This patch reapplies r193676 with an additional fix for the Hexagon backend. The
SystemZ backend has already been fixed by r194148.
The Type Legalizer recognizes that VSELECT needs to be split, because the type
is to wide for the given target. The same does not always apply to SETCC,
because less space is required to encode the result of a comparison. As a result
VSELECT is split and SETCC is unrolled into scalar comparisons.
This commit fixes the issue by checking for VSELECT-SETCC patterns in the DAG
Combiner. If a matching pattern is found, then the result mask of SETCC is
promoted to the expected vector mask type for the given target. Now the type
legalizer will split both VSELECT and SETCC.
This allows the following X86 DAG Combine code to sucessfully detect the MIN/MAX
pattern. This fixes PR16695, PR17002, and <rdar://problem/14594431>.
Reviewed by Nadav
llvm-svn: 194542
This patch moves the jump address materialization inside the noop slide. This
enables patching of the materialization itself or its complete removal. This
patch also adds the ability to define scratch registers that can be used safely
by the code called from the patchpoint intrinsic. At least one scratch register
is required, because that one is used for the materialization of the jump
address. This patch depends on D2009.
Differential Revision: http://llvm-reviews.chandlerc.com/D2074
Reviewed by Andy
llvm-svn: 194306
The idea of the AnyReg Calling Convention is to provide the call arguments in
registers, but not to force them to be placed in a paticular order into a
specified set of registers. Instead it is up tp the register allocator to assign
any register as it sees fit. The same applies to the return value (if
applicable).
Differential Revision: http://llvm-reviews.chandlerc.com/D2009
Reviewed by Andy
llvm-svn: 194293
The Type Legalizer recognizes that VSELECT needs to be split, because the type
is to wide for the given target. The same does not always apply to SETCC,
because less space is required to encode the result of a comparison. As a result
VSELECT is split and SETCC is unrolled into scalar comparisons.
This commit fixes the issue by checking for VSELECT-SETCC patterns in the DAG
Combiner. If a matching pattern is found, then the result mask of SETCC is
promoted to the expected vector mask type for the given target. This mask has
usually the same size as the VSELECT return type (except for Intel KNL). Now the
type legalizer will split both VSELECT and SETCC.
This allows the following X86 DAG Combine code to sucessfully detect the MIN/MAX
pattern. This fixes PR16695, PR17002, and <rdar://problem/14594431>.
Reviewed by Nadav
llvm-svn: 193676
This optimization is not SSE specific so I am moving it to DAGco.
The new scalar_to_vector dag node exposed a missing pattern in the AArch64 target that I needed to add.
llvm-svn: 193393
Calling _chkstk is required on ELF as well as COFF on Windows. Without
_chkstk, functions requiring large stack crash in initialization code.
Previous code tested for COFF format but not Mach-O and this patch modifies
the code to test for Windows OS (both Windows target and MingW target)
but not Mach-O object format: Looks like macho environment was used to
build some EFI code.
Credits to Andrew MacPherson.
llvm-svn: 193289
Without _chkstk functions requiring large stack crash in
initialization code. Previous code tested for COFF format but
not Mach-O and this patch modifies the code to test for Windows.
Credits to Andrew MacPherson.
llvm-svn: 193263
On sandy bridge (PR17654) we now get
vpxor %xmm1, %xmm1, %xmm1
vpunpckhbw %xmm1, %xmm0, %xmm2
vpunpcklbw %xmm1, %xmm0, %xmm0
vinsertf128 $1, %xmm2, %ymm0, %ymm0
On haswell it's a simple
vpmovzxbw %xmm0, %ymm0
There is a maze of duplicated and dead transforms and patterns in this
area. Remove the dead custom lowering of zext v8i16 to v8i32, that's
already handled by LowerAVXExtend.
llvm-svn: 193262
the instruction defenitions and ISEL reflect this.
Prior to this patch these instructions took an i32i8imm, and the high bits were
dropped during encoding. This led to incorrect behavior for shifts by
immediates higher than 255. This patch fixes that issue by detecting large
immediate shifts and returning constant zero (for logical shifts) or capping
the shift amount at an encodable value (for arithmetic shifts).
Fixes <rdar://problem/14968098>
llvm-svn: 193096
Consider the following:
typedef unsigned short ushort4U __attribute__((ext_vector_type(4),
aligned(2)));
typedef unsigned short ushort4 __attribute__((ext_vector_type(4)));
typedef unsigned short ushort8 __attribute__((ext_vector_type(8)));
typedef int int4 __attribute__((ext_vector_type(4)));
int4 __bbase_cvt_int(ushort4 v) {
ushort8 a;
a.lo = v;
return _mm_cvtepu16_epi32(a);
}
This generates the, not unreasonable, IR:
define <4 x i32> @foo0(double %v.coerce) nounwind ssp {
%tmp = bitcast double %v.coerce to <4 x i16>
%tmp1 = shufflevector <4 x i16> %tmp, <4 x i16> undef, <8 x i32> <i32
%0, i32 1, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
%tmp2 = tail call <4 x i32> @llvm.x86.sse41.pmovzxwd(<8 x i16> %tmp1)
ret <4 x i32> %tmp2
}
The problem is when type legalization gets hold of the v4i16. It
legalizes that by spilling to the stack, then doing a zero-extending
load. Things go even more silly from there, ending up with something
like:
_foo0:
movsd %xmm0, -8(%rsp) <== Spill to the stack.
movq -8(%rsp), %xmm0 <== Reload it right back out.
pmovzxwd %xmm0, %xmm1 <== Here's what we actually asked for.
pblendw $1, %xmm1, %xmm0 <== We don't need this at all
pmovzxwd %xmm0, %xmm0 <== We already did this
ret
The v8i8 to v8i16 zext intrinsic gives even worse results, with two
table lookups via pshufb instructions(!!).
To avoid all that, we can move the bitcasting until after we've formed
the wider (legal) vector type. Then our normal codegen flows along
nicely and we get the expected:
_foo0:
pmovzxwd %xmm0, %xmm0
ret
rdar://15245794
llvm-svn: 192866
- Type of index used in extract_vector_elt or insert_vector_elt supposes
to be TLI.getVectorIdxTy() which is pointer type on most targets. It'd
better to truncate (or zero-extend in case it's changed later) it to
mask element type to guarantee they are matching instead of asserting
that.
llvm-svn: 192722
- Lower signed division by constant powers-of-2 to target-independent
DAG operators instead of target-dependent ones to support them better
on targets where vector types are legal but shift operators on that
types are illegal. E.g., on AVX, PSRAW is only available on <8 x i16>
though <16 x i16> is a legal type.
llvm-svn: 192721
In AVX 256bit vectors are valid vectors and therefore the Type Legalizer doesn't
split the VSELECT and SETCC nodes. AVX only supports MIN/MAX on 128bit vectors
and this fix enables vector splitting for this special case in the X86 DAG
Combiner.
This fix is related to PR16695, PR17002, and <rdar://problem/14594431>.
llvm-svn: 191131
The Type Legalizer recognizes that VSELECT needs to be split, because the type
is to wide for the given target. The same does not always apply to SETCC,
because less space is required to encode the result of a comparison. As a result
VSELECT is split and SETCC is unrolled into scalar comparisons.
This commit fixes the issue by checking for VSELECT-SETCC patterns in the DAG
Combiner. If a matching pattern is found, then the result mask of SETCC is
promoted to the expected vector mask for the given target. This mask has usually
te same size as the VSELECT return type (except for Intel KNL). Now the type
legalizer will split both VSELECT and SETCC.
This allows the following X86 DAG Combine code to sucessfully detect the MIN/MAX
pattern. This fixes PR16695, PR17002, and <rdar://problem/14594431>.
llvm-svn: 191130
If the DAG already has only legal types, then the second round of DAG combines
is skipped. In this case VSELECT+SETCC patterns that match a more efficient
instruction (e.g. min/max) are never recognized.
This fix allows VSELECT+SETCC combines if the types are already legal before DAG
type legalization.
Reviewer: Nadav
llvm-svn: 190105
This change came about primarily because of two issues in the existing code.
Niether of:
define i64 @test1(i64 %val) {
%in = trunc i64 %val to i32
tail call i32 @ret32(i32 returned %in)
ret i64 %val
}
define i64 @test2(i64 %val) {
tail call i32 @ret32(i32 returned undef)
ret i32 42
}
should be tail calls, and the function sameNoopInput is responsible. The main
problem is that it is completely symmetric in the "tail call" and "ret" value,
but in reality different things are allowed on each side.
For these cases:
1. Any truncation should lead to a larger value being generated by "tail call"
than needed by "ret".
2. Undef should only be allowed as a source for ret, not as a result of the
call.
Along the way I noticed that a mismatch between what this function treats as a
valid truncation and what the backends see can lead to invalid calls as well
(see x86-32 test case).
This patch refactors the code so that instead of being based primarily on
values which it recurses into when necessary, it starts by inspecting the type
and considers each fundamental slot that the backend will see in turn. For
example, given a pathological function that returned {{}, {{}, i32, {}}, i32}
we would consider each "real" i32 in turn, and ask if it passes through
unchanged. This is much closer to what the backend sees as a result of
ComputeValueVTs.
Aside from the bug fixes, this eliminates the recursion that's going on and, I
believe, makes the bulk of the code significantly easier to understand. The
trade-off is the nasty iterators needed to find the real types inside a
returned value.
llvm-svn: 187787
Due to the weird and wondeful usual arithmetic conversions, some
calculations involving negative values were getting performed in
uint32_t and then promoted to int64_t, which is really not a good
idea.
Patch by Katsuhiro Ueno.
llvm-svn: 187703
All insertf*/extractf* functions replaced with insert/extract since we have insertf and inserti forms.
Added lowering for INSERT_VECTOR_ELT / EXTRACT_VECTOR_ELT for 512-bit vectors.
Added lowering for EXTRACT/INSERT subvector for 512-bit vectors.
Added a test.
llvm-svn: 187491
CustomLowerNode was not being called during SplitVectorOperand,
meaning custom legalization could not be used by targets.
This also adds a test case for NVPTX that depends on this custom
legalization.
Differential Revision: http://llvm-reviews.chandlerc.com/D1195
Attempt to fix the buildbots by making the X86 test I just added platform independent
llvm-svn: 187202
This reverts commit 187198. It broke the bots.
The soft float test probably needs a -triple because of name differences.
On the hard float test I am getting a "roundss $1, %xmm0, %xmm0", instead of
"vroundss $1, %xmm0, %xmm0, %xmm0".
llvm-svn: 187201
CustomLowerNode was not being called during SplitVectorOperand,
meaning custom legalization could not be used by targets.
This also adds a test case for NVPTX that depends on this custom
legalization.
Differential Revision: http://llvm-reviews.chandlerc.com/D1195
llvm-svn: 187198
Use PMIN/PMAX for UGE/ULE vector comparions to reduce the number of required
instructions. This trick also works for UGT/ULT, but there is no advantage in
doing so. It wouldn't reduce the number of instructions and it would actually
reduce performance.
Reviewer: Ben
radar:5972691
llvm-svn: 186432
Summary:
This patch adds explicit calling convention types for the Win64 and
System V/x86-64 ABIs. This allows code to override the default, and use
the Win64 convention on a target that wants to use SysV (and
vice-versa). This is needed to implement the `ms_abi` and `sysv_abi` GNU
attributes.
Reviewers:
CC:
llvm-svn: 186144
in-tree implementations of TargetLoweringBase::isFMAFasterThanMulAndAdd in
order to resolve the following issues with fmuladd (i.e. optional FMA)
intrinsics:
1. On X86(-64) targets, ISD::FMA nodes are formed when lowering fmuladd
intrinsics even if the subtarget does not support FMA instructions, leading
to laughably bad code generation in some situations.
2. On AArch64 targets, ISD::FMA nodes are formed for operations on fp128,
resulting in a call to a software fp128 FMA implementation.
3. On PowerPC targets, FMAs are not generated from fmuladd intrinsics on types
like v2f32, v8f32, v4f64, etc., even though they promote, split, scalarize,
etc. to types that support hardware FMAs.
The function has also been slightly renamed for consistency and to force a
merge/build conflict for any out-of-tree target implementing it. To resolve,
see comments and fixed in-tree examples.
llvm-svn: 185956
Fixes PR16146: gdb.base__call-ar-st.exp fails after
pre-RA-sched=source fixes.
Patch by Xiaoyi Guo!
This also fixes an unsupported dbg.value test case. Codegen was
previously incorrect but the test was passing by luck.
llvm-svn: 182885
Shuffles that only move an element into position 0 of the vector are common in
the output of the loop vectorizer and often generate suboptimal code when SSSE3
is not available. Lower them to vector shifts if possible.
We still prefer palignr over psrldq because it has higher throughput on
sandybridge.
llvm-svn: 182102
X86ISelLowering has support to treat:
(icmp ne (and (xor %flags, -1), (shl 1, flag)), 0)
as if it were actually:
(icmp eq (and %flags, (shl 1, flag)), 0)
However, r179386 has code at the InstCombine level to handle this.
llvm-svn: 181145
I think it's almost impossible to fold atomic fences profitably under
LLVM/C++11 semantics. As a result, this is now unused and just
cluttering up the target interface.
llvm-svn: 179940
As packed comparisons in AVX/SSE produce all 0s or all 1s in each SIMD lane,
vector select could be simplified to AND/OR or removed if one or both values
being selected is all 0s or all 1s.
llvm-svn: 179267
This patch is revised based on patch from Victor Umansky
<victor.umansky@intel.com>. More cases are handled in X86's bool
simplification, i.e.
- SETCC_CARRY
- value is truncated to i1 with AND
As a by-product, PR5443 is also fixed.
llvm-svn: 179265
During LTO, the target options on functions within the same Module may
change. This would necessitate resetting some of the back-end. Do this for X86,
because it's a Friday afternoon.
llvm-svn: 178917
- RDRAND always clears the destination value when a random value is not
available (i.e. CF == 0). This value is truncated or zero-extended as
the false boolean value to be returned. Boolean simplification needs
to skip this 'zext' or 'trunc' node.
llvm-svn: 178312
To enable a load of a call address to be folded with that call, this
load is moved from outside of callseq into callseq. Such a moving
adds a non-glued node (that load) into a glued sequence. This non-glue
load is only removed when DAG selection folds them into a memory form
call instruction. When such instruction selection is disabled, it breaks
DAG schedule.
To prevent that, such moving is disabled when target favors register
indirect call.
Previous workaround disabling CALL32m/CALL64m insn selection is removed.
llvm-svn: 178308
indirect through a memory address is to load the memory address into
a register and then call indirect through the register.
This patch implements this improvement by modifying SelectionDAG to
force a function address which is a memory reference to be loaded
into a virtual register.
Patch by Sriram Murali.
llvm-svn: 178171
- It's still considered aligned when the specified alignment is larger
than the natural alignment;
- The new alignment for the high 128-bit vector should be min(16,
alignment) as the pointer is advanced by 16, a power-of-2 offset.
llvm-svn: 177947
MinGW is almost completely compatible to MSVC, with the exception of the _tls_array global not being available.
Patch by David Nadlinger!
llvm-svn: 177257
LegalizeDAG.cpp uses the value of the comparison operands when checking
the legality of BR_CC, so DAGCombiner should do the same.
v2:
- Expand more BR_CC value types for NVPTX
v3:
- Expand correct BR_CC value types for Hexagon, Mips, and XCore.
llvm-svn: 176694
That can usually be lowered efficiently and is common in sandybridge code.
It would be nice to do this in DAGCombiner but we can't insert arbitrary
BUILD_VECTORs this late.
Fixes PR15462.
llvm-svn: 176634
- Phi nodes should be replaced/updated after lowering CMOV into branch
because 'mainMBB' updating operand in Phi node is changed.
- Add EFLAGS in livein before lowering the 2nd CMOV. It's necessary as
we will reuse the EFLAGS generated before the 1st lowered CMOV, which
won't clobber EFLAGS. However, we need explicitly specify that.
- '-attr=-cmov' test case are added.
llvm-svn: 176598
- Clear 'mayStore' flag when loading from the atomic variable before the
spin loop
- Clear kill flag from one use to multiple use in registers forming the
address to that atomic variable
- don't use a physical register as live-in register in BB (neither entry
nor landing pad.) by copying it into virtual register
(patch by Cameron Zwarich)
llvm-svn: 176538
* Only apply divide bypass optimization when not optimizing for size.
* Fixed bug caused by constant for 0 value of type Int32,
used dividend type to generate the constant instead.
* For atom x86-64 apply the divide bypass to use 16-bit divides instead of
64-bit divides when operand values are small enough.
* Added lit tests for 64-bit divide bypass.
Patch by Tyler Nowicki!
llvm-svn: 176442
- ISD::SHL/SRL/SRA must have either both scalar or both vector operands
but TLI.getShiftAmountTy() so far only return scalar type. As a
result, backend logic assuming that breaks.
- Rename the original TLI.getShiftAmountTy() to
TLI.getScalarShiftAmountTy() and re-define TLI.getShiftAmountTy() to
return target-specificed scalar type or the same vector type as the
1st operand.
- Fix most TICG logic assuming TLI.getShiftAmountTy() a simple scalar
type.
llvm-svn: 176364
sext <4 x i1> to <4 x i64>
sext <4 x i8> to <4 x i64>
sext <4 x i16> to <4 x i64>
I'm running Combine on SIGN_EXTEND_IN_REG and revert SEXT patterns:
(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) -> (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
The sext_in_reg (v4i32 x) may be lowered to shl+sar operations.
The "sar" does not exist on 64-bit operation, so lowering sext_in_reg (v4i64 x) has no vector solution.
I also added a cost of this operations to the AVX costs table.
llvm-svn: 175619
conditions are met:
1. They share the same operand and are in the same BB.
2. Both outputs are used.
3. The target has a native instruction that maps to ISD::FSINCOS node or
the target provides a sincos library call.
Implemented the generic optimization in sdisel and enabled it for
Mac OSX. Also added an additional optimization for x86_64 Mac OSX by
using an alternative entry point __sincos_stret which returns the two
results in xmm0 / xmm1.
rdar://13087969
PR13204
llvm-svn: 173755
This catches many cases where we can emit a more efficient shuffle for a
specific mask or when the mask contains undefs. Once the splat is lowered to
unpacks we can't do that anymore.
There is a possibility of moving the promotion after pshufb matching, but I'm
not sure if pshufb with a mask loaded from memory is faster than 3 shuffles, so
I avoided that for now.
llvm-svn: 173569
(defined by the x32 ABI) mode, in which case its pointers are 32-bits
in size. This knowledge is also added to X86RegisterInfo that now
returns the appropriate registers in getPointerRegClass.
There are many outcomes to this change. In order to keep the patches
separate and manageable, we start by focusing on some simple testable
cases. The patch adds a test with passing a pointer to a function -
focusing on the difference between the two data models for x86-64.
Another test is added for handling of 'sret' arguments (and
functionality is added in X86ISelLowering to make it work).
A note on naming: the "x32 ABI" document refers to the AMD64
architecture (in LLVM it's distinguished by being is64Bits() in the
x86 subtarget) with two variations: the LP64 (default) data model, and
the ILP32 data model. This patch adds predicates to the subtarget
which are consistent with this naming scheme.
llvm-svn: 173503
- Add list of physical registers clobbered in pseudo atomic insts
Physical registers are clobbered when pseudo atomic instructions are
expanded. Add them in clobber list to prevent DAG scheduler to
mis-schedule them after these insns are declared side-effect free.
- Add test case from Michael Kuperstein <michael.m.kuperstein@intel.com>
llvm-svn: 173200
Previously we tried to infer it from the bit width size, with an added
IsIEEE argument for the PPC/IEEE 128-bit case, which had a default
value. This default value allowed bugs to creep in, where it was
inappropriate.
llvm-svn: 173138
The optimization handles esoteric cases but adds a lot of complexity both to the X86 backend and to other backends.
This optimization disables an important canonicalization of chains of SEXT nodes and makes SEXT and ZEXT asymmetrical.
Disabling the canonicalization of consecutive SEXT nodes into a single node disables other DAG optimizations that assume
that there is only one SEXT node. The AVX mask optimizations is one example. Additionally this optimization does not update the cost model.
llvm-svn: 172968
PR 14848. The lowered sequence is based on the existing sequence the target-independent
DAG Combiner creates for the scalar case.
Patch by Zvi Rackover.
llvm-svn: 171953
a TargetMachine to construct (and thus isn't always available), to an
analysis group that supports layered implementations much like
AliasAnalysis does. This is a pretty massive change, with a few parts
that I was unable to easily separate (sorry), so I'll walk through it.
The first step of this conversion was to make TargetTransformInfo an
analysis group, and to sink the nonce implementations in
ScalarTargetTransformInfo and VectorTargetTranformInfo into
a NoTargetTransformInfo pass. This allows other passes to add a hard
requirement on TTI, and assume they will always get at least on
implementation.
The TargetTransformInfo analysis group leverages the delegation chaining
trick that AliasAnalysis uses, where the base class for the analysis
group delegates to the previous analysis *pass*, allowing all but tho
NoFoo analysis passes to only implement the parts of the interfaces they
support. It also introduces a new trick where each pass in the group
retains a pointer to the top-most pass that has been initialized. This
allows passes to implement one API in terms of another API and benefit
when some other pass above them in the stack has more precise results
for the second API.
The second step of this conversion is to create a pass that implements
the TargetTransformInfo analysis using the target-independent
abstractions in the code generator. This replaces the
ScalarTargetTransformImpl and VectorTargetTransformImpl classes in
lib/Target with a single pass in lib/CodeGen called
BasicTargetTransformInfo. This class actually provides most of the TTI
functionality, basing it upon the TargetLowering abstraction and other
information in the target independent code generator.
The third step of the conversion adds support to all TargetMachines to
register custom analysis passes. This allows building those passes with
access to TargetLowering or other target-specific classes, and it also
allows each target to customize the set of analysis passes desired in
the pass manager. The baseline LLVMTargetMachine implements this
interface to add the BasicTTI pass to the pass manager, and all of the
tools that want to support target-aware TTI passes call this routine on
whatever target machine they end up with to add the appropriate passes.
The fourth step of the conversion created target-specific TTI analysis
passes for the X86 and ARM backends. These passes contain the custom
logic that was previously in their extensions of the
ScalarTargetTransformInfo and VectorTargetTransformInfo interfaces.
I separated them into their own file, as now all of the interface bits
are private and they just expose a function to create the pass itself.
Then I extended these target machines to set up a custom set of analysis
passes, first adding BasicTTI as a fallback, and then adding their
customized TTI implementations.
The fourth step required logic that was shared between the target
independent layer and the specific targets to move to a different
interface, as they no longer derive from each other. As a consequence,
a helper functions were added to TargetLowering representing the common
logic needed both in the target implementation and the codegen
implementation of the TTI pass. While technically this is the only
change that could have been committed separately, it would have been
a nightmare to extract.
The final step of the conversion was just to delete all the old
boilerplate. This got rid of the ScalarTargetTransformInfo and
VectorTargetTransformInfo classes, all of the support in all of the
targets for producing instances of them, and all of the support in the
tools for manually constructing a pass based around them.
Now that TTI is a relatively normal analysis group, two things become
straightforward. First, we can sink it into lib/Analysis which is a more
natural layer for it to live. Second, clients of this interface can
depend on it *always* being available which will simplify their code and
behavior. These (and other) simplifications will follow in subsequent
commits, this one is clearly big enough.
Finally, I'm very aware that much of the comments and documentation
needs to be updated. As soon as I had this working, and plausibly well
commented, I wanted to get it committed and in front of the build bots.
I'll be doing a few passes over documentation later if it sticks.
Commits to update DragonEgg and Clang will be made presently.
llvm-svn: 171681
1. Add code to estimate register pressure.
2. Add code to select the unroll factor based on register pressure.
3. Add bits to TargetTransformInfo to provide the number of registers.
llvm-svn: 171469
In order to cost subvector insertion and extraction, we need to know
the type of the subvector being extracted.
No functionality change.
llvm-svn: 171453
into their new header subdirectory: include/llvm/IR. This matches the
directory structure of lib, and begins to correct a long standing point
of file layout clutter in LLVM.
There are still more header files to move here, but I wanted to handle
them in separate commits to make tracking what files make sense at each
layer easier.
The only really questionable files here are the target intrinsic
tablegen files. But that's a battle I'd rather not fight today.
I've updated both CMake and Makefile build systems (I think, and my
tests think, but I may have missed something).
I've also re-sorted the includes throughout the project. I'll be
committing updates to Clang, DragonEgg, and Polly momentarily.
llvm-svn: 171366
directly.
This is in preparation for removing the use of the 'Attribute' class as a
collection of attributes. That will shift to the AttributeSet class instead.
llvm-svn: 171253
register. In most cases we actually compare or select YMM-sized registers
and mixing the two types creates horrible code. This commit optimizes
some of the transition sequences.
PR14657.
llvm-svn: 171148
The vector truncs were scalarized during LegalizeVectorOps, later vectorized again by some DAGCombine optimization
and finally, lowered by a dagcombing optimization. Now, they are properly lowered during LegalizeVectorOps.
No new testcase because the original testcases still work.
llvm-svn: 171146
pmuludq is slow, but it turns out that all the unpacking and packing of the
scalarized mul is even slower. 10% speedup on loop-vectorized paq8p.
llvm-svn: 170985
We match the pattern "x >= y ? x-y : 0" into "subus x, y" and two special cases
if y is a constant. DAGCombiner canonicalizes those so we first have to undo the
canonicalization for those cases. The pattern occurs in gzip when the loop
vectorizer is enabled. Part of PR14613.
llvm-svn: 170273
mention the inline memcpy / memset expansion code is a mess?
This patch split the ZeroOrLdSrc argument into two: IsMemset and ZeroMemset.
The first indicates whether it is expanding a memset or a memcpy / memmove.
The later is whether the memset is a memset of zero. It's totally possible
(likely even) that targets may want to do different things for memcpy and
memset of zero.
llvm-svn: 169959
Also added more comments to explain why it is generally ok to return true.
- Rename getOptimalMemOpType argument IsZeroVal to ZeroOrLdSrc. It's meant to
be true for loaded source (memcpy) or zero constants (memset). The poor name
choice is probably some kind of legacy issue.
llvm-svn: 169954
1. Teach it to use overlapping unaligned load / store to copy / set the trailing
bytes. e.g. On 86, use two pairs of movups / movaps for 17 - 31 byte copies.
2. Use f64 for memcpy / memset on targets where i64 is not legal but f64 is. e.g.
x86 and ARM.
3. When memcpy from a constant string, do *not* replace the load with a constant
if it's not possible to materialize an integer immediate with a single
instruction (required a new target hook: TLI.isIntImmLegal()).
4. Use unaligned load / stores more aggressively if target hooks indicates they
are "fast".
5. Update ARM target hooks to use unaligned load / stores. e.g. vld1.8 / vst1.8.
Also increase the threshold to something reasonable (8 for memset, 4 pairs
for memcpy).
This significantly improves Dhrystone, up to 50% on ARM iOS devices.
rdar://12760078
llvm-svn: 169791
There are still bugs in this pass, as well as other issues that are
being worked on, but the bugs are crashers that occur pretty easily in
the wild. Test cases have been sent to the original commit's review
thread.
This reverts the commits:
r169671: Fix a logic error.
r169604: Move the popcnt tests to an X86 subdirectory.
r168931: Initial commit adding the pass.
llvm-svn: 169683
understand target implementation of any_extend / extload, just generate
zero_extend in place of any_extend for liveouts when the target knows the
zero_extend will be implicit (e.g. ARM ldrb / ldrh) or folded (e.g. x86 movz).
rdar://12771555
llvm-svn: 169536
Sooooo many of these had incorrect or strange main module includes.
I have manually inspected all of these, and fixed the main module
include to be the nearest plausible thing I could find. If you own or
care about any of these source files, I encourage you to take some time
and check that these edits were sensible. I can't have broken anything
(I strictly added headers, and reordered them, never removed), but they
may not be the headers you'd really like to identify as containing the
API being implemented.
Many forward declarations and missing includes were added to a header
files to allow them to parse cleanly when included first. The main
module rule does in fact have its merits. =]
llvm-svn: 169131
This revision attempts to recognize following population-count pattern:
while(a) { c++; ... ; a &= a - 1; ... },
where <c> and <a>could be used multiple times in the loop body.
TODO: On X8664 and ARM, __buildin_ctpop() are not expanded to a efficent
instruction sequence, which need to be improved in the following commits.
Reviewed by Nadav, really appreciate!
llvm-svn: 168931
The stack realignment code was fixed to work when there is stack realignment and
a dynamic alloca is present so this shouldn't cause correctness issues anymore.
Note that this also enables generation of AVX instructions for memset
under the assumptions:
- Unaligned loads/stores are always fast on CPUs supporting AVX
- AVX is not slower than SSE
We may need some tweaked heuristics if one of those assumptions turns out not to
be true.
Effectively reverts r58317. Part of PR2962.
llvm-svn: 167967
chain is correctly setup.
As an example, if the original load must happen before later stores, we need
to make sure the constructed VZEXT_LOAD is constrained to be before the stores.
rdar://12684358
llvm-svn: 167859
- Add RTM code generation support throught 3 X86 intrinsics:
xbegin()/xend() to start/end a transaction region, and xabort() to abort a
tranaction region
llvm-svn: 167573
r165941: Resubmit the changes to llvm core to update the functions to
support different pointer sizes on a per address space basis.
Despite this commit log, this change primarily changed stuff outside of
VMCore, and those changes do not carry any tests for correctness (or
even plausibility), and we have consistently found questionable or flat
out incorrect cases in these changes. Most of them are probably correct,
but we need to devise a system that makes it more clear when we have
handled the address space concerns correctly, and ideally each pass that
gets updated would receive an accompanying test case that exercises that
pass specificaly w.r.t. alternate address spaces.
However, from this commit, I have retained the new C API entry points.
Those were an orthogonal change that probably should have been split
apart, but they seem entirely good.
In several places the changes were very obvious cleanups with no actual
multiple address space code added; these I have not reverted when
I spotted them.
In a few other places there were merge conflicts due to a cleaner
solution being implemented later, often not using address spaces at all.
In those cases, I've preserved the new code which isn't address space
dependent.
This is part of my ongoing effort to clean out the partial address space
code which carries high risk and low test coverage, and not likely to be
finished before the 3.2 release looms closer. Duncan and I would both
like to see the above issues addressed before we return to these
changes.
llvm-svn: 167222
The adc/sbb optimization is to able to convert following expression
into a single adc/sbb instruction:
(ult) ... = x + 1 // where the ult is unsigned-less-than comparison
(ult) ... = x - 1
This change is to flip the "x >u y" (i.e. ugt comparison) in order
to expose the adc/sbb opportunity.
llvm-svn: 167180
%0 = load <8 x i16>* %dest
%1 = shufflevector <8 x i16> %0, <8 x i16> %in,
<8 x i32> < i32 0, i32 1, i32 2, i32 3, i32 13, i32 undef, i32 14, i32 14>
store <8 x i16> %1, <8 x i16>* %dest
We get:
vmovlpd (%eax), %xmm0, %xmm0
instead of:
vmovaps (%eax), %xmm1
vmovsd %xmm1, %xmm0, %xmm0
No extra test-case is added. I just fixed the existing one
(also it uses FileCheck now).
llvm-svn: 166971
- As there's no 64-bit GPRs in 32-bit mode, a custom conversion from v2u32 to
v2f32 is added to improve the efficiency of the code generated.
llvm-svn: 166545
- Replace v4i8/v8i8 -> v8f32 DAG combine with custom lowering to reduce
DAG combine overhead.
- Extend the support to v4i16/v8i16 as well.
llvm-svn: 166487
which is supposed to consistently raise SIGTRAP across all systems. In contrast,
__builtin_trap() behave differently on different systems. e.g. it raises SIGTRAP on ARM, and
SIGILL on X86. The purpose of __builtin_debugtrap() is to consistently provide "trap"
functionality, in the mean time preserve the compatibility with on gcc on __builtin_trap().
The X86 backend is already able to handle debugtrap(). This patch is to:
1) make front-end recognize "__builtin_debugtrap()" (emboddied in the one-line change to Clang).
2) In DAG legalization phase, by default, "debugtrap" will be replaced with "trap", which
make the __builtin_debugtrap() "available" to all existing ports without the hassle of
changing their code.
3) If trap-function is specified (via -trap-func=xyz to llc), both __builtin_debugtrap() and
__builtin_trap() will be expanded into the function call of the specified trap function.
This behavior may need change in the future.
The provided testing-case is to make sure 2) and 3) are working for ARM port, and we
already have a testing case for x86.
llvm-svn: 166300
- If INSERT_VECTOR_ELT is supported (above SSE2, either by custom
sequence of legal insn), transform BUILD_VECTOR into SHUFFLE +
INSERT_VECTOR_ELT if most of elements could be built from SHUFFLE with few
(so far 1) elements being inserted.
llvm-svn: 166288
- MBB address is only valid as an immediate value in Small & Static
code/relocation models. On other models, LEA is needed to load IP address of
the restore MBB.
- A minor fix of MBB in MC lowering is added as well to enable target
relocation flag being propagated into MC.
llvm-svn: 166084
- Add custom FP_TO_SINT on v8i16 (and v8i8 which is legalized as v8i16 due to
vector element-wise widening) to reduce DAG combiner and its overhead added
in X86 backend.
llvm-svn: 166036
Original message:
The attached is the fix to radar://11663049. The optimization can be outlined by following rules:
(select (x != c), e, c) -> select (x != c), e, x),
(select (x == c), c, e) -> select (x == c), x, e)
where the <c> is an integer constant.
The reason for this change is that : on x86, conditional-move-from-constant needs two instructions;
however, conditional-move-from-register need only one instruction.
While the LowerSELECT() sounds to be the most convenient place for this optimization, it turns out to be a bad place. The reason is that by replacing the constant <c> with a symbolic value, it obscure some instruction-combining opportunities which would otherwise be very easy to spot. For that reason, I have to postpone the change to last instruction-combining phase.
The change passes the test of "make check-all -C <build-root/test" and "make -C project/test-suite/SingleSource".
Original message since r165661:
My previous change has a bug: I negated the condition code of a CMOV, and go ahead creating a new CMOV using the *ORIGINAL* condition code.
llvm-svn: 166017
- Besides used in SjLj exception handling, __builtin_setjmp/__longjmp is also
used as a light-weight replacement of setjmp/longjmp which are used to
implementation continuation, user-level threading, and etc. The support added
in this patch ONLY addresses this usage and is NOT intended to support SjLj
exception handling as zero-cost DWARF exception handling is used by default
in X86.
llvm-svn: 165989
X86 doesn't have i8 cmovs so isel would emit a branch. Emitting branches at this
level is often not a good idea because it's too late for many optimizations to
kick in. This solution doesn't add any extensions (truncs are free) and tries
to avoid introducing partial register stalls by filtering direct copyfromregs.
I'm seeing a ~10% speedup on reading a random .png file with libpng15 via
graphicsmagick on x86_64/westmere, but YMMV depending on the microarchitecture.
llvm-svn: 165868
Original message:
The attached is the fix to radar://11663049. The optimization can be outlined by following rules:
(select (x != c), e, c) -> select (x != c), e, x),
(select (x == c), c, e) -> select (x == c), x, e)
where the <c> is an integer constant.
The reason for this change is that : on x86, conditional-move-from-constant needs two instructions;
however, conditional-move-from-register need only one instruction.
While the LowerSELECT() sounds to be the most convenient place for this optimization, it turns out to be a bad place. The reason is that by replacing the constant <c> with a symbolic value, it obscure some instruction-combining opportunities which would otherwise be very easy to spot. For that reason, I have to postpone the change to last instruction-combining phase.
The change passes the test of "make check-all -C <build-root/test" and "make -C project/test-suite/SingleSource".
llvm-svn: 165661
- Due to the current matching vector elements constraints in
ISD::FP_ROUND, rounding from v2f64 to v4f32 (after legalization from
v2f32) is scalarized. Add a customized v2f32 widening to convert it
into a target-specific X86ISD::VFPROUND to work around this
constraints.
llvm-svn: 165631
- Due to the current matching vector elements constraints in ISD::FP_EXTEND,
rounding from v2f32 to v2f64 is scalarized. Add a customized v2f32 widening
to convert it into a target-specific X86ISD::VFPEXT to work around this
constraints. This patch also reverts a previous attempt to fix this issue by
recovering the scalarized ISD::FP_EXTEND pattern and thus significantly
reduces the overhead of supporting non-power-2 vector FP extend.
llvm-svn: 165625
We use the enums to query whether an Attributes object has that attribute. The
opaque layer is responsible for knowing where that specific attribute is stored.
llvm-svn: 165488
- Add 'HwEncoding' for X86 registers and call getEncodingValue() to
retrieve their encoding values.
- This's the first step to adopt new scheme. Furthur revising is onging.
llvm-svn: 165241
- Rewrite/merge pseudo-atomic instruction emitters to address the
following issue:
* Reduce one unnecessary load in spin-loop
previously the spin-loop looks like
thisMBB:
newMBB:
ld t1 = [bitinstr.addr]
op t2 = t1, [bitinstr.val]
not t3 = t2 (if Invert)
mov EAX = t1
lcs dest = [bitinstr.addr], t3 [EAX is implicit]
bz newMBB
fallthrough -->nextMBB
the 'ld' at the beginning of newMBB should be lift out of the loop
as lcs (or CMPXCHG on x86) will load the current memory value into
EAX. This loop is refined as:
thisMBB:
EAX = LOAD [MI.addr]
mainMBB:
t1 = OP [MI.val], EAX
LCMPXCHG [MI.addr], t1, [EAX is implicitly used & defined]
JNE mainMBB
sinkMBB:
* Remove immopc as, so far, all pseudo-atomic instructions has
all-register form only, there is no immedidate operand.
* Remove unnecessary attributes/modifiers in pseudo-atomic instruction
td
* Fix issues in PR13458
- Add comprehensive tests on atomic ops on various data types.
NOTE: Some of them are turned off due to missing functionality.
- Revise tests due to the new spin-loop generated.
llvm-svn: 164281
- Enhance the fix to PR12312 to support wider integer, such as 256-bit
integer. If more than 1 fully evaluated vectors are found, POR them
first followed by the final PTEST.
llvm-svn: 163832
- BlockAddress has no support of BA + offset form and there is no way to
propagate that offset into machine operand;
- Add BA + offset support and a new interface 'getTargetBlockAddress' to
simplify target block address forming;
- All targets are modified to use new interface and X86 backend is enhanced to
support BA + offset addressing.
llvm-svn: 163743
- If a boolean value is generated from CMOV and tested as boolean value,
simplify the use of test result by referencing the original condition.
RDRAND intrinisc is one of such cases.
llvm-svn: 163516
- CodeGenPrepare pass for identifying div/rem ops
- Backend specifies the type mapping using addBypassSlowDivType
- Enabled only for Intel Atom with O2 32-bit -> 8-bit
- Replace IDIV with instructions which test its value and use DIVB if the value
is positive and less than 256.
- In the case when the quotient and remainder of a divide are used a DIV
and a REM instruction will be present in the IR. In the non-Atom case
they are both lowered to IDIVs and CSE removes the redundant IDIV instruction,
using the quotient and remainder from the first IDIV. However,
due to this optimization CSE is not able to eliminate redundant
IDIV instructions because they are located in different basic blocks.
This is overcome by calculating both the quotient (DIV) and remainder (REM)
in each basic block that is inserted by the optimization and reusing the result
values when a subsequent DIV or REM instruction uses the same operands.
- Test cases check for the presents of the optimization when calculating
either the quotient, remainder, or both.
Patch by Tyler Nowicki!
llvm-svn: 163150
output chain is correctly setup.
As an example, if the original load must happen before later stores, we need
to make sure the constructed VZEXT_LOAD is constrained to be before the stores.
rdar://11457792
llvm-svn: 163036
- In addition to undefined, if V2 is zero vector, skip 2nd PSHUFB and POR as
well as PSHUFB will zero elements with negative indices.
Patch by Sriram Murali <sriram.murali@intel.com>
llvm-svn: 163018
- Add a target-specific DAG optimization to recognize a pattern PTEST-able.
Such a pattern is a OR'd tree with X86ISD::OR as the root node. When
X86ISD::OR node has only its flag result being used as a boolean value and
all its leaves are extracted from the same vector, it could be folded into an
X86ISD::PTEST node.
llvm-svn: 162735
this allows for better code generation.
Added a new DAGCombine transformation to convert FMAX and FMIN to FMANC and
FMINC, which are commutative.
For example:
movaps %xmm0, %xmm1
movsd LC(%rip), %xmm0
minsd %xmm1, %xmm0
becomes:
minsd LC(%rip), %xmm0
llvm-svn: 162187
arithmetic instructions. However, when small data types are used, a truncate
node appears between the SETCC node and the arithmetic operation. This patch
adds support for this pattern.
Before:
xorl %esi, %edi
testb %dil, %dil
setne %al
ret
After:
xorb %dil, %sil
setne %al
ret
rdar://12081007
llvm-svn: 162160
- FP_EXTEND only support extending from vectors with matching elements.
This results in the scalarization of extending to v2f64 from v2f32,
which will be legalized to v4f32 not matching with v2f64.
- add X86-specific VFPEXT supproting extending from v4f32 to v2f64.
- add BUILD_VECTOR lowering helper to recover back the original
extending from v4f32 to v2f64.
- test case is enhanced to include different vector width.
llvm-svn: 161894
- FCMOV only supports a subset of X86 conditions. Skip boolean
simplification if X86 condition is not valid for FCMOV.
- add a minimal test case for PR13577.
llvm-svn: 161732
- if a boolean test (X86ISD::CMP or X86ISD:SUB) checks a boolean value
generated from X86ISD::SETCC, try to simplify the boolean value
generation and checking by reusing the original EFLAGS with proper
condition code
- add hooks to X86 specific SETCC/BRCOND/CMOV, the major 3 places
consuming EFLAGS
part of patches fixing PR12312
llvm-svn: 161687
We perform the following:
1> Use SUB instead of CMP for i8,i16,i32 and i64 in ISel lowering.
2> Modify MachineCSE to correctly handle implicit defs.
3> Convert SUB back to CMP if possible at peephole.
Removed pattern matching of (a>b) ? (a-b):0 and like, since they are handled
by peephole now.
rdar://11873276
llvm-svn: 161462
Fast isel doesn't currently have support for translating builtin function
calls to target instructions. For embedded environments where the library
functions are not available, this is a matter of correctness and not
just optimization. Most of this patch is just arranging to make the
TargetLibraryInfo available in fast isel. <rdar://problem/12008746>
llvm-svn: 161232
large immediates. Add dag combine logic to recover in case the large
immediates doesn't fit in cmp immediate operand field.
int foo(unsigned long l) {
return (l>> 47) == 1;
}
we produce
%shr.mask = and i64 %l, -140737488355328
%cmp = icmp eq i64 %shr.mask, 140737488355328
%conv = zext i1 %cmp to i32
ret i32 %conv
which codegens to
movq $0xffff800000000000,%rax
andq %rdi,%rax
movq $0x0000800000000000,%rcx
cmpq %rcx,%rax
sete %al
movzbl %al,%eax
ret
TargetLowering::SimplifySetCC would transform
(X & -256) == 256 -> (X >> 8) == 1
if the immediate fails the isLegalICmpImmediate() test. For x86,
that's immediates which are not a signed 32-bit immediate.
Based on a patch by Eli Friedman.
PR10328
rdar://9758774
llvm-svn: 160346
uint32_t hi(uint64_t res)
{
uint_32t hi = res >> 32;
return !hi;
}
llvm IR looks like this:
define i32 @hi(i64 %res) nounwind uwtable ssp {
entry:
%lnot = icmp ult i64 %res, 4294967296
%lnot.ext = zext i1 %lnot to i32
ret i32 %lnot.ext
}
The optimizer has optimize away the right shift and truncate but the resulting
constant is too large to fit in the 32-bit immediate field. The resulting x86
code is worse as a result:
movabsq $4294967296, %rax ## imm = 0x100000000
cmpq %rax, %rdi
sbbl %eax, %eax
andl $1, %eax
This patch teaches the x86 lowering code to handle ult against a large immediate
with trailing zeros. It will issue a right shift and a truncate followed by
a comparison against a shifted immediate.
shrq $32, %rdi
testl %edi, %edi
sete %al
movzbl %al, %eax
It also handles a ugt comparison against a large immediate with trailing bits
set. i.e. X > 0x0ffffffff -> (X >> 32) >= 1
rdar://11866926
llvm-svn: 160312
multiple scalars and insert them into a vector. Next, we shuffle the elements
into the correct places, as before.
Also fix a small dagcombine bug in SimplifyBinOpWithSameOpcodeHands, when the
migration of bitcasts happened too late in the SelectionDAG process.
llvm-svn: 159991
The CopyToReg nodes that set up the argument registers before a call
must be glued to the call instruction. Otherwise, the scheduler may emit
the physreg copies long before the call, causing long live ranges for
the fixed registers.
Besides disabling good register allocation, that can also expose
problems when EmitInstrWithCustomInserter() splits a basic block during
the live range of a physreg.
llvm-svn: 159721
Before this patch in pic 32 bit code we would add the global base register
and not load from that address. This is a really old bug, but before the
introduction of the tls attributes we would never select initial exec for
pic code.
llvm-svn: 159409
The function live-out registers must be live at all function returns,
and %RCX is only used by eh.return. When a function also has a normal
return, only %RAX holds a return value.
This fixes PR13188.
llvm-svn: 159116
TargetLoweringObjectFileELF. Use this to support it on X86. Unlike ARM,
on X86 it is not easy to find out if .init_array should be used or not, so
the decision is made via TargetOptions and defaults to off.
Add a command line option to llc that enables it.
llvm-svn: 158692
This patch will generate the following for integer ABS:
movl %edi, %eax
negl %eax
cmovll %edi, %eax
INSTEAD OF
movl %edi, %ecx
sarl $31, %ecx
leal (%rdi,%rcx), %eax
xorl %ecx, %eax
There exists a target-independent DAG combine for integer ABS, which converts
integer ABS to sar+add+xor. For X86, we match this pattern back to neg+cmov.
This is implemented in PerformXorCombine.
rdar://10695237
llvm-svn: 158175
This patch will optimize the following
movq %rdi, %rax
subq %rsi, %rax
cmovsq %rsi, %rdi
movq %rdi, %rax
to
cmpq %rsi, %rdi
cmovsq %rsi, %rdi
movq %rdi, %rax
Perform this optimization if the actual result of SUB is not used.
rdar: 11540023
llvm-svn: 158126
This implements codegen support for accesses to thread-local variables
using the local-dynamic model, and adds a clean-up pass so that the base
address for the TLS block can be re-used between local-dynamic access on
an execution path.
llvm-svn: 157818
to pass around a struct instead of a large set of individual values. This
cleans up the interface and allows more information to be added to the struct
for future targets without requiring changes to each and every target.
NV_CONTRIB
llvm-svn: 157479
This patch will optimize -(x != 0) on X86
FROM
cmpl $0x01,%edi
sbbl %eax,%eax
notl %eax
TO
negl %edi
sbbl %eax %eax
In order to generate negl, I added patterns in Target/X86/X86InstrCompiler.td:
def : Pat<(X86sub_flag 0, GR32:$src), (NEG32r GR32:$src)>;
rdar: 10961709
llvm-svn: 156312
This will be used to determine whether it's profitable to turn a select into a
branch when the branch is likely to be predicted.
Currently enabled for everything but Atom on X86 and Cortex-A9 devices on ARM.
I'm not entirely happy with the name of this flag, suggestions welcome ;)
llvm-svn: 156233
This patch will optimize the following cases on X86
(a > b) ? (a-b) : 0
(a >= b) ? (a-b) : 0
(b < a) ? (a-b) : 0
(b <= a) ? (a-b) : 0
FROM
movl %edi, %ecx
subl %esi, %ecx
cmpl %edi, %esi
movl $0, %eax
cmovll %ecx, %eax
TO
xorl %eax, %eax
subl %esi, %edi
cmovll %eax, %edi
movl %edi, %eax
rdar: 10734411
llvm-svn: 155919
x == -y --> x+y == 0
x != -y --> x+y != 0
On x86, the generated code goes from
negl %esi
cmpl %esi, %edi
je .LBB0_2
to
addl %esi, %edi
je .L4
This case is correctly handled for ARM with "cmn".
Patch by Manman Ren.
rdar://11245199
PR12545
llvm-svn: 155739
* Model FPSW (the FPU status word) as a register.
* Add ISel patterns for the FUCOM*, FNSTSW and SAHF instructions.
* During Legalize/Lowering, build a node sequence to transfer the comparison
result from FPSW into EFLAGS. If you're wondering about the right-shift: That's
an implicit sub-register extraction (%ax -> %ah) which is handled later on by
the instruction selector.
Fixes PR6679. Patch by Christoph Erhardt!
llvm-svn: 155704
immediate. We can't use it here because the shuffle code does not check that
the lower part of the word is identical to the upper part.
llvm-svn: 155440
Original message:
Modify the code that lowers shuffles to blends from using blendvXX to vblendXX.
blendV uses a register for the selection while Vblend uses an immediate.
On sandybridge they still have the same latency and execute on the same execution ports.
llvm-svn: 154483
blendv uses a register for the selection while vblend uses an immediate.
On sandybridge they still have the same latency and execute on the same execution ports.
llvm-svn: 154396
legalizer always use the DAG entry node. This is wrong when the libcall is
emitted as a tail call since it effectively folds the return node. If
the return node's input chain is not the entry (i.e. call, load, or store)
use that as the tail call input chain.
PR12419
rdar://9770785
rdar://11195178
llvm-svn: 154370
in TargetLowering. There was already a FIXME about this location being
odd. The interface is simplified as a consequence. This will also make
it easier to change TLS models when compiling with PIE.
llvm-svn: 154292
Previously we used three instructions to broadcast an immediate value into a
vector register.
On Sandybridge we continue to load the broadcasted value from the constant pool.
llvm-svn: 154284
This allows us to keep passing reduced masks to SimplifyDemandedBits, but
know about all the bits if SimplifyDemandedBits fails. This allows instcombine
to simplify cases like the one in the included testcase.
llvm-svn: 154011
Specifically, remove the magic number when checking to see if the copy has a
glue operand and simplify the checking logic.
rdar://10930395
llvm-svn: 152041
In this instance we are generating the tail-call during legalizeDAG. The 2nd
floor call can't be a tail call because it clobbers %xmm1, which is defined by
the first floor call. The first floor call can't be a tail-call because it's
not in the tail position. The only reasonable way I could think to fix this
in a target-independent manner was to check for glue logic on the copy reg.
rdar://10930395
llvm-svn: 151877
the processor keeps a return addresses stack (RAS) which stores the address
and the instruction execution state of the instruction after a function-call
type branch instruction.
Calling a "noreturn" function with normal call instructions (e.g. bl) can
corrupt RAS and causes 100% return misprediction so LLVM should use a
unconditional branch instead. i.e.
mov lr, pc
b _foo
The "mov lr, pc" is issued in order to get proper backtrace.
rdar://8979299
llvm-svn: 151623
[Joe Groff] Hi everyone. My previous patch applied as r151382 had a few problems:
Clang raised a warning, and X86 LowerOperation would assert out for
fptoui f64 to i32 because it improperly lowered to an illegal
BUILD_PAIR. Here's a patch that addresses these issues. Let me know if
any other changes are necessary. Thanks.
llvm-svn: 151432
Call instructions no longer have a list of 43 call-clobbered registers.
Instead, they get a single register mask operand with a bit vector of
call-preserved registers.
This saves a lot of memory, 42 x 32 bytes = 1344 bytes per call
instruction, and it speeds up building call instructions because those
43 imp-def operands no longer need to be added to use-def lists. (And
removed and shifted and re-added for every explicit call operand).
Passes like LiveVariables, LiveIntervals, RAGreedy, PEI, and
BranchFolding are significantly faster because they can deal with call
clobbers in bulk.
Overall, clang -O2 is between 0% and 8% faster, uniformly distributed
depending on call density in the compiled code. Debug builds using
clang -O0 are 0% - 3% faster.
I have verified that this patch doesn't change the assembly generated
for the LLVM nightly test suite when building with -disable-copyprop
and -disable-branch-fold.
Branch folding behaves slightly differently in a few cases because call
instructions have different hash values now.
Copy propagation flushes its data structures when it crosses a register
mask operand. This causes it to leave a few dead copies behind, on the
order of 20 instruction across the entire nightly test suite, including
SPEC. Fixing this properly would require the pass to use different data
structures.
llvm-svn: 150638
Adds an instruction itinerary to all x86 instructions, giving each a default latency of 1, using the InstrItinClass IIC_DEFAULT.
Sets specific latencies for Atom for the instructions in files X86InstrCMovSetCC.td, X86InstrArithmetic.td, X86InstrControl.td, and X86InstrShiftRotate.td. The Atom latencies for the remainder of the x86 instructions will be set in subsequent patches.
Adds a test to verify that the scheduler is working.
Also changes the scheduling preference to "Hybrid" for i386 Atom, while leaving x86_64 as ILP.
Patch by Preston Gurd!
llvm-svn: 149558
It adds register mask operands to x86 call instructions. Once all the
backend passes support register mask operands, this will be permanently
enabled.
llvm-svn: 148438
In CanXFormVExtractWithShuffleIntoLoad we assumed that EXTRACT_VECTOR_ELT can be later handled by the DAGCombiner.
However, in some cases on AVX, the EXTRACT_VECTOR_ELT is legalized to EXTRACT_SUBVECTOR + EXTRACT_VECTOR_ELT, which
currently is not handled by the DAGCombiner. In this patch I added a check that we only extract from the XMM part.
llvm-svn: 148298
We know that the blend instructions only use the MSB, so if the mask is
sign-extended then we can convert it into a SHL instruction. This is a
common pattern because the type-legalizer sign-extends the i1 type which
is used by the LLVM-IR for the condition.
Added a new optimization in SimplifyDemandedBits for SIGN_EXTEND_INREG -> SHL.
llvm-svn: 148225
lc: X86ISelLowering.cpp:6480: llvm::SDValue llvm::X86TargetLowering::LowerVECTOR_SHUFFLE(llvm::SDValue, llvm::SelectionDAG&) const: Assertion `V1.getOpcode() != ISD::UNDEF&& "Op 1 of shuffle should not be undef"' failed.
Added a test.
llvm-svn: 148044
As the comment around 7746 says, it's better to use the x87 extended precision
here than SSE. And the generic code doesn't know how to do that. It also regains
the speed lost for the uint64_to_float.c testcase.
<rdar://problem/10669858>
llvm-svn: 147869
Testing: passed 'make check' including LIT tests for all sequences being handled (both SSE and AVX)
Reviewers: Evan Cheng, David Blaikie, Bruno Lopes, Elena Demikhovsky, Chad Rosier, Anton Korobeynikov
llvm-svn: 147601
This small bit of ASM code is sufficient to do what the old algorithm did:
movq %rax, %xmm0
punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
#ifdef __SSE3__
haddpd %xmm0, %xmm0
#else
pshufd $0x4e, %xmm0, %xmm1
addpd %xmm1, %xmm0
#endif
It's arguably faster. One caveat, the 'haddpd' instruction isn't very fast on
all processors.
<rdar://problem/7719814>
llvm-svn: 147593
(x > y) ? x : y
=>
(x >= y) ? x : y
So for something like
(x - y) > 0 : (x - y) ? 0
It will be
(x - y) >= 0 : (x - y) ? 0
This makes is possible to test sign-bit and eliminate a comparison against
zero. e.g.
subl %esi, %edi
testl %edi, %edi
movl $0, %eax
cmovgl %edi, %eax
=>
xorl %eax, %eax
subl %esi, $edi
cmovsl %eax, %edi
rdar://10633221
llvm-svn: 147512
LZCNT instructions are available. Force promotion to i32 to get
a smaller encoding since the fix-ups necessary are just as complex for
either promoted type
We can't do standard promotion for CTLZ when lowering through BSR
because it results in poor code surrounding the 'xor' at the end of this
instruction. Essentially, if we promote the entire CTLZ node to i32, we
end up doing the xor on a 32-bit CTLZ implementation, and then
subtracting appropriately to get back to an i8 value. Instead, our
custom logic just uses the knowledge of the incoming size to compute
a perfect xor. I'd love to know of a way to fix this, but so far I'm
drawing a blank. I suspect the legalizer could be more clever and/or it
could collude with the DAG combiner, but how... ;]
llvm-svn: 147251
'bsf' instructions here.
This one is actually debatable to my eyes. It's not clear that any chip
implementing 'tzcnt' would have a slow 'bsf' for any reason, and unless
EFLAGS or a zero input matters, 'tzcnt' is just a longer encoding.
Still, this restores the old behavior with 'tzcnt' enabled for now.
llvm-svn: 147246
X86ISelLowering C++ code. Because this is lowered via an xor wrapped
around a bsr, we want the dagcombine which runs after isel lowering to
have a chance to clean things up. In particular, it is very common to
see code which looks like:
(sizeof(x)*8 - 1) ^ __builtin_clz(x)
Which is trying to compute the most significant bit of 'x'. That's
actually the value computed directly by the 'bsr' instruction, but if we
match it too late, we'll get completely redundant xor instructions.
The more naive code for the above (subtracting rather than using an xor)
still isn't handled correctly due to the dagcombine getting confused.
Also, while here fix an issue spotted by inspection: we should have been
expanding the zero-undef variants to the normal variants when there is
an 'lzcnt' instruction. Do so, and test for this. We don't want to
generate unnecessary 'bsr' instructions.
These two changes fix some regressions in encoding and decoding
benchmarks. However, there is still a *lot* to be improve on in this
type of code.
llvm-svn: 147244
use the zero-undefined variants of CTTZ and CTLZ. These are just simple
patterns for now, there is more to be done to make real world code using
these constructs be optimized and codegen'ed properly on X86.
The existing tests are spiffed up to check that we no longer generate
unnecessary cmov instructions, and that we generate the very important
'xor' to transform bsr which counts the index of the most significant
one bit to the number of leading (most significant) zero bits. Also they
now check that when the variant with defined zero result is used, the
cmov is still produced.
llvm-svn: 146974
undefined result. This adds new ISD nodes for the new semantics,
selecting them when the LLVM intrinsic indicates that the undef behavior
is desired. The new nodes expand trivially to the old nodes, so targets
don't actually need to do anything to support these new nodes besides
indicating that they should be expanded. I've done this for all the
operand types that I could figure out for all the targets. Owners of
various targets, please review and let me know if any of these are
incorrect.
Note that the expand behavior is *conservatively correct*, and exactly
matches LLVM's current behavior with these operations. Ideally this
patch will not change behavior in any way. For example the regtest suite
finds the exact same instruction sequences coming out of the code
generator. That's why there are no new tests here -- all of this is
being exercised by the existing test suite.
Thanks to Duncan Sands for reviewing the various bits of this patch and
helping me get the wrinkles ironed out with expanding for each target.
Also thanks to Chris for clarifying through all the discussions that
this is indeed the approach he was looking for. That said, there are
likely still rough spots. Further review much appreciated.
llvm-svn: 146466
This was actually a bit of a mess. TLI.setPrefLoopAlignment was clearly
documented as taking log2(bytes) units, but the x86 target would still
set a preferred loop alignment of '16'.
CodePlacementOpt passed this number on to the basic block, and
AsmPrinter interpreted it as bytes.
Now both MachineFunction and MachineBasicBlock use logarithmic
alignments.
Obviously, MachineConstantPool still measures alignments in bytes, so we
can emulate the thrill of using as.
llvm-svn: 145889
change, now you need a TargetOptions object to create a TargetMachine. Clang
patch to follow.
One small functionality change in PTX. PTX had commented out the machine
verifier parts in their copy of printAndVerify. That now calls the version in
LLVMTargetMachine. Users of PTX who need verification disabled should rely on
not passing the command-line flag to enable it.
llvm-svn: 145714
Before:
movabsq $4294967296, %rax ## encoding: [0x48,0xb8,0x00,0x00,0x00,0x00,0x01,0x00,0x00,0x00]
testq %rax, %rdi ## encoding: [0x48,0x85,0xf8]
jne LBB0_2 ## encoding: [0x75,A]
After:
btq $32, %rdi ## encoding: [0x48,0x0f,0xba,0xe7,0x20]
jb LBB0_2 ## encoding: [0x72,A]
btq is usually slower than testq because it doesn't fuse with the jump, but here we're better off
saving one register and a giant movabsq.
llvm-svn: 145103
When this field is true it means that the load is from constant (runt-time or compile-time) and so can be hoisted from loops or moved around other memory accesses
llvm-svn: 144100
On x86: (shl V, 1) -> add V,V
Hardware support for vector-shift is sparse and in many cases we scalarize the
result. Additionally, on sandybridge padd is faster than shl.
llvm-svn: 143311
fixes: Use a separate register, instead of SP, as the
calling-convention resource, to avoid spurious conflicts with
actual uses of SP. Also, fix unscheduling of calling sequences,
which can be triggered by pseudo-two-address dependencies.
llvm-svn: 143206
it fixes the dragonegg self-host (it looks like gcc is miscompiled).
Original commit messages:
Eliminate LegalizeOps' LegalizedNodes map and have it just call RAUW
on every node as it legalizes them. This makes it easier to use
hasOneUse() heuristics, since unneeded nodes can be removed from the
DAG earlier.
Make LegalizeOps visit the DAG in an operands-last order. It previously
used operands-first, because LegalizeTypes has to go operands-first, and
LegalizeTypes used to be part of LegalizeOps, but they're now split.
The operands-last order is more natural for several legalization tasks.
For example, it allows lowering code for nodes with floating-point or
vector constants to see those constants directly instead of seeing the
lowered form (often constant-pool loads). This makes some things
somewhat more complicated today, though it ought to allow things to be
simpler in the future. It also fixes some bugs exposed by Legalizing
using RAUW aggressively.
Remove the part of LegalizeOps that attempted to patch up invalid chain
operands on libcalls generated by LegalizeTypes, since it doesn't work
with the new LegalizeOps traversal order. Instead, define what
LegalizeTypes is doing to be correct, and transfer the responsibility
of keeping calls from having overlapping calling sequences into the
scheduler.
Teach the scheduler to model callseq_begin/end pairs as having a
physical register definition/use to prevent calls from having
overlapping calling sequences. This is also somewhat complicated, though
there are ways it might be simplified in the future.
This addresses rdar://9816668, rdar://10043614, rdar://8434668, and others.
Please direct high-level questions about this patch to management.
Delete #if 0 code accidentally left in.
llvm-svn: 143188
on every node as it legalizes them. This makes it easier to use
hasOneUse() heuristics, since unneeded nodes can be removed from the
DAG earlier.
Make LegalizeOps visit the DAG in an operands-last order. It previously
used operands-first, because LegalizeTypes has to go operands-first, and
LegalizeTypes used to be part of LegalizeOps, but they're now split.
The operands-last order is more natural for several legalization tasks.
For example, it allows lowering code for nodes with floating-point or
vector constants to see those constants directly instead of seeing the
lowered form (often constant-pool loads). This makes some things
somewhat more complicated today, though it ought to allow things to be
simpler in the future. It also fixes some bugs exposed by Legalizing
using RAUW aggressively.
Remove the part of LegalizeOps that attempted to patch up invalid chain
operands on libcalls generated by LegalizeTypes, since it doesn't work
with the new LegalizeOps traversal order. Instead, define what
LegalizeTypes is doing to be correct, and transfer the responsibility
of keeping calls from having overlapping calling sequences into the
scheduler.
Teach the scheduler to model callseq_begin/end pairs as having a
physical register definition/use to prevent calls from having
overlapping calling sequences. This is also somewhat complicated, though
there are ways it might be simplified in the future.
This addresses rdar://9816668, rdar://10043614, rdar://8434668, and others.
Please direct high-level questions about this patch to management.
llvm-svn: 143177
http://lab.llvm.org:8011/builders/llvm-x86_64-linux/builds/101
--- Reverse-merging r141854 into '.':
U test/MC/Disassembler/X86/x86-32.txt
U test/MC/Disassembler/X86/simple-tests.txt
D test/CodeGen/X86/bmi.ll
U lib/Target/X86/X86InstrInfo.td
U lib/Target/X86/X86ISelLowering.cpp
U lib/Target/X86/X86.td
U lib/Target/X86/X86Subtarget.h
llvm-svn: 141857
- x87: no min or max.
- SSE1: min/max for single precision scalars and vectors.
- SSE2: min/max for single and double precision scalars and vectors.
- AVX: as SSE2, but also supports the wider ymm vectors. (this is covered by the isTypeLegal check)
llvm-svn: 140296
dag-combine optimization to implement the ext-load efficiently (using shuffles).
For example the type <4 x i8> is stored in memory as i32, but it needs to
find its way into a <4 x i32> register. Previously we scalarized the memory
access, now we use shuffles.
llvm-svn: 139995
maxps and maxpd). This broke the sse41-blend.ll testcase by causing
maxpd to be produced rather than a cmp+blend pair, which is the reason
I tweaked it. Gives a small speedup on doduc with dragonegg when the
GCC vectorizer is used.
llvm-svn: 139986
take into consideration the presence of AVX. This change, together with
the SSEDomainFix enabled for AVX, makes AVX codegen to always (hopefully)
emit the same code as SSE for 128-bit vector ops. I don't
have a testcase for this, but AVX now beats SSE in performance for
128-bit ops in the majority of programas in the llvm testsuite
llvm-svn: 139817
However with this fix it does now.
Basically the operand order for the x86 target specific node
is not the same as the instruction, but since the intrinsic need that
specific order at the instruction definition, just change the order
during legalization. Also, there were some wrong invertions of condition
codes, such as GE => LE, GT => LT, fix that too. Fix PR10907.
llvm-svn: 139528
assert("not implemented for target shuffle node");
to:
assert(0 && "not implemented for target shuffle node");
This causes a test failure in CodeGen/X86/palignr.ll which has
been marked as XFAIL for the time being.
Test failure filed at PR10901.
llvm-svn: 139454
in Nadav's r139285 and r139287 commits.
1) Rename vsel.ll to a more descriptive name
2) Change the order of BLEND operands to "Op1, Op2, Cond", this is
necessary because PBLENDVB is already used in different places with
this order, and it was being emitted in the wrong way for vselect
3) Add AVX patterns and tests for the same SSE41 instructions
llvm-svn: 139305
with a vector condition); such selects become VSELECT codegen nodes.
This patch also removes VSETCC codegen nodes, unifying them with SETCC
nodes (codegen was actually often using SETCC for vector SETCC already).
This ensures that various DAG combiner optimizations kick in for vector
comparisons. Passes dragonegg bootstrap with no testsuite regressions
(nightly testsuite as well as "make check-all"). Patch mostly by
Nadav Rotem.
llvm-svn: 139159
init.trampoline and adjust.trampoline intrinsics, into two intrinsics
like in GCC. While having one combined intrinsic is tempting, it is
not natural because typically the trampoline initialization needs to
be done in one function, and the result of adjust trampoline is needed
in a different (nested) function. To get around this llvm-gcc hacks the
nested function lowering code to insert an additional parent variable
holding the adjust.trampoline result that can be accessed from the child
function. Dragonegg doesn't have the luxury of tweaking GCC code, so it
stored the result of adjust.trampoline in the memory GCC set aside for
the trampoline itself (this is always available in the child function),
and set up some new memory (using an alloca) to hold the trampoline.
Unfortunately this breaks Go which allocates trampoline memory on the
heap and wants to use it even after the parent has exited (!). Rather
than doing even more hacks to get Go working, it seemed best to just use
two intrinsics like in GCC. Patch mostly by Sanjoy Das.
llvm-svn: 139140
- Duplicate some store patterns to their AVX forms!
- Catched a bug while restricting the patterns subtarget, fix it
and update a testcase to check it properly
llvm-svn: 138851
code is inserted to first check if the current stacklet has enough
space. If so, space is allocated by simply decrementing the stack
pointer. Otherwise a runtime routine (__morestack_allocate_stack_space
in libgcc) is called which allocates the required memory from the
heap.
Patch by Sanjoy Das.
llvm-svn: 138818
from DYNAMIC_STACKALLOC.
Two new pseudo instructions (SEG_ALLOCA_32 and SEG_ALLOCA_64) which
will match X86SegAlloca (based on word size) are also added. They
will be custom emitted to inject the actual stack handling code.
Patch by Sanjoy Das.
llvm-svn: 138814
X86. Modify the pass added in the previous patch to call this new
code.
This new prologues generated will call a libgcc routine (__morestack)
to allocate more stack space from the heap when required
Patch by Sanjoy Das.
llvm-svn: 138812
explicit about which subtarget they refer to, and add AVX versions of
the ones we currently don't. Make the mask check more strict, to be
clear it won't be used to match to 256-bit versions!
llvm-svn: 138514
match splats in the form (splat (scalar_to_vector (load ...))) whenever
the load can be folded. All the logic and instruction emission is
working but because of PR8156, there are no ways to match loads, cause
they can never be folded for splats. Thus, the tests are XFAILed, but
I've tested and exercised all the logic using a relaxed version for
checking the foldable loads, as if the bug was already fixed. This
should work out of the box once PR8156 gets fixed since MayFoldLoad will
work as expected.
llvm-svn: 137810
vinsertf128 $1 + vpermilps $0, remove the old code that used to first
do the splat in a 128-bit vector and then insert it into a larger one.
This is better because the handling code gets simpler and also makes a
better room for the upcoming vbroadcast!
llvm-svn: 137807
there is no support for native 256-bit shuffles, be more smart in some
cases, for example, when you can extract specific 128-bit parts and use
regular 128-bit shuffles for them. Example:
For this shuffle:
shufflevector <4 x i64> %a, <4 x i64> %b, <4 x i32>
<i32 1, i32 0, i32 7, i32 6>
This was expanded to:
vextractf128 $1, %ymm1, %xmm2
vpextrq $0, %xmm2, %rax
vmovd %rax, %xmm1
vpextrq $1, %xmm2, %rax
vmovd %rax, %xmm2
vpunpcklqdq %xmm1, %xmm2, %xmm1
vpextrq $0, %xmm0, %rax
vmovd %rax, %xmm2
vpextrq $1, %xmm0, %rax
vmovd %rax, %xmm0
vpunpcklqdq %xmm2, %xmm0, %xmm0
vinsertf128 $1, %xmm1, %ymm0, %ymm0
ret
Now we get:
vshufpd $1, %xmm0, %xmm0, %xmm0
vextractf128 $1, %ymm1, %xmm1
vshufpd $1, %xmm1, %xmm1, %xmm1
vinsertf128 $1, %xmm1, %ymm0, %ymm0
llvm-svn: 137733
vectors. It operates on 128-bit elements instead of regular scalar
types. Recognize shuffles that are suitable for VPERM2F128 and teach
the x86 legalizer how to handle them.
llvm-svn: 137519
(for example, after integer operation), do not pack the registers into a YMM
before saving. Its better to save as two XMM registers.
Before:
vinsertf128 $1, %xmm3, %ymm0, %ymm3
vinsertf128 $0, %xmm1, %ymm3, %ymm1
vmovaps %ymm1, 416(%rsp)
After:
vmovaps %xmm3, 416+16(%rsp)
vmovaps %xmm1, 416(%rsp)
llvm-svn: 137308
data in-register prior to saving to memory. When we reorder the data in memory
we prevent the need to save multiple scalars to memory, making a single regular
store.
llvm-svn: 137238
avoid returning early for v8i32 types, which would only be valid for
vector with all zeros. Also split the handling of zeros and ones into separate
checking logic since they are handled differently. This fixes PR10547
llvm-svn: 136642
working on x86 (at least for trivial testcases); other architectures will
need more work so that they actually emit the appropriate instructions for
orderings stricter than 'monotonic'. (As far as I can tell, the ARM, PPC,
Mips, and Alpha backends need such changes.)
llvm-svn: 136457
usage of the shuffle bitmask. Both work in 128-bit lanes without
crossing, but in the former the mask of the high part is the same
used by the low part while in the later both lanes have independent
masks. Handle this properly and and add support for vpermilpd.
llvm-svn: 136200
On x86 we can't encode an immediate LHS of a sub directly. If the RHS comes from a XOR with a constant we can
fold the negation into the xor and add one to the immediate of the sub. Then we can turn the sub into an add,
which can be commuted and encoded efficiently.
This code is generated for __builtin_clz and friends.
llvm-svn: 136167
shuffle before inserting on a 256-bit vector.
- Add AVX versions of movd/movq instructions
- Introduce a few COPY patterns to match insert_subvector instructions.
This turns a trivial insert_subvector instruction into a register copy,
coalescing the xmm into a ymm and avoid emiting on more instruction.
llvm-svn: 136002
the way to go. Doing this here will prevent several node matches later,
and would have to force looking all the way through several
VINSERTF128/VEXTRACTF128 chains to optimize simple things.
llvm-svn: 135730
and was actually very wrong, fix it and make it simpler. Also remove the
ConcatVectors function, which is unused now.
- Fix a introduction of useless nodes in r126664 and r126264. The
VUNPCKL* should never be introduced cause we don't want duplicate
nodes for 128 AVX and non-AVX modes, the actual instruction
difference only exists during isel, but not for target specific DAG
nodes. We only introduce V* target nodes when there is no 128-bit
version already there.
- Fix a fragile test and make it more useful.
llvm-svn: 135729
instruction introduced in AVX, which can operate on 128 and 256-bit vectors.
It considers a 256-bit vector as two independent 128-bit lanes. It can permute
any 32 or 64 elements inside a lane, and restricts the second lane to
have the same permutation of the first one. With the improved splat support
introduced early today, adding codegen for this instruction enable more
efficient 256-bit code:
Instead of:
vextractf128 $0, %ymm0, %xmm0
punpcklbw %xmm0, %xmm0
punpckhbw %xmm0, %xmm0
vinsertf128 $0, %xmm0, %ymm0, %ymm1
vinsertf128 $1, %xmm0, %ymm1, %ymm0
vextractf128 $1, %ymm0, %xmm1
shufps $1, %xmm1, %xmm1
movss %xmm1, 28(%rsp)
movss %xmm1, 24(%rsp)
movss %xmm1, 20(%rsp)
movss %xmm1, 16(%rsp)
vextractf128 $0, %ymm0, %xmm0
shufps $1, %xmm0, %xmm0
movss %xmm0, 12(%rsp)
movss %xmm0, 8(%rsp)
movss %xmm0, 4(%rsp)
movss %xmm0, (%rsp)
vmovaps (%rsp), %ymm0
We get:
vextractf128 $0, %ymm0, %xmm0
punpcklbw %xmm0, %xmm0
punpckhbw %xmm0, %xmm0
vinsertf128 $0, %xmm0, %ymm0, %ymm1
vinsertf128 $1, %xmm0, %ymm1, %ymm0
vpermilps $85, %ymm0, %ymm0
llvm-svn: 135662
refactor the code and add a bunch of comments. The final shuffle
emitted by handling 256-bit types is suitable for the VPERM shuffle
instruction which is going to be introduced in a next commit (with
a testcase which cover this commit)
llvm-svn: 135661
to MCRegisterInfo. Also initialize the mapping at construction time.
This patch eliminate TargetRegisterInfo from TargetAsmInfo. It's another step
towards fixing the layering violation.
llvm-svn: 135424
1) Make non-legal 256-bit loads to be promoted to v4i64. This lets us
canonize the loads and handle things the same way we use to handle
for 128-bit registers. Despite of what one of the removed comments
explained, the load promotion would not mess with VPERM, it's only a
matter of doing the appropriate bitcasts when this instructions comes
to be introduced. Also make LOAD v8i32 legal.
2) Doing 1) exposed two bugs:
- v4i64 was being promoted to itself for several opcodes (introduced
in r124447 by David Greene) causing endless recursion and the stack to
explode.
- there was no support for allOnes BUILD_VECTORs and ANDNP would fail to
match because it was generating early target constant pools during
lowering.
3) The testcases are already checked-in, doing 1) exposed the
bugs in the current testcases.
4) Tidy up code to be more clear and explicit about AVX.
llvm-svn: 135313
when determining validity of matching constraint. Allow i1
types access to the GR8 reg class for x86.
Fixes PR10352 and rdar://9777108
llvm-svn: 135180
During type legalization we often use the SIGN_EXTEND_INREG SDNode.
When this SDNode is legalized during the LegalizeVector phase, it is
scalarized because non-simple types are automatically marked to be expanded.
In this patch we add support for lowering SIGN_EXTEND_INREG manually.
This fixes CodeGen/X86/vec_sext.ll when running with the '-promote-elements'
flag.
llvm-svn: 135144
Drop the FpMov instructions, use plain COPY instead.
Drop the FpSET/GET instruction for accessing fixed stack positions.
Instead use normal COPY to/from ST registers around inline assembly, and
provide a single new FpPOP_RETVAL instruction that can access the return
value(s) from a call. This is still necessary since you cannot tell from
the CALL instruction alone if it returns anything on the FP stack. Teach
fast isel to use this.
This provides a much more robust way of handling fixed stack registers -
we can tolerate arbitrary FP stack instructions inserted around calls
and inline assembly. Live range splitting could sometimes break x87 code
by inserting spill code in unfortunate places.
As a bonus we handle floating point inline assembly correctly now.
llvm-svn: 134018
optimizations when emitting calls to the function; instead those calls may
use faster relocations which require the function to be immediately resolved
upon loading the dynamic object featuring the call. This is useful when it
is known that the function will be called frequently and pervasively and
therefore there is no merit in delaying binding of the function.
Currently only implemented for x86-64, where it turns into a call through
the global offset table.
Patch by Dan Gohman, who assures me that he's going to add LangRef documentation
for this once it's committed.
llvm-svn: 133080
floating-point comparison, generate a mask of 0s or 1s, and generally
DTRT with NaNs. Only profitable when the user wants a materialized 0
or 1 at runtime. rdar://problem/5993888
llvm-svn: 132404
non-zero.
- Teach X86 cmov optimization to eliminate the cmov from ctlz, cttz extension
when the source of X86ISD::BSR / X86ISD::BSF is proven to be non-zero.
rdar://9490949
llvm-svn: 131948
to have single return block (at least getting there) for optimizations. This
is general goodness but it would prevent some tailcall optimizations.
One specific case is code like this:
int f1(void);
int f2(void);
int f3(void);
int f4(void);
int f5(void);
int f6(void);
int foo(int x) {
switch(x) {
case 1: return f1();
case 2: return f2();
case 3: return f3();
case 4: return f4();
case 5: return f5();
case 6: return f6();
}
}
=>
LBB0_2: ## %sw.bb
callq _f1
popq %rbp
ret
LBB0_3: ## %sw.bb1
callq _f2
popq %rbp
ret
LBB0_4: ## %sw.bb3
callq _f3
popq %rbp
ret
This patch teaches codegenprep to duplicate returns when the return value
is a phi and where the phi operands are produced by tail calls followed by
an unconditional branch:
sw.bb7: ; preds = %entry
%call8 = tail call i32 @f5() nounwind
br label %return
sw.bb9: ; preds = %entry
%call10 = tail call i32 @f6() nounwind
br label %return
return:
%retval.0 = phi i32 [ %call10, %sw.bb9 ], [ %call8, %sw.bb7 ], ... [ 0, %entry ]
ret i32 %retval.0
This allows codegen to generate better code like this:
LBB0_2: ## %sw.bb
jmp _f1 ## TAILCALL
LBB0_3: ## %sw.bb1
jmp _f2 ## TAILCALL
LBB0_4: ## %sw.bb3
jmp _f3 ## TAILCALL
rdar://9147433
llvm-svn: 127953
not have native support for this operation (such as X86).
The legalized code uses two vector INT_TO_FP operations and is faster
than scalarizing.
llvm-svn: 127951
comparisons on x86. Essentially, the way this works is that SUB+SBB sets
the relevant flags the same way a double-width CMP would.
This is a substantial improvement over the generic lowering in LLVM. The output
is also shorter than the gcc-generated output; I haven't done any detailed
benchmarking, though.
llvm-svn: 127852
rather than an int. Thankfully, this only causes LLVM to miss optimizations, not
generate incorrect code.
This just fixes the zext at the return. We still insert an i32 ZextAssert when
reading a function's arguments, but it is followed by a truncate and another i8
ZextAssert so it is not optimized.
llvm-svn: 127766
testcases accordingly. Some are currently xfailed and will be filed
as bugs to be fixed or understood.
Performance results:
roughly neutral on SPEC
some micro benchmarks in the llvm suite are up between 100 and 150%, only
a pair of regressions that are due to be investigated
john-the-ripper saw:
10% improvement in traditional DES
8% improvement in BSDI DES
59% improvement in FreeBSD MD5
67% improvement in OpenBSD Blowfish
14% improvement in LM DES
Small compile time impact.
llvm-svn: 127208
regs. This is the only change in this checkin that may affects the
default scheduler. With better register tracking and heuristics, it
doesn't make sense to artificially lower the register limit so much.
Added -sched-high-latency-cycles and X86InstrInfo::isHighLatencyDef to
give the scheduler a way to account for div and sqrt on targets that
don't have an itinerary. It is currently defaults to 10 (the actual
number doesn't matter much), but only takes effect on non-default
schedulers: list-hybrid and list-ilp.
Added several heuristics that can be individually disabled for the
non-default sched=list-ilp mode. This helps us determine how much
better we can do on a given benchmark than the default
scheduler. Certain compute intensive loops run much faster in this
mode with the right set of heuristics, and it doesn't seem to have
much negative impact elsewhere. Not all of the heuristics are needed,
but we still need to experiment to decide which should be disabled by
default for sched=list-ilp.
llvm-svn: 127067
and 256-bit forms. Because the number of elements in a vector
does not determine the vector type (4 elements could be v4f32 or
v4f64), pass the full type of the vector to decode routines.
llvm-svn: 126664
In other words, do not keep track of argument's location. The debugger (gdb) is not prepared to see line table entries for arguments. For the debugger, "second" line table entry marks beginning of function body.
This requires some coordination with debugger to get this working.
- The debugger needs to be aware of prolog_end attribute attached with line table entries.
- The compiler needs to accurately mark prolog_end in line table entries (at -O0 and at -O1+)
llvm-svn: 126155
(LLVMX86Utils.a) to break cyclic library dependencies between
LLVMX86CodeGen.a and LLVMX86AsmParser.a. Previously this code was in
a header file and marked static but AVX requires some additional
functionality here that won't be used by all clients. Since including
unused static functions causes a gcc compiler warning, keeping it as a
header would break builds that use -Werror. Putting this in its own
library solves both problems at once.
llvm-svn: 125765
have their low bits set to zero. This allows us to optimize
out explicit stack alignment code like in stack-align.ll:test4 when
it is redundant.
Doing this causes the code generator to start turning FI+cst into
FI|cst all over the place, which is general goodness (that is the
canonical form) except that various pieces of the code generator
don't handle OR aggressively. Fix this by introducing a new
SelectionDAG::isBaseWithConstantOffset predicate, and using it
in places that are looking for ADD(X,CST). The ARM backend in
particular was missing a lot of addressing mode folding opportunities
around OR.
llvm-svn: 125470
This allows us to easily support 256-bit operations that don't have
native 256-bit support. This applies to integer operations, certain
types of shuffles and various othher things.
llvm-svn: 124910
matching EXTRACT_SUBVECTOR to VEXTRACTF128 along with support routines
to examine and translate index values. VINSERTF128 comes next. With
these two in place we can begin supporting more AVX operations as
INSERT/EXTRACT can be used as a fallback when 256-bit support is not
available.
llvm-svn: 124797
Reversing the operands allows us to fold, but doesn't force us to. Also, at
this point the DAG is still being optimized, so the check for hasOneUse is not
very precise.
llvm-svn: 124773
default implementation for x86, going through the stack in a similr
fashion to how the codegen implements BUILD_VECTOR. Eventually this
will get matched to VINSERTF128 if AVX is available.
llvm-svn: 124307
implementation of EXTRACT_SUBVECTOR for x86, going through the stack
in a similr fashion to how the codegen implements BUILD_VECTOR.
Eventually this will get matched to VEXTRACTF128 if AVX is available.
llvm-svn: 124292
The theory is it's still faster than a pair of movq / a quad of movl. This
will probably hurt older chips like P4 but should run faster on current
and future Intel processors. rdar://8817010
llvm-svn: 122955
the same as setcc. Optimize ADDC(0,0,FLAGS) -> SET_CARRY(FLAGS). This is
a step towards finishing off PR5443. In the testcase in that bug we now get:
movq %rdi, %rax
addq %rsi, %rax
sbbq %rcx, %rcx
testb $1, %cl
setne %dl
ret
instead of:
movq %rdi, %rax
addq %rsi, %rax
movl $0, %ecx
adcq $0, %rcx
testq %rcx, %rcx
setne %dl
ret
llvm-svn: 122219
their carry depenedencies with MVT::Flag operands) and use clean and beautiful
EFLAGS dependences instead.
We do this by changing the modelling of SBB/ADC to have EFLAGS input and outputs
(which is what requires the previous scheduler change) and change X86 ISelLowering
to custom lower ADDC and friends down to X86ISD::ADD/ADC/SUB/SBB nodes.
With the previous series of changes, this causes no changes in the testsuite, woo.
llvm-svn: 122213
the output to the correct register. Fixes a hidden problem uncovered
by the last patch where we'd try to DAG combine our MVT::Other node
oddly.
llvm-svn: 121358
result. This allows us to compile:
void *test12(long count) {
return new int[count];
}
into:
test12:
movl $4, %ecx
movq %rdi, %rax
mulq %rcx
movq $-1, %rdi
cmovnoq %rax, %rdi
jmp __Znam ## TAILCALL
instead of:
test12:
movl $4, %ecx
movq %rdi, %rax
mulq %rcx
seto %cl
testb %cl, %cl
movq $-1, %rdi
cmoveq %rax, %rdi
jmp __Znam
Of course it would be even better if the regalloc inverted the cmov to 'cmovoq',
which would eliminate the need for the 'movq %rdi, %rax'.
llvm-svn: 120936
backend that they were all implemented except umul. This one fell back
to the default implementation that did a hi/lo multiply and compared the
top. Fix this to check the overflow flag that the 'mul' instruction
sets, so we can avoid an explicit test. Now we compile:
void *func(long count) {
return new int[count];
}
into:
__Z4funcl: ## @_Z4funcl
movl $4, %ecx ## encoding: [0xb9,0x04,0x00,0x00,0x00]
movq %rdi, %rax ## encoding: [0x48,0x89,0xf8]
mulq %rcx ## encoding: [0x48,0xf7,0xe1]
seto %cl ## encoding: [0x0f,0x90,0xc1]
testb %cl, %cl ## encoding: [0x84,0xc9]
movq $-1, %rdi ## encoding: [0x48,0xc7,0xc7,0xff,0xff,0xff,0xff]
cmoveq %rax, %rdi ## encoding: [0x48,0x0f,0x44,0xf8]
jmp __Znam ## TAILCALL
instead of:
__Z4funcl: ## @_Z4funcl
movl $4, %ecx ## encoding: [0xb9,0x04,0x00,0x00,0x00]
movq %rdi, %rax ## encoding: [0x48,0x89,0xf8]
mulq %rcx ## encoding: [0x48,0xf7,0xe1]
testq %rdx, %rdx ## encoding: [0x48,0x85,0xd2]
movq $-1, %rdi ## encoding: [0x48,0xc7,0xc7,0xff,0xff,0xff,0xff]
cmoveq %rax, %rdi ## encoding: [0x48,0x0f,0x44,0xf8]
jmp __Znam ## TAILCALL
Other than the silly seto+test, this is using the o bit directly, so it's going in the right
direction.
llvm-svn: 120935
The user (i.e. whoever generated a call to the intrinsic in the first place) is
essentially asking for a particular instruction to be placed in the assembler.
If that instruction won't execute on the target machine, that's their problem
not ours. Two buildbots with processors that don't support SSE3 were barfing
on the apm.ll test in CodeGen/X86 because of this assertion.
llvm-svn: 120574
legalization time. Since at legalization time there is no mapping from
SDNode back to the corresponding LLVM instruction and the return
SDNode is target specific, this requires a target hook to check for
eligibility. Only x86 and ARM support this form of sibcall optimization
right now.
rdar://8707777
llvm-svn: 120501
nodes to indicate when ha16/lo16 modifiers should be used. This lets
us pass PowerPC/indirectbr.ll.
The one annoying thing about this patch is that the MCSymbolExpr isn't
expressive enough to represent ha16(label1-label2) which we need on
PowerPC. I have a terrible hack in the meantime, but this will have
to be revisited at some point.
Last major conversion item left is global variable references.
llvm-svn: 119105
The x86_mmx type is used for MMX intrinsics, parameters and
return values where these use MMX registers, and is also
supported in load, store, and bitcast.
Only the above operations generate MMX instructions, and optimizations
do not operate on or produce MMX intrinsics.
MMX-sized vectors <2 x i32> etc. are lowered to XMM or split into
smaller pieces. Optimizations may occur on these forms and the
result casted back to x86_mmx, provided the result feeds into a
previous existing x86_mmx operation.
The point of all this is prevent optimizations from introducing
MMX operations, which is unsafe due to the EMMS problem.
llvm-svn: 115243
ARM cross-compiler on x86, because the MMO size did not match the type size.
This fixes the MMO size and also the size of the stack object to match the
type size.
llvm-svn: 114554
(sbbl x, x) sets the registers to 0 or ~0. Combined with two's complement arithmetic, we can fold
the intermediate AND and the ADD into a single SUB.
This fixes <rdar://problem/8449754>.
llvm-svn: 114460
"getFixedStack" on the MachinePointerInfo class. While
this isn't the problem I'm setting out to solve, it is the
right way to eliminate PseudoSourceValue, so lets go with it.
llvm-svn: 114406
nodes to emit shuffles and don't do isel mask matching anymore.
- Add the selection of the remaining shuffle opcode (movddup)
- Introduce two new functions to "recognize" where we may get
potential folds and add several comments to them explaining why
they are not yet in the desidered shape.
- Add more patterns to fallback the case where we select
a specific shuffle opcode as if it could fold a load, but it
can't, so remap to a valid instruction.
- Add a couple of FIXMEs to address in the following days once
there's a good solution to the current folding problem.
llvm-svn: 113369
NAKAMURA Takumi