the RegStackify pass. If the instruction which
would be the pop for what would be a push is removed, don't use a push.
llvm-svn: 258694
When generating calls to memcpy, memmove, and memset, use void* as the return
type rather than void, to match the standard signatures for these functions.
This has no practical effect for most targets, since the return values of
these calls aren't being used anyway, and most calling conventions tolerate
this kind of mismatch. However, this change will help support future
optimizations to utilize the return value to avoid holding the argument
value live across a call.
llvm-svn: 258691
This patch was originally committed as r257885, but was reverted due to windows
failures. The cause of these failures has been fixed under r258677, hence
re-committing the original patch.
llvm-svn: 258683
This patch was originally committed as r257884, but was reverted due to windows
failures. The cause of these failures has been fixed under r258677, hence
re-committing the original patch.
llvm-svn: 258682
VPMADD52LUQ - Packed Multiply of Unsigned 52-bit Integers and Add the Low 52-bit Products to Qword Accumulators
VPMADD52HUQ - Packed Multiply of Unsigned 52-bit Unsigned Integers and Add High 52-bit Products to 64-bit Accumulators
Differential Revision: http://reviews.llvm.org/D16407
llvm-svn: 258680
Use AVX1 FP instructions (vmaskmovps/pd) in place of the AVX2 int instructions (vpmaskmovd/q).
Differential Revision: http://reviews.llvm.org/D16528
llvm-svn: 258675
Changes in X86.td:
I set features of Intel processors in incremental form: IVB = SNB + X HSW = IVB + X ..
I added Skylake client processor and defined it's features
FeatureADX was missing on KNL
Added some new features to appropriate processors SMAP, IFMA, PREFETCHWT1, VMFUNC and others
Differential Revision: http://reviews.llvm.org/D16357
llvm-svn: 258659
A cleanup can have paths which unwind or end up in unreachable.
If there is an unreachable path *and* a path which unwinds to caller,
we would mistakenly inject an unwind path to a catchswitch on the
unreachable path. This results in a verifier assertion firing because
the cleanup unwinds to two different places: to the caller and to the
catchswitch.
This occured because we used getCleanupRetUnwindDest to determine if the
cleanuppad had no cleanuprets.
This is incorrect, getCleanupRetUnwindDest returns null for cleanuprets
which unwind to caller.
llvm-svn: 258651
Summary:
Previously, we would just output "foo = bar" in the assembly, and then
ptxas would choke. Now we die before emitting any invalid code.
Reviewers: echristo
Subscribers: jholewinski, llvm-commits, jhen, tra
Differential Revision: http://reviews.llvm.org/D16490
llvm-svn: 258638
For the moment, this file takes way too long to run (see inline comments), but
that should be a temporary problem. The fact that the compile time is so slow
for a target that doesn't support maskmov may be a bug worth investigating too.
llvm-svn: 258629
If the INSERTPS zeroes out all the referenced elements from either of the 2 input vectors (and the input is not already UNDEF), then set that input to UNDEF to reduce dependencies.
llvm-svn: 258622
Some of the conditions necessary to produce ccmp sequences were only
checked in recursive calls to emitConjunctionDisjunctionTree() after
some of the earlier expressions were already built. Move all checks over
to isConjunctionDisjunctionTree() so they are all checked before we
start emitting instructions.
Also rename some variable to better reflect their usage.
llvm-svn: 258605
Cleanups in C++ are a little weird. They are only guaranteed to be
reliably executed if, and only if, there is a viable catch handler which
can handle the exception.
This means that reachability of a cleanup is lexically determined by it
being nested with a try-block which unwinds to a catch. It is *cannot*
be reasoned about by examining the control flow edges leaving a cleanup.
Usually this is not a problem. It becomes a problem when there are *no*
edges out of a cleanup because we believed that code post-dominated by
the cleanup is dead. In LLVM's case, this code is what informs the
personality routine about the presence of a suitable catch handler.
However, the lack of edges to that catch handler makes the handler
become unreachable which causes us to remove it. By removing the
handler, the cleanup becomes unreachable.
Instead, inject a catch-all handler with every cleanup that has no
unwind edges. This will allow us to properly unwind the stack.
This fixes PR25997.
llvm-svn: 258580
The intrinsic target prefix should match the target name
as it appears in the triple.
This is not yet complete, but gets most of the important ones.
llvm.AMDGPU.* intrinsics used by mesa and libclc are still handled
for compatability for now.
llvm-svn: 258557
The promote alloca pass didn't handle these intrinsics and crashed.
These intrinsics should accept any address space, but for now just
erase them to avoid breaking.
llvm-svn: 258537
The current behavior is incorrect, as the two CCs returned by
changeFPCCToAArch64CC, intended to be OR'ed, are instead used
in an AND ccmp chain.
Consider:
define i32 @t(float %a, float %b, float %c, float %d, i32 %e, i32 %f) {
%cc1 = fcmp one float %a, %b
%cc2 = fcmp olt float %c, %d
%and = and i1 %cc1, %cc2
%r = select i1 %and, i32 %e, i32 %f
ret i32 %r
}
Assuming (%a < %b) and (%c < %d); we used to do:
fcmp s0, s1 # nzcv <- 1000
orr w8, wzr, #0x1 # w8 <- 1
csel w9, w8, wzr, mi # w9 <- 1
csel w8, w8, w9, gt # w8 <- 1
fcmp s2, s3 # nzcv <- 1000
cset w9, mi # w9 <- 1
tst w8, w9 # (w8 & w9) == 1, so: nzcv <- 0000
csel w0, w0, w1, ne # w0 <- w0
We now do:
fcmp s2, s3 # nzcv <- 1000
fccmp s0, s1, #0, mi # mi, so: nzcv <- 1000
fccmp s0, s1, #8, le # !le, so: nzcv <- 1000
csel w0, w0, w1, pl # !pl, so: w0 <- w1
In other words, we transformed:
(c < d) && ((a < b) || (a > b))
into:
(c < d) && (a u>= b) && (a u<= b)
whereas, per De Morgan's, we wanted:
(c < d) && !((a u>= b) && (a u<= b))
Note that this problem doesn't occur in the test-suite.
changeFPCCToAArch64CC produces disjunct CCs; here, one -> mi/gt.
We can't represent that in the fccmp chain; it can't express
arbitrary OR sequences, as one comment explains:
In general we can create code for arbitrary "... (and (and A B) C)"
sequences. We can also implement some "or" expressions, because
"(or A B)" is equivalent to "not (and (not A) (not B))" and we can
implement some negation operations. [...] However there is no way
to negate the result of a partial sequence.
Instead, introduce changeFPCCToANDAArch64CC, which produces the
conjunct cond codes:
- (a one b)
== ((a olt b) || (a ogt b))
== ((a ord b) && (a une b))
- (a ueq b)
== ((a uno b) || (a oeq b))
== ((a ule b) && (a uge b))
Note that, at first, one might think that, when PushNegate is true,
we should use the disjunct CCs, in effect doing:
(a || b)
= !(!a && !(b))
= !(!a && !(b1 || b2)) <- changeFPCCToAArch64CC(b, b1, b2)
= !(!a && !b1 && !b2)
However, we can take advantage of the fact that the CC is already
negated, which lets us avoid special-casing PushNegate and doing
the simpler to reason about:
(a || b)
= !(!a && (!b))
= !(!a && (b1 && b2)) <- changeFPCCToANDAArch64CC(!b, b1, b2)
= !(!a && b1 && b2)
This makes both emitConditionalCompare cases behave identically,
and produces correct ccmp sequences for the 2-CC fcmps.
llvm-svn: 258533
This reapplies r258296 and r258366, and also fixes an existing bug in
SelectionDAG.cpp's isMemSrcFromString, neglecting to account for the
offset in a GlobalAddressSDNode, which is uncovered by those patches.
llvm-svn: 258482
Summary:
SETCC with f16 vectors has OperationAction set to Expand but still gets
lowered to FCM* intrinsics based on its result type. This patch skips
lowering of VSETCC if the operand is an f16 vector.
v4 and v8 tests included.
Reviewers: ab, jmolloy
Subscribers: srhines, llvm-commits
Differential Revision: http://reviews.llvm.org/D15361
llvm-svn: 258471
This reverts r258296 and the follow up r258366. With this change, we
miscompiled the following program on Windows:
#include <string>
#include <iostream>
static const char kData[] = "asdf jkl;";
int main() {
std::string s(kData + 3, sizeof(kData) - 3);
std::cout << s << '\n';
}
llvm-svn: 258465
The X86 musttail implementation finds register parameters to forward by
running the calling convention algorithm until a non-register location
is returned. However, assigning a vector memory location has the side
effect of increasing the function's stack alignment. We shouldn't
increase the stack alignment when we are only looking for register
parameters, so this change conditionalizes it.
llvm-svn: 258442
Better handling of the annoying pshuflw/pshufhw ops which only shuffle lower/upper halves of a vector.
Added vXi16 unary shuffle support for cases where i16 elements (from the same half of the source) are being splatted to the whole of one of the halves. This avoids the general lowering case which must shuffle the 32-bit elements first - meaning that we used to end up with unnecessary duplicate pshuflw/pshufhw shuffles.
Note this has the side effect of a lot of SSSE3 test cases no longer needing to use PSHUFB, as it falls below the 3 op combine threshold for when PSHUFB is typically worth it. I've raised PR26183 to discuss if the threshold should be changed and whether we need to make it more specific to the target CPU.
Differential Revision: http://reviews.llvm.org/D14901
llvm-svn: 258440
Summary:
While working on uniform branching, I've hit a few cases where we emit
i1 SETCC operations.
Reviewers: arsenm
Subscribers: arsenm, llvm-commits
Differential Revision: http://reviews.llvm.org/D16233
llvm-svn: 258352
Ahmed Bougacha