cases from Halide folks. This initial step was extracted from
a prototype change by Clay Wood to try and address regressions found
with Halide and the new vector shuffle lowering.
llvm-svn: 221779
With this patch MCDisassembler::getInstruction takes an ArrayRef<uint8_t>
instead of a MemoryObject.
Even on X86 there is a maximum size an instruction can have. Given
that, it seems way simpler and more efficient to just pass an ArrayRef
to the disassembler instead of a MemoryObject and have it do a virtual
call every time it wants some extra bytes.
llvm-svn: 221751
This commit adds a new pass that can inject checks before indirect calls to
make sure that these calls target known locations. It supports three types of
checks and, at compile time, it can take the name of a custom function to call
when an indirect call check fails. The default failure function ignores the
error and continues.
This pass incidentally moves the function JumpInstrTables::transformType from
private to public and makes it static (with a new argument that specifies the
table type to use); this is so that the CFI code can transform function types
at call sites to determine which jump-instruction table to use for the check at
that site.
Also, this removes support for jumptables in ARM, pending further performance
analysis and discussion.
Review: http://reviews.llvm.org/D4167
llvm-svn: 221708
This is a first step for generating SSE rcp instructions for reciprocal
calcs when fast-math allows it. This is very similar to the rsqrt optimization
enabled in D5658 ( http://reviews.llvm.org/rL220570 ).
For now, be conservative and only enable this for AMD btver2 where performance
improves significantly both in terms of latency and throughput.
We may never enable this codegen for Intel Core* chips because the divider circuits
are just too fast. On SandyBridge, divss can be as fast as 10 cycles versus the 21
cycle critical path for the rcp + mul + sub + mul + add estimate.
Follow-on patches may allow configuration of the number of Newton-Raphson refinement
steps, add AVX512 support, and enable the optimization for more chips.
More background here: http://llvm.org/bugs/show_bug.cgi?id=21385
Differential Revision: http://reviews.llvm.org/D6175
llvm-svn: 221706
The ISel lowering for global TLS access in PIC mode was creating a pseudo
instruction that is later expanded to a call, but the code was not
setting the hasCalls flag in the MachineFrameInfo alongside the adjustsStack
flag. This caused some functions to be mistakenly recognized as leaf functions,
and this in turn affected the decision to eliminate the frame pointer.
With the fix, hasCalls is properly set and the leaf frame pointer is correctly
preserved.
llvm-svn: 221695
This fixes an issue with matching trunc -> assertsext -> zext on x86-64, which would not zero the high 32-bits. See PR20494 for details.
Recommitting - This time, with a hopefully working test.
Differential Revision: http://reviews.llvm.org/D6128
llvm-svn: 221672
AVX2 is available.
According to IACA, the new lowering has a throughput of 8 cycles instead of 13
with the previous one.
Althought this lowering kicks in some SPECs benchmarks, the performance
improvement was within the noise.
Correctness testing has been done for the whole range of uint32_t with the
following program:
uint4 v = (uint4) {0,1,2,3};
uint32_t i;
//Check correctness over entire range for uint4 -> float4 conversion
for( i = 0; i < 1U << (32-2); i++ )
{
float4 t = test(v);
float4 c = correct(v);
if( 0xf != _mm_movemask_ps( t == c ))
{
printf( "Error @ %vx: %vf vs. %vf\n", v, c, t);
return -1;
}
v += 4;
}
Where "correct" is the old lowering and "test" the new one.
The patch adds a test case for the two custom lowering instruction.
It also modifies the vector cost model, which is why cast.ll and uitofp.ll are
modified.
2009-02-26-MachineLICMBug.ll is also modified because we now hoist 7
instructions instead of 4 (3 more constant loads).
rdar://problem/18153096>
llvm-svn: 221657
This fixes an issue with matching trunc -> assertsext -> zext on x86-64, which would not zero the high 32-bits.
See PR20494 for details.
Differential Revision: http://reviews.llvm.org/D6128
llvm-svn: 221626
This fixes a few cases of:
* Wrong variable name style.
* Lines longer than 80 columns.
* Repeated names in comments.
* clang-format of the above.
This make the next patch a lot easier to read.
llvm-svn: 221615
Fixed an issue with the (v)cvttps2dq and (v)cvttpd2dq instructions being incorrectly put in the 2 source operand folding tables instead of the 1 source operand and added the missing SSE/AVX versions.
Also added missing (v)cvtps2dq and (v)cvtpd2dq instructions to the folding tables.
Differential Revision: http://reviews.llvm.org/D6001
llvm-svn: 221489
Example:
define <4 x i32> @test(<4 x i32> %a, <4 x i32> %b) {
%shuffle = shufflevector <4 x i32> %a, <4 x i32> %b, <4 x i32> <i32 4, i32 5, i32 6, i32 3>
ret <4 x i32> %shuffle
}
Before llc (-mattr=+sse4.1), produced the following assembly instruction:
pblendw $4294967103, %xmm1, %xmm0
After
pblendw $63, %xmm1, %xmm0
llvm-svn: 221455
condition to match a blend.
This prevents optimizations that work on VSELECT to perform invalid
transformations. Indeed, the optimized condition does not match the vector
boolean content that is expected and bad things may happen.
This patch yields the exact same code on the whole test-suite + specs (-O3 and
-O3 -march=core-avx2), it improves one test case (vector-blend.ll) and fixes a
bug reduced in vselect-avx.ll.
<rdar://problem/18819506>
llvm-svn: 221429
Added missing memory folding for the (V)CVTDQ2PS instructions - we can safely fold these (but not the (V)CVTDQ2PD versions which have a register/memory size discrepancy in the source operand). I've added a test case demonstrating that stack folding now works.
Differential Revision: http://reviews.llvm.org/D5981
llvm-svn: 221407
Summary:
X86FastISel::fastMaterializeAlloca was incorrectly conditioning its
opcode selection on subtarget bitness rather than pointer size.
Differential Revision: http://reviews.llvm.org/D6136
llvm-svn: 221386
This patch improves the folding of vector AND nodes into blend operations for
targets that feature SSE4.1. A vector AND node where one of the operands is
a constant build_vector with elements that are either zero or all-ones can be
converted into a blend.
This allows for example to simplify the following code:
define <4 x i32> @test(<4 x i32> %A, <4 x i32> %B) {
%1 = and <4 x i32> %A, <i32 0, i32 0, i32 0, i32 -1>
%2 = and <4 x i32> %B, <i32 -1, i32 -1, i32 -1, i32 0>
%3 = or <4 x i32> %1, %2
ret <4 x i32> %3
}
Before this patch llc (-mcpu=corei7) generated:
andps LCPI1_0(%rip), %xmm0, %xmm0
andps LCPI1_1(%rip), %xmm1, %xmm1
orps %xmm1, %xmm0, %xmm0
retq
With this patch we generate a single 'vpblendw'.
llvm-svn: 221343
Patch to allow (v)blendps, (v)blendpd, (v)pblendw and vpblendd instructions to be commuted - swaps the src registers and inverts the blend mask.
This is primarily to improve memory folding (see new tests), but it also improves the quality of shuffles (see modified tests).
Differential Revision: http://reviews.llvm.org/D6015
llvm-svn: 221313
This patch adds 'FeatureSlowSHLD' to 'bdver3'.
According to the official AMD optimization guide for amdfam15: "Using
alternative code in place of SHLD achieves lower overall latency and
requires fewer execution resources. The 32-bit and 64-bit forms of
ADD, ADC, SHR, and LEA (except 16-bit form) are DirectPath
instructions, while SHLD is a VectorPath instruction."
This patch also explicitly sets feature AVX and SSE4A for all the bdver*
cpus. This part of the patch is a non-functional change and it is mainly
done for clarity reasons (Both XOP and FMA4 already imply AVX and SSE4A).
llvm-svn: 221296
For 8-bit divrems where the remainder is used, we used to generate:
divb %sil
shrw $8, %ax
movzbl %al, %eax
That was to avoid an H-reg access, which is problematic mainly because
it isn't possible in REX-prefixed instructions.
This patch optimizes that to:
divb %sil
movzbl %ah, %eax
To do that, we explicitly extend AH, and extract the L-subreg in the
resulting register. The extension is done using the NOREX variants of
MOVZX. To support signed operations, MOVSX_NOREX is also added.
Further, this introduces a new SDNode type, [us]divrem_ext_hreg, which is
then lowered to a sequence containing a single zext (rather than 2).
Differential Revision: http://reviews.llvm.org/D6064
llvm-svn: 221176
This removes calls to isMaterializable in the following cases:
* It was redundant with a call to isDeclaration now that isDeclaration returns
the correct answer for materializable functions.
* It was followed by a call to Materialize. Just call Materialize and check EC.
llvm-svn: 221050
"[x86] Simplify vector selection if condition value type matches vselect value type and true value is all ones or false value is all zeros."
llvm-svn: 221028
It appears to ignore or find ambiguous MachineInstrBuilder's conversion
operators that allow conversion to MachineInstr* and
MachineBasicBlock::bundle_iterator.
As a workaround, add an explicit way to get the MachineInstr.
llvm-svn: 221017
This transformation worked if selector is produced by SETCC, however SETCC is needed only if we consider to swap operands. So I replaced SETCC check for this case.
Added tests for vselect of <X x i1> values.
llvm-svn: 220777
Ffter commit at rev219046 512-bit broadcasts lowering become non-optimal. Most of tests on broadcasting and embedded broadcasting were changed and they doesn’t produce efficient code.
Example below is from commit changes (it’s the first test from test/CodeGen/X86/avx512-vbroadcast.ll):
define <16 x i32> @_inreg16xi32(i32 %a) {
; CHECK-LABEL: _inreg16xi32:
; CHECK: ## BB#0:
-; CHECK-NEXT: vpbroadcastd %edi, %zmm0
+; CHECK-NEXT: vmovd %edi, %xmm0
+; CHECK-NEXT: vpbroadcastd %xmm0, %ymm0
+; CHECK-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; CHECK-NEXT: retq
%b = insertelement <16 x i32> undef, i32 %a, i32 0
%c = shufflevector <16 x i32> %b, <16 x i32> undef, <16 x i32> zeroinitializer
ret <16 x i32> %c
}
Here, 256-bit broadcast was generated instead of 512-bit one.
In this patch
1) I added vector-shuffle lowering through broadcasts
2) Removed asserts and branches likes because this is incorrect
- assert(Subtarget->hasDQI() && "We can only lower v8i64 with AVX-512-DQI");
3) Fixed lowering tests
llvm-svn: 220774
This is a Microsoft calling convention that supports both x86 and x86_64
subtargets. It passes vector and floating point arguments in XMM0-XMM5,
and passes them indirectly once they are consumed.
Homogenous vector aggregates of up to four elements can be passed in
sequential vector registers, but this part is not implemented in LLVM
and will be handled in Clang.
On 32-bit x86, it is similar to fastcall in that it uses ecx:edx as
integer register parameters and is callee cleanup. On x86_64, it
delegates to the normal win64 calling convention.
Reviewers: majnemer
Differential Revision: http://reviews.llvm.org/D5943
llvm-svn: 220745
No functionality change. No change in X86.td.expanded except that we only set
the CD8 attributes for the memory variants. (This shouldn't be used unless we
have a memory operand.)
llvm-svn: 220736
1) i512mem -> f512mem (this is the packed FP input being permuted)
2) element size is 64 bits in EVEX_CD8 for PD.
(A good illustration why X86VectorVTInfo is useful)
llvm-svn: 220734
For a call to not return in to the stackmap shadow, the shadow must end with the call.
To do this, we must insert any required nops *before* the call, and not after it.
llvm-svn: 220728
To avoid emitting too many nops, a stackmap shadow can include emitted instructions in the shadow, but these must not include branch targets.
A return from a call should count as a branch target as patching over the instructions after the call would lead to incorrect behaviour for threads currently making that call, when they return.
llvm-svn: 220710
Tidied up some entries in the folding tables so that they are under the correct comment section (they were categorised as AVX2 instructions when they're AVX1).
Minor patch agreed with qcolombet.
llvm-svn: 220613
In a Mach-O object file a relocatable expression of the form
SymbolA - SymbolB + constant is allowed when both symbols are
defined in a section. But when either symbol is undefined it
is an error.
The code was crashing when it had an undefined symbol in this case.
And should have printed a error message using the location information
in the relocation entry.
rdar://18678402
llvm-svn: 220599
Minor patch to fix an issue in XFormVExtractWithShuffleIntoLoad where a load is unary shuffled, then bitcast (to a type with the same number of elements) before extracting an element.
An undef was created for the second shuffle operand using the original (post-bitcasted) vector type instead of the pre-bitcasted type like the rest of the shuffle node - this was then causing an assertion on the different types later on inside SelectionDAG::getVectorShuffle.
Differential Revision: http://reviews.llvm.org/D5917
llvm-svn: 220592
This is a first step for generating SSE rsqrt instructions for
reciprocal square root calcs when fast-math is allowed.
For now, be conservative and only enable this for AMD btver2
where performance improves significantly - for example, 29%
on llvm/projects/test-suite/SingleSource/Benchmarks/BenchmarkGame/n-body.c
(if we convert the data type to single-precision float).
This patch adds a two constant version of the Newton-Raphson
refinement algorithm to DAGCombiner that can be selected by any target
via a parameter returned by getRsqrtEstimate()..
See PR20900 for more details:
http://llvm.org/bugs/show_bug.cgi?id=20900
Differential Revision: http://reviews.llvm.org/D5658
llvm-svn: 220570
This is asm/diasm-only support, similar to AVX.
For ISeling the register variant, they are no different from 213 other than
whether the multiplication or the addition operand is destructed.
For ISeling the memory variant, i.e. to fold a load, they are no different
than the 132 variant. The addition operand (op3) in both cases can come from
memory. Again the ony difference is which operand is destructed.
There could be a post-RA pass that would convert a 213 or 132 into a 231.
Part of <rdar://problem/17082571>
llvm-svn: 220540
This multiclass generates the different forms: 213, 231, 132 in AVX.
132 in AVX512 is a separate class but I am planning to use this same
multiclass to generate 231 relying on the nice the null_frag trick from AVX to
disable codegen pattern for 231.
No functionality change, no change in X86.td.expanded except for the different
instruction definition names.
llvm-svn: 220539
Currently, @llvm.smul.with.overflow.i8 expands to 9 instructions, where
3 are really needed.
This adds X86ISD::UMUL8/SMUL8 SD nodes, and custom lowers them to
MUL8/IMUL8 + SETO.
i8 is a special case because there is no two/three operand variants of
(I)MUL8, so the first operand and return value need to go in AL/AX.
Also, we can't write patterns for these instructions: TableGen refuses
patterns where output operands don't match SDNode results. In this case,
instructions where the output operand is an implicitly defined register.
A related special case (and FIXME) exists for MUL8 (X86InstrArith.td):
// FIXME: Used for 8-bit mul, ignore result upper 8 bits.
// This probably ought to be moved to a def : Pat<> if the
// syntax can be accepted.
[(set AL, (mul AL, GR8:$src)), (implicit EFLAGS)]
Ideally, these go away with UMUL8, but we still need to improve TableGen
support of implicit operands in patterns.
Before this change:
movsbl %sil, %eax
movsbl %dil, %ecx
imull %eax, %ecx
movb %cl, %al
sarb $7, %al
movzbl %al, %eax
movzbl %ch, %esi
cmpl %eax, %esi
setne %al
After:
movb %dil, %al
imulb %sil
seto %al
Also, remove a made-redundant testcase for PR19858, and enable more FastISel
ALU-overflow tests for SelectionDAG too.
Differential Revision: http://reviews.llvm.org/D5809
llvm-svn: 220516
Summary: Fixed memory accesses with rbp as a base or an index register.
Reviewers: eugenis
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5819
llvm-svn: 220283
Every target we support has support for assembly that looks like
a = b - c
.long a
What is special about MachO is that the above combination suppresses the
production of a relocation.
With this change we avoid producing the intermediary labels when they don't
add any value.
llvm-svn: 220256
X86 code to lower VSELECT messed a bit with the bits set in the mask of VSELECT
when it knows it can be lowered into BLEND. Indeed, only the high bits need to be
set for those and it optimizes those accordingly.
However, when the mask is a compile time constant, the lowering will be handled
by the generic optimizer and those modifications will generate bad code in the
generic optimizer.
This patch fixes that by preventing the optimization if the VSELECT will be
handled by the generic optimizer.
<rdar://problem/18675020>
llvm-svn: 220242
This patch improves support for commutative instructions in the x86 memory folding implementation by attempting to fold a commuted version of the instruction if the original folding fails - if that folding fails as well the instruction is 're-commuted' back to its original order before returning.
Updated version of r219584 (reverted in r219595) - the commutation attempt now explicitly ensures that neither of the commuted source operands are tied to the destination operand / register, which was the source of all the regressions that occurred with the original patch attempt.
Added additional regression test case provided by Joerg Sonnenberger.
Differential Revision: http://reviews.llvm.org/D5818
llvm-svn: 220239
When the input to a store instruction was a zero vector, the backend
always selected a normal vector store regardless of the non-temporal
hint. This is fixed by this patch.
This fixes PR19370.
llvm-svn: 220054
In AVX512f we support 64x2 and 32x8 inserts via matching them to 32x4 and 64x4
respectively. These are matched by "Alt" Pat<>'s (Alt stands for alternative
VTs).
Since DQ has native support for these intructions, I peeled off the non-"Alt"
part of the baseclass into vinsert_for_size_no_alt. The DQ instructions are
derived from this multiclass. The "Alt" Pat<>'s are disabled with DQ.
Fixes <rdar://problem/18426089>
llvm-svn: 219874
The new attributes are NumElts and the CD8TupleForm. This prepares the code
to enable x8 and x2 inserts.
NFC, no change in X86.td.expanded except for the new attributes.
llvm-svn: 219871
It's the W bit that selects between 32 or 64 elt type and not the opcode. The
opcode selects between the width of the insert (128 or 256).
llvm-svn: 219870
Patch to provide shuffle decodes and asm comments for the sse pslldq/psrldq SSE2/AVX2 byte shift instructions.
Differential Revision: http://reviews.llvm.org/D5598
llvm-svn: 219738
Summary: [asan-asm-instrumentation] Fixed memory references which includes %rsp as a base or an index register.
Reviewers: eugenis
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5599
llvm-svn: 219602
This patch improves support for commutative instructions in the x86 memory folding implementation by attempting to fold a commuted version of the instruction if the original folding fails - if that folding fails as well the instruction is 're-commuted' back to its original order before returning.
This mainly helps the stack inliner better fold reloads of 3 (or more) operand instructions (VEX encoded SSE etc.) but by performing this in the lowest foldMemoryOperandImpl implementation it also replaces the X86InstrInfo::optimizeLoadInstr version and is now used by FastISel too.
Differential Revision: http://reviews.llvm.org/D5701
llvm-svn: 219584
On x86_64 this brings it from 80 bytes to 64 bytes. Also make any member
variables private and clean up uses to go through the existing accessors.
NFC.
llvm-svn: 219573
This is dangerous for numerous reasons. The primary risk here is with
floating point or double types where if the wrong header files are
included in a strange order this can implicitly convert to integers and
then call the C abs function on the integers. There is a secondary risk
that even impacts integers where if the namespace the code is written in
ever defines an abs overload for types within that namespace the global
abs will be hidden. The correct form is to call std::abs or write 'using
std::abs' for builtin types (and only the latter is correct in any
generic context).
I've also added the requisite header to be a bit more explicit here.
llvm-svn: 219484
No functional change.
This is the current AVX512_maskable multiclass hierarchy:
maskable_custom
/ \
/ \
maskable_common maskable_in_asm
/ \
/ \
maskable maskable_3src
llvm-svn: 219363
This adds the Pat<>'s for the intrinsics. These are necessary because we
don't lower these intrinsics to SDNodes but match them directly. See the
rational in the previous commit.
llvm-svn: 219362
These derive from the new asm-only masking definitions.
Unfortunately I wasn't able to find a ISel pattern that we could legally
generate for the masking variants. The problem is that since the destination
is v4* we would need VK4 register classes and v4i1 value types to express the
masking. These are however not legal types/classes in AVX512f but only in VL,
so things get complicated pretty quickly. We can revisit this question later
if we have a more pressing need to express something like this.
So the ISel patterns are empty for the masking instructions and the next patch
will add Pat<>s instead to match the intrinsics calls with instructions.
llvm-svn: 219361
No functional change.
No change in X86.td.expanded except for the appearance of the new attributes.
The new attributes will be used in the subsequent patch.
llvm-svn: 219360
Summary:
I had forgotten to check for NotSlowIncDec in the patterns that can generate
inc/dec for the above pattern (added in D4796).
This currently applies to Atom Silvermont, KNL and SKX.
Test Plan: New checks on atomic_mi.ll
Reviewers: jfb, nadav
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5677
llvm-svn: 219336
Summary:
Fix pr21099
The pseudocode of what we were doing (spread through two functions) was:
if (operand.doesNotFitIn32Bits())
Opc.initializeWithFoo();
if (operand < 0)
operand = -operand;
if (operand.doesFitIn8Bits())
Opc.initializeWithBar();
else if (operand.doesFitIn32Bits())
Opc.initializeWithBlah();
doStuff(Opc);
So for operand == INT32_MIN, Opc was never initialized because the operand changes
from fitting in 32 bits to not fitting, causing the various bugs/error messages
noted by pr21099.
This patch adds an extra test at the beginning for this case, and an
llvm_unreachable to have better error message if the operand ends up
not fitting in 32-bits at the end.
Test Plan: new test + make check
Reviewers: jfb
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5655
llvm-svn: 219257
Unfortunately, this isn't easy to fix since there's no simple way to figure out from the disassembler tables whether the W-bit is being used to select a 64-bit GPR or if its a required part of the opcode. The fix implemented here just looks for "64" in the instruction name and ignores the W-bit in 32-bit mode if its present.
Fixes PR21169.
llvm-svn: 219194
These will make it easier to test further changes to the
code generation and optimization pipelines as those are
moved to subtargets initialized with target feature and
target cpu.
llvm-svn: 219106
This trades a (register-renamer-friendly) movaps for a floating point
/ integer domain cross. That is a very bad trade, even on architectures
where domain crossing is relatively fast. On any chip where there is
even a cycle stall, this is a Very Bad Idea. It doesn't even seem likely
to cause a spill to be introduced because the reason for the copy is to
destructively shuffle in place.
Thanks to Ben Kramer for fixing a bug in this code that my new shuffle
lowering exposed and highlighting that perhaps it should just go away.
=]
llvm-svn: 219090
new vector shuffle lowering.
This is loosely based on a patch by Marius Wachtler to the PR (thanks!).
I refactored it a bi to use std::count_if and a mutable array ref but
the core idea was exactly right. I also added some direct testing of
this case.
I believe PR21137 is now the only remaining regression.
llvm-svn: 219081
shuffles using AVX and AVX2 instructions. This fixes PR21138, one of the
few remaining regressions impacting benchmarks from the new vector
shuffle lowering.
You may note that it "regresses" many of the vperm2x128 test cases --
these were actually "improved" by the naive lowering that the new
shuffle lowering previously did. This regression gave me fits. I had
this patch ready-to-go about an hour after flipping the switch but
wasn't sure how to have the best of both worlds here and thought the
correct solution might be a completely different approach to lowering
these vector shuffles.
I'm now convinced this is the correct lowering and the missed
optimizations shown in vperm2x128 are actually due to missing
target-independent DAG combines. I've even written most of the needed
DAG combine and will submit it shortly, but this part is ready and
should help some real-world benchmarks out.
llvm-svn: 219079
Update the entire regression test suite for the new shuffles. Remove
most of the old testing which was devoted to the old shuffle lowering
path and is no longer relevant really. Also remove a few other random
tests that only really exercised shuffles and only incidently or without
any interesting aspects to them.
Benchmarking that I have done shows a few small regressions with this on
LNT, zero measurable regressions on real, large applications, and for
several benchmarks where the loop vectorizer fires in the hot path it
shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy
Bridge machines. Running on AMD machines shows even more dramatic
improvements.
When using newer ISA vector extensions the gains are much more modest,
but the code is still better on the whole. There are a few regressions
being tracked (PR21137, PR21138, PR21139) but by and large this is
expected to be a win for x86 generated code performance.
It is also more correct than the code it replaces. I have fuzz tested
this extensively with ISA extensions up through AVX2 and found no
crashes or miscompiles (yet...). The old lowering had a few miscompiles
and crashers after a somewhat smaller amount of fuzz testing.
There is one significant area where the new code path lags behind and
that is in AVX-512 support. However, there was *extremely little*
support for that already and so this isn't a significant step backwards
and the new framework will probably make it easier to implement lowering
that uses the full power of AVX-512's table-based shuffle+blend (IMO).
Many thanks to Quentin, Andrea, Robert, and others for benchmarking
assistance. Thanks to Adam and others for help with AVX-512. Thanks to
Hal, Eric, and *many* others for answering my incessant questions about
how the backend actually works. =]
I will leave the old code path in the tree until the 3 PRs above are at
least resolved to folks' satisfaction. Then I will rip it (and 1000s of
lines of code) out. =] I don't expect this flag to stay around for very
long. It may not survive next week.
llvm-svn: 219046
It turns out this combine was always somewhat flawed -- there are cases
where nested VZEXT nodes *can't* be combined: if their types have
a mismatch that can be observed in the result. While none of these show
up in currently, once I switch to the new vector shuffle lowering a few
test cases actually form such nested VZEXT nodes. I've not come up with
any IR pattern that I can sensible write to exercise this, but it will
be covered by tests once I flip the switch.
llvm-svn: 219044
nodes to the DAG combining of them.
This will allow the combine to fire on both old vector shuffle lowering
and the new vector shuffle lowering and generally seems like a cleaner
design. I've trimmed down the code a bit and tried to make it and the
surrounding combine fairly clean while moving it around.
llvm-svn: 219042
the various ways in which blends can be used to do vector element
insertion for lowering with the scalar math instruction forms that
effectively re-blend with the high elements after performing the
operation.
This then allows me to bail on the element insertion lowering path when
we have SSE4.1 and are going to be doing a normal blend, which in turn
restores the last of the blends lost from the new vector shuffle
lowering when I got it to prioritize insertion in other cases (for
example when we don't *have* a blend instruction).
Without the patterns, using blends here would have regressed
sse-scalar-fp-arith.ll *completely* with the new vector shuffle
lowering. For completeness, I've added RUN-lines with the new lowering
here. This is somewhat superfluous as I'm about to flip the default, but
hey, it shows that this actually significantly changed behavior.
The patterns I've added are just ridiculously repetative. Suggestions on
making them better very much welcome. In particular, handling the
commuted form of the v2f64 patterns is somewhat obnoxious.
llvm-svn: 219033
perform a load to use blendps rather than movss when it is available.
For non-loads, blendps is *much* faster. It can execute on two ports in
Sandy Bridge and Ivy Bridge, and *three* ports on Haswell. This fixes
one of the "regressions" from aggressively taking the "insertion" path
in the new vector shuffle lowering.
This does highlight one problem with blendps -- it isn't commuted as
heavily as it should be. That's future work though.
llvm-svn: 219022
In the X86 backend, matching an address is initiated by the 'addr' complex
pattern and its friends. During this process we may reassociate and-of-shift
into shift-of-and (FoldMaskedShiftToScaledMask) to allow folding of the
shift into the scale of the address.
However as demonstrated by the testcase, this can trigger CSE of not only the
shift and the AND which the code is prepared for but also the underlying load
node. In the testcase this node is sitting in the RecordedNode and MatchScope
data structures of the matcher and becomes a deleted node upon CSE. Returning
from the complex pattern function, we try to access it again hitting an assert
because the node is no longer a load even though this was checked before.
Now obviously changing the DAG this late is bending the rules but I think it
makes sense somewhat. Outside of addresses we prefer and-of-shift because it
may lead to smaller immediates (FoldMaskAndShiftToScale is an even better
example because it create a non-canonical node). We currently don't recognize
addresses during DAGCombiner where arguably this canonicalization should be
performed. On the other hand, having this in the matcher allows us to cover
all the cases where an address can be used in an instruction.
I've also talked a little bit to Dan Gohman on llvm-dev who added the RAUW for
the new shift node in FoldMaskedShiftToScaledMask. This RAUW is responsible
for initiating the recursive CSE on users
(http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-September/076903.html) but it
is not strictly necessary since the shift is hooked into the visited user. Of
course it's safer to keep the DAG consistent at all times (e.g. for accurate
number of uses, etc.).
So rather than changing the fundamentals, I've decided to continue along the
previous patches and detect the CSE. This patch installs a very targeted
DAGUpdateListener for the duration of a complex-pattern match and updates the
matching state accordingly. (Previous patches used HandleSDNode to detect the
CSE but that's not practical here). The listener is only installed on X86.
I tested that there is no measurable overhead due to this while running
through the spec2k BC files with llc. The only thing we pay for is the
creation of the listener. The callback never ever triggers in spec2k since
this is a corner case.
Fixes rdar://problem/18206171
llvm-svn: 219009
and MOVSD nodes for single element vector inserts.
This is particularly important because a number of patterns in the
backend detect these patterns and leverage them to simplify things. It
also fixes quite a few of the insertion bad code examples. However, it
regresses a specific area: when available, blendps and blendpd are
*dramatically* faster than movss and movsd respectively. But it doesn't
really work to form the blend logic first because the blends *aren't* as
crazy efficient when the data is coming from memory anyways, and thus
will have a movss or movsd regardless. Also, doing that would block
a bunch of the patterns that this is designed to hit.
So my plan is to go into the patterns for lowering MOVSS and MOVSD and
lower them via blends when available. However that's a pretty invasive
restructuring so it will need to be a follow-up patch.
I have already gone into the patterns to lower MOVSS and MOVSD from
memory using MOVLPD, etc. Without that, several of the test cases
I already have regress.
llvm-svn: 218985
lowering to handle the potential mirroring of 2-element vectors (because
we can't reliably sort them one way) in the caller rather than in the
insertion logic.
This will simplify things considerably as more ways to fail to match the
insertion are added because now we have a nice try and retry point.
llvm-svn: 218980
lowering to match VZEXT_MOVL patterns.
I hadn't realized that these had sufficient pattern smarts in the
backend to lower zext-ing from the low element of a vector without it
being a scalar_to_vector node. They do, and this is how to match a bunch
of patterns for movq, movss, etc.
There is a weird propensity to end up using pshufd to place the element
afterward even though it means domain crossing (or rather, to use
xorps+movss to zext the element rather than movq) but that's an
orthogonal problem with VZEXT_MOVL that someone should probably look at.
llvm-svn: 218977
element types to form illegal vector types.
I've added a special SSE1 test case here that makes sure we don't break
this going forward.
llvm-svn: 218974
No functional change intended.
Very similar to the change I made for subvector extract in r218480.
test/CodeGen/X86/avx512-insert-extract.ll covers this.
llvm-svn: 218928
Do not eliminate the frame pointer if there is a stackmap or patchpoint in the
function. All stackmap references should be FP relative.
This fixes PR21107.
llvm-svn: 218920
elements as well as integer elements in order to form simpler shuffle
patterns.
This is the primary reason why we were failing to match some of the
2-and-2 floating point shuffles such as PR21140. Even after fixing this
we need to support some extra patterns in the backend in order to match
the resulting X86ISD::UNPCKL nodes into the correct instructions. This
commit should fix PR21140 and includes more comprehensive testing of
insertion patterns in v4 shuffles.
Not all of the added tests are beautiful. For example, we don't have
clever instructions to insert-via-load in the integer domain. There are
also some places where we aren't sufficiently cunning with our use of
movq and movd, but that's future work.
llvm-svn: 218911
matching and lowering 64-bit insertions.
The first problem was that we weren't looking through bitcasts to
discover that we *could* lower as insertions. Once fixed, we in turn
weren't looking through bitcasts to discover that we could fold a load
into the lowering. Once fixed, we weren't forming a SCALAR_TO_VECTOR
node around the inserted element and instead were passing a scalar to
a DAG node that expected a vector. It turns out there are some patterns
that will "lower" this into the correct asm, but the rest of the X86
backend is very unhappy with such antics.
This should fix a few more edge case regressions I've spotted going
through the regression test suite to enable the new vector shuffle
lowering.
llvm-svn: 218839
Negative FABS of either a scalar or vector should be handled the same way
on x86 with SSE/AVX: a single OR instruction of the FP operand with a
constant to light up the sign bit(s).
http://llvm.org/bugs/show_bug.cgi?id=20578
Differential Revision: http://reviews.llvm.org/D5201
llvm-svn: 218822
argument of the llvm.dbg.declare/llvm.dbg.value intrinsics.
Previously, DIVariable was a variable-length field that has an optional
reference to a Metadata array consisting of a variable number of
complex address expressions. In the case of OpPiece expressions this is
wasting a lot of storage in IR, because when an aggregate type is, e.g.,
SROA'd into all of its n individual members, the IR will contain n copies
of the DIVariable, all alike, only differing in the complex address
reference at the end.
By making the complex address into an extra argument of the
dbg.value/dbg.declare intrinsics, all of the pieces can reference the
same variable and the complex address expressions can be uniqued across
the CU, too.
Down the road, this will allow us to move other flags, such as
"indirection" out of the DIVariable, too.
The new intrinsics look like this:
declare void @llvm.dbg.declare(metadata %storage, metadata %var, metadata %expr)
declare void @llvm.dbg.value(metadata %storage, i64 %offset, metadata %var, metadata %expr)
This patch adds a new LLVM-local tag to DIExpressions, so we can detect
and pretty-print DIExpression metadata nodes.
What this patch doesn't do:
This patch does not touch the "Indirect" field in DIVariable; but moving
that into the expression would be a natural next step.
http://reviews.llvm.org/D4919
rdar://problem/17994491
Thanks to dblaikie and dexonsmith for reviewing this patch!
Note: I accidentally committed a bogus older version of this patch previously.
llvm-svn: 218787
argument of the llvm.dbg.declare/llvm.dbg.value intrinsics.
Previously, DIVariable was a variable-length field that has an optional
reference to a Metadata array consisting of a variable number of
complex address expressions. In the case of OpPiece expressions this is
wasting a lot of storage in IR, because when an aggregate type is, e.g.,
SROA'd into all of its n individual members, the IR will contain n copies
of the DIVariable, all alike, only differing in the complex address
reference at the end.
By making the complex address into an extra argument of the
dbg.value/dbg.declare intrinsics, all of the pieces can reference the
same variable and the complex address expressions can be uniqued across
the CU, too.
Down the road, this will allow us to move other flags, such as
"indirection" out of the DIVariable, too.
The new intrinsics look like this:
declare void @llvm.dbg.declare(metadata %storage, metadata %var, metadata %expr)
declare void @llvm.dbg.value(metadata %storage, i64 %offset, metadata %var, metadata %expr)
This patch adds a new LLVM-local tag to DIExpressions, so we can detect
and pretty-print DIExpression metadata nodes.
What this patch doesn't do:
This patch does not touch the "Indirect" field in DIVariable; but moving
that into the expression would be a natural next step.
http://reviews.llvm.org/D4919
rdar://problem/17994491
Thanks to dblaikie and dexonsmith for reviewing this patch!
llvm-svn: 218778
that keep cropping up in the regression test suite.
This also addresses one of the issues raised on the mailing list with
failing to form 'movsd' in as many cases as we realistically should.
There will be corresponding patches forthcoming for v4f32 at least. This
was a lot of fuss for a relatively small gain, but all the fuss was on
my end trying different ways of holding the pieces of the x86 fragment
patterns *just right*. Now that it works, the code is reasonably simple.
In the new test cases I'm adding here, v2i64 sticks out as just plain
horrible. I've not come up with any great ideas here other than that it
would be nice to recognize when we're *going* to take a domain crossing
hit and cross earlier to get the decent instructions. At least with AVX
it is slightly less silly....
llvm-svn: 218756
Nothing was relying on this and there are potentially some edge cases
that it would not be correct under. Removing it seems better than trying
to "fix" it as nothing was relying on it.
llvm-svn: 218755
in exposing the scalar value to the broadcast DAG fragment so that we
can catch even reloads and fold them into the broadcast.
This is somewhat magical I'm afraid but seems to work. It is also what
the old lowering did, and I've switched an old test to run both
lowerings demonstrating that we get the same result.
Unlike the old code, I'm not lowering f32 or f64 scalars through this
path when we only have AVX1. The target patterns include pretty heinous
code to re-cast those as shuffles when the scalar happens to not be
spilled because AVX1 provides no broadcast mechanism from registers
what-so-ever. This is terribly brittle. I'd much rather go through our
generic lowering code to get this. If needed, we can add a peephole to
get even more opportunities to broadcast-from-spill-slots that are
exposed post-RA, but my suspicion is this just doesn't matter that much.
llvm-svn: 218734
the same speed as pshufd but we can fold loads into the pmovzx
instructions.
This fixes some regressions that came up in the regression test suite
for the new vector shuffle lowering.
llvm-svn: 218733
VPBROADCAST.
This has the somewhat expected pervasive impact. I don't know why
I forgot about this. Everything seems good with lots of significant
improvements in the tests.
llvm-svn: 218724
Fixed lowering of this intrinsics in case when mask is v2i1 and v4i1.
Now cmp intrinsics lower in the following way:
(i8 (int_x86_avx512_mask_pcmpeq_q_128
(v2i64 %a), (v2i64 %b), (i8 %mask))) ->
(i8 (bitcast
(v8i1 (insert_subvector undef,
(v2i1 (and (PCMPEQM %a, %b),
(extract_subvector
(v8i1 (bitcast %mask)), 0))), 0))))
llvm-svn: 218669
a flawed direction and causing miscompiles. Read on for details.
Fundamentally, the premise of this patch series was to map
VECTOR_SHUFFLE DAG nodes into VSELECT DAG nodes for all blends because
we are going to *have* to lower to VSELECT nodes for some blends to
trigger the instruction selection patterns of variable blend
instructions. This doesn't actually work out so well.
In order to match performance with the existing VECTOR_SHUFFLE
lowering code, we would need to re-slice the blend in order to fit it
into either the integer or floating point blends available on the ISA.
When coming from VECTOR_SHUFFLE (or other vNi1 style VSELECT sources)
this works well because the X86 backend ensures that these types of
operands to VSELECT get sign extended into '-1' and '0' for true and
false, allowing us to re-slice the bits in whatever granularity without
changing semantics.
However, if the VSELECT condition comes from some other source, for
example code lowering vector comparisons, it will likely only have the
required bit set -- the high bit. We can't blindly slice up this style
of VSELECT. Reid found some code using Halide that triggers this and I'm
hopeful to eventually get a test case, but I don't need it to understand
why this is A Bad Idea.
There is another aspect that makes this approach flawed. When in
VECTOR_SHUFFLE form, we have very distilled information that represents
the *constant* blend mask. Converting back to a VSELECT form actually
can lose this information, and so I think now that it is better to treat
this as VECTOR_SHUFFLE until the very last moment and only use VSELECT
nodes for instruction selection purposes.
My plan is to:
1) Clean up and formalize the target pre-legalization DAG combine that
converts a VSELECT with a constant condition operand into
a VECTOR_SHUFFLE.
2) Remove any fancy lowering from VSELECT during *legalization* relying
entirely on the DAG combine to catch cases where we can match to an
immediate-controlled blend instruction.
One additional step that I'm not planning on but would be interested in
others' opinions on: we could add an X86ISD::VSELECT or X86ISD::BLENDV
which encodes a fully legalized VSELECT node. Then it would be easy to
write isel patterns only in terms of this to ensure VECTOR_SHUFFLE
legalization only ever forms the fully legalized construct and we can't
cycle between it and VSELECT combining.
llvm-svn: 218658
No functionality change.
Makes the code more compact (see the FMA part).
This needs a new type attribute MemOpFrag in X86VectorVTInfo. For now I only
defined this in the simple cases. See the commment before the attribute.
Diff of X86.td.expanded before and after is empty except for the appearance of
the new attribute.
llvm-svn: 218637
nodes, and rely exclusively on its logic. This removes a ton of
duplication from the blend lowering and centralizes it in one place.
One downside is that it requires a bunch of hacks to make this work with
the current legalization framework. We have to manually speculate one
aspect of legalizing VSELECT nodes to get everything to work nicely
because the existing legalization framework isn't *actually* bottom-up.
The other grossness is that we somewhat duplicate the analysis of
constant blends. I'm on the fence here. If reviewers thing this would
look better with VSELECT when it has constant operands dumping over tho
VECTOR_SHUFFLE, we could go that way. But it would be a substantial
change because currently all of the actual blend instructions are
matched via patterns in the TD files based around VSELECT nodes (despite
them not being perfect fits for that). Suggestions welcome, but at least
this removes the rampant duplication in the backend.
llvm-svn: 218600
X86 target-specific DAG combining that tried to convert VSELECT nodes
into VECTOR_SHUFFLE nodes that it "knew" would lower into
immediate-controlled blend nodes.
Turns out, we have perfectly good lowering of all these VSELECT nodes,
and indeed that lowering already knows how to handle lowering through
BLENDI to immediate-controlled blend nodes. The code just wasn't getting
used much because this thing forced the world to go through the vector
shuffle lowering. Yuck.
This also exposes that I was too aggressive in avoiding domain crossing
in v218588 with that lowering -- when the other option is to expand into
two 128-bit vectors, it is worth domain crossing. Restore that behavior
now that we have nice tests covering it.
The test updates here fall into two camps. One is where previously we
ended up with an unsigned encoding of the blend operand and now we get
a signed encoding. In most of those places there were elaborate comments
explaining exactly what these operands really mean. Rather than that,
just switch these tests to use the nicely decoded comments that make it
obvious that the final shuffle matches.
The other updates are just removing pointless domain crossing by
blending integers with PBLENDW rather than BLENDPS.
llvm-svn: 218589
crossing and generally work more like the blend emission code in the new
vector shuffle lowering.
My goal is to have the new vector shuffle lowering just produce VSELECT
nodes that are either matched here to BLENDI or are legal and matched in
the .td files to specific blend instructions. That seems much cleaner as
there are other ways to produce a VSELECT anyways. =]
No *observable* functionality changed yet, mostly because this code
appears to be near-dead. The behavior of this lowering routine did
change though. This code being mostly dead and untestable will change
with my next commit which will also point some new tests at it.
llvm-svn: 218588
AVX-512.
There is no interesting logic yet. Everything ends up eventually
delegating to the generic code to split the vector and shuffle the
halves. Interestingly, that logic does a significantly better job of
lowering all of these types than the generic vector expansion code does.
Mostly, it lets most of the cases fall back to nice AVX2 code rather
than all the way back to SSE code paths.
Step 2 of basic AVX-512 support in the new vector shuffle lowering. Next
up will be to incrementally add direct support for the basic instruction
set to each type (adding tests first).
llvm-svn: 218585
assertion, making the name generic, and improving the documentation.
Step 1 in adding very primitive support for AVX-512. No functionality
changed yet.
llvm-svn: 218584
vectors.
Someone will need to build the AVX512 lowering, which should follow
AVX1 and AVX2 *very* closely for AVX512F and AVX512BW resp. I've added
a dummy test which is a port of the v8f32 and v8i32 tests from AVX and
AVX2 to v8f64 and v8i64 tests for AVX512F and AVX512BW. Hopefully this
is enough information for someone to implement proper lowering here. If
not, I'll be happy to help, but right now the AVX-512 support isn't
a priority for me.
llvm-svn: 218583
lowerings.
This was hopelessly broken. First, the x86 backend wants '-1' to be the
element value representing true in a boolean vector, and second the
operand order for VSELECT is backwards from the actual x86 instructions.
To make matters worse, the backend is just using '-1' as the true value
to get the high bit to be set. It doesn't actually symbolically map the
'-1' to anything. But on x86 this isn't quite how it works: there *only*
the high bit is relevant. As a consequence weird non-'-1' values like
0x80 actually "work" once you flip the operands to be backwards.
Anyways, thanks to Hal for helping me sort out what these *should* be.
llvm-svn: 218582
new vector shuffle target DAG combines -- it helps to actually test for
the value you want rather than just using an integer in a boolean
context.
Have I mentioned that I loathe implicit conversions recently? :: sigh ::
llvm-svn: 218576
of widening masks.
We can't widen a zeroing mask unless both elements that would be merged
are either zeroed or undef. This is the only way to widen a mask if it
has a zeroed element.
Also clean up the code here by ordering the checks in a more logical way
and by using the symoblic values for undef and zero. I'm actually torn
on using the symbolic values because the existing code is littered with
the assumption that -1 is undef, and moreover that entries '< 0' are the
special entries. While that works with the values given to these
constants, using the symbolic constants actually makes it a bit more
opaque why this is the case.
llvm-svn: 218575
I spotted this by inspection when debugging something else, so I have no
test case what-so-ever, and am not even sure it is possible to
realistically trigger the bug. But this is what was intended here.
llvm-svn: 218565
and in the target shuffle combining when trying to widen vector
elements.
Previously only one of these was correct, and we didn't correctly
propagate zeroing target shuffle masks (which have a different sentinel
value from undef in non- target shuffle masks now). This isn't just
a missed optimization, this caused us to drop zeroing shuffles on the
floor and miscompile code. The added test case is one example of that.
There are other fixes to the test suite as a consequence of this as well
as restoring the undef elements in some of the masks that were lost when
I brought sanity to the actual *value* of the undef and zero sentinels.
I've also just cleaned up some of the PSHUFD and PSHUFLW and PSHUFHW
combining code, but that code really needs to go. It was a nice initial
attempt, but it isn't very principled and the recursive shuffle combiner
is much more powerful.
llvm-svn: 218562
to significantly more sane sentinels. Notably, everywhere else in the
backend's representation of shuffles uses '-1' to represent undef. The
target shuffle masks really shouldn't diverge from that, especially as
in a few places they are manipulated by shared code.
This causes us to lose some undef lanes in various test masks. I want to
get these back, but technically it isn't invalid and there are a *lot*
of bugs here so I want to try to establish a saner baseline for fixing
some of the bugs by aligning the specific senitnel values used.
llvm-svn: 218561
that managed to elude all of my fuzz testing historically. =/
Something changed to allow this code path to actually be exercised and
it was doing bad things. It is especially heavily exercised by the
patterns that emerge when doing AVX shuffles that end up lowered through
the 128-bit code path.
llvm-svn: 218540
layer of tie-breaking sorting, it really helps to check that you're in
a tie first. =] Otherwise the whole thing cycles infinitely. Test case
added, another one found through fuzz testing.
llvm-svn: 218523
AVX support.
New test cases included. Note that none of the existing test cases
covered these buggy code paths. =/ Also, it is clear from this that
SHUFPS and SHUFPD are the most bug prone shuffle instructions in x86. =[
These were all detected by fuzz-testing. (I <3 fuzz testing.)
llvm-svn: 218522
The SSE rsqrt instruction (a fast reciprocal square root estimate) was
grouped in the same scheduling IIC_SSE_SQRT* class as the accurate (but very
slow) SSE sqrt instruction. For code which uses rsqrt (possibly with
newton-raphson iterations) this poor scheduling was affecting performances.
This patch splits off the rsqrt instruction from the sqrt instruction scheduling
classes and creates new IIC_SSE_RSQER* classes with latency values based on
Agner's table.
Differential Revision: http://reviews.llvm.org/D5370
Patch by Simon Pilgrim.
llvm-svn: 218517
No functional change.
I initially thought that pulling the Pat<> into the instruction pattern was
not possible because it was doing a transform on the index in order to convert
it from a per-element (extract_subvector) index into a per-chunk (vextract*x4)
index.
Turns out this also works inside the pattern because the vextract_extract
PatFrag has an OperandTransform EXTRACT_get_vextract{128,256}_imm, so the
index in $idx goes through the same conversion.
The existing test CodeGen/X86/avx512-insert-extract.ll extended in the
previous commit provides coverage for this change.
llvm-svn: 218480
No functional change.
These are now implemented as two levels of multiclasses heavily relying on the
new X86VectorVTInfo class. The multiclass at the first level that is called
with float or int provides the 128 or 256 bit subvector extracts. The second
level provides the register and memory variants and some more Pat<>s.
I've compared the td.expanded files before and after. One change is that
ExeDomain for 64x4 is SSEPackedDouble now. I think this is correct, i.e. a
bugfix.
(BTW, this is the change that was blocked on the recent tablegen fix. The
class-instance values X86VectorVTInfo inside vextract_for_type weren't
properly evaluated.)
Part of <rdar://problem/17688758>
llvm-svn: 218478
Summary:
I originally tried doing this specifically for X86 in the backend in D5091,
but it was rather brittle and generally running too late to be general.
Furthermore, other targets may want to implement similar optimizations.
So I reimplemented it at the IR-level, fitting it into AtomicExpandPass
as it interacts with that pass (which could not be cleanly done before
at the backend level).
This optimization relies on a new target hook, which is only used by X86
for now, as the correctness of the optimization on other targets remains
an open question. If it is found correct on other targets, it should be
trivial to enable for them.
Details of the optimization are discussed in D5091.
Test Plan: make check-all + a new test
Reviewers: jfb
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5422
llvm-svn: 218455
v4f64 and v8f32 shuffles when they are lane-crossing. We have fully
general lane-crossing permutation functions in AVX2 that make this easy.
Part of this also changes exactly when and how these vectors are split
up when we don't have AVX2. This isn't always a win but it usually is
a win, so on the balance I think its better. The primary regressions are
all things that just need to be fixed anyways such as modeling when
a blend can be completely accomplished via VINSERTF128, etc.
Also, this highlights one of the few remaining big features: we do
a really poor job of inserting elements into AVX registers efficiently.
This completes almost all of the big tricks I have in mind for AVX2. The
only things left that I plan to add:
1) element insertion smarts
2) palignr and other fairly specialized lowerings when they happen to
apply
llvm-svn: 218449
256-bit vectors with lane-crossing.
Rather than immediately decomposing to 128-bit vectors, try flipping the
256-bit vector lanes, shuffling them and blending them together. This
reduces our worst case shuffle by a pretty significant margin across the
board.
llvm-svn: 218446
lowering where it only used the mask of the low 128-bit lane rather than
the entire mask.
This allows the new lowering to correctly match the unpack patterns for
v8i32 vectors.
For reference, the reason that we check for the the entire mask rather
than checking the repeated mask is because the repeated masks don't
abide by all of the invariants of normal masks. As a consequence, it is
safer to use the full mask with functions like the generic equivalence
test.
llvm-svn: 218442
reduce the amount of checking we do here.
The first realization is that only non-crossing cases between 128-bit
lanes are handled by almost the entire function. It makes more sense to
handle the crossing cases first.
THe second is that until we actually are going to generate fancy shared
lowering strategies that use the repeated semantics of the v8i16
lowering, we should waste time checking for repeated masks. It is
simplest to directly test for the entire unpck masks anyways, so we
gained nothing from this.
This also matches the structure of v32i8 more closely.
No functionality changed here.
llvm-svn: 218441
lowering.
This completes the basic AVX2 feature support, but there are still some
improvements I'd like to do to really get the last mile of performance
here.
llvm-svn: 218440
for this now.
Should prevent folks from running afoul of this and not knowing why
their code won't instruction select the way I just did...
llvm-svn: 218436
missing test cases for it.
Unsurprisingly, without test cases, there were bugs here. Surprisingly,
this bug wasn't caught at compile time. Yep, there is an X86ISD::BLENDV.
It isn't wired to anything. Oops. I'll fix than next.
llvm-svn: 218434
lowering.
This also implements the fancy blend lowering for v16i16 using AVX2 and
teaches the X86 backend to print shuffle masks for 256-bit PSHUFB
and PBLENDW instructions. It also makes the mask decoding correct for
PBLENDW instructions. The yaks, they are legion.
Tests are updated accordingly. There are some missing tests for the
VBLENDVB lowering, but I'll add those in a follow-up as this commit has
accumulated enough cruft already.
llvm-svn: 218430
into unblended shuffles and a blend.
This is the consistent fallback for the lowering paths that have fast
blend operations available, and its getting quite repetitive.
No functionality changed.
llvm-svn: 218399
pool data being loaded into a vector register.
The comments take the form of:
# ymm0 = [a,b,c,d,...]
# xmm1 = <x,y,z...>
The []s are used for generic sequential data and the <>s are used for
specifically ConstantVector loads. Undef elements are printed as the
letter 'u', integers in decimal, and floating point values as floating
point values. Suggestions on improving the formatting or other aspects
of the display are very welcome.
My primary use case for this is to be able to FileCheck test masks
passed to vector shuffle instructions in-register. It isn't fantastic
for that (no decoding special zeroing semantics or other tricks), but it
at least puts the mask onto an instruction line that could reasonably be
checked. I've updated many of the new vector shuffle lowering tests to
leverage this in their test cases so that we're actually checking the
shuffle masks remain as expected.
Before implementing this, I tried a *bunch* of different approaches.
I looked into teaching the MCInstLower code to scan up the basic block
and find a definition of a register used in a shuffle instruction and
then decode that, but this seems incredibly brittle and complex.
I talked to Hal a lot about the "right" way to do this: attach the raw
shuffle mask to the instruction itself in some form of unencoded
operands, and then use that to emit the comments. I still think that's
the optimal solution here, but it proved to be beyond what I'm up for
here. In particular, it seems likely best done by completing the
plumbing of metadata through these layers and attaching the shuffle mask
in metadata which could have fully automatic dropping when encoding an
actual instruction.
llvm-svn: 218377
attempt didn't work out so well. It looks like it will be much better
for introducing extra logic to find a shuffle mask if the finding logic
is totally separate. This also makes it easy to sink the opcode logic
completely out of the routine so we don't re-dispatch across it.
Still no functionality changed.
llvm-svn: 218363
asm. This can be somewhat expensive and there is no reason to do it
outside of tests or debugging sessions. I'm also likely to make it
significantly more expensive to support more styles of shuffles.
llvm-svn: 218362
from the MachineInstr into the caller which is already doing a switch
over the instruction.
This will make it more clear how to compute different operands to feed
the comment selection for example.
Also, in a drive-by-fix, don't append an empty comment string (which is
a no-op ultimately).
No functionality changed.
llvm-svn: 218361
vector shuffles.
This is just the beginning by hoisting it into its own function and
making use of early exit to dramatically simplify the flow of the
function. I'm going to be incrementally refactoring this until it is
a bit less magical how this applies to other instructions, and I can
teach it how to dig a shuffle mask out of a register. Then I plan to
hook it up to VPERMD so we get our mask comments for it.
No functionality changed yet.
llvm-svn: 218357
the native AVX2 instructions.
Note that the test case is really frustrating here because VPERMD
requires the mask to be in the register input and we don't produce
a comment looking through that to the constant pool. I'm going to
attempt to improve this in a subsequent commit, but not sure if I will
succeed.
llvm-svn: 218347
detection. It was incorrectly handling undef lanes by actually treating
an undef lane in the first 128-bit lane as a *numeric* shuffle value.
Fortunately, this almost always DTRT and disabled detecting repeated
patterns. But not always. =/ This patch introduces a much more
principled approach and fixes the miscompiles I spotted by inspection
previously.
llvm-svn: 218346
shuffles using the AVX2 instructions. This is the first step of cutting
in real AVX2 support.
Note that I have spotted at least one bug in the test cases already, but
I suspect it was already present and just is getting surfaced. Will
investigate next.
llvm-svn: 218338
add VPBLENDD to the InstPrinter's comment generation so we get nice
comments everywhere.
Now that we have the nice comments, I can see the bug introduced by
a silly typo in the commit that enabled VPBLENDD, and have fixed it. Yay
tests that are easy to inspect.
llvm-svn: 218335
Summary:
AtomicExpand already had logic for expanding wide loads and stores on LL/SC
architectures, and for expanding wide stores on CmpXchg architectures, but
not for wide loads on CmpXchg architectures. This patch fills this hole,
and makes use of this new feature in the X86 backend.
Only one functionnal change: we now lose the SynchScope attribute.
It is regrettable, but I have another patch that I will submit soon that will
solve this for all of AtomicExpand (it seemed better to split it apart as it
is a different concern).
Test Plan: make check-all (lots of tests for this functionality already exist)
Reviewers: jfb
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5404
llvm-svn: 218332
VPBLENDD where appropriate even on 128-bit vectors.
According to Agner's tables, this instruction is significantly higher
throughput (can execute on any port) on Haswell chips so we should
aggressively try to form it when available.
Sadly, this loses our delightful shuffle comments. I'll add those back
for VPBLENDD next.
llvm-svn: 218322
undef in the shuffle mask. This shows up when we're printing comments
during lowering and we still have an IR-level constant hanging around
that models undef.
A nice consequence of this is *much* prettier test cases where the undef
lanes actually show up as undef rather than as a particular set of
values. This also allows us to print shuffle comments in cases that use
undef such as the recently added variable VPERMILPS lowering. Now those
test cases have nice shuffle comments attached with their details.
The shuffle lowering for PSHUFB has been augmented to use undef, and the
shuffle combining has been augmented to comprehend it.
llvm-svn: 218301
trick that I missed.
VPERMILPS has a non-immediate memory operand mode that allows it to do
asymetric shuffles in the two 128-bit lanes. Use this rather than two
shuffles and a blend.
However, it turns out the variable shuffle path to VPERMILPS (and
VPERMILPD, although that one offers no functional differenc from the
immediate operand other than variability) wasn't even plumbed through
codegen. Do such plumbing so that we can reasonably emit
a variable-masked VPERMILP instruction. Also plumb basic comment parsing
and printing through so that the tests are reasonable.
There are still a few tests which don't show the shuffle pattern. These
are tests with undef lanes. I'll teach the shuffle decoding and printing
to handle undef mask entries in a follow-up. I've looked at the masks
and they seem reasonable.
llvm-svn: 218300
td pattern). Currently we only model the immediate operand variation of
VPERMILPS and VPERMILPD, we should make that clear in the pseudos used.
Will be adding support for the variable mask variant in my next commit.
llvm-svn: 218282
Summary:
This fixes a couple of issues. One is ensuring that AOK_Label rewrite
rules have a lower priority than AOK_Skip rules, as AOK_Skip needs to
be able to skip the brackets properly. The other part of the fix ensures
that we don't overwrite Identifier when looking up the identifier, and
that we use the locally available information to generate the AOK_Label
rewrite in ParseIntelIdentifier. Doing that in CreateMemForInlineAsm
would be problematic since the Start location there may point to the
beginning of a bracket expression, and not necessarily the beginning of
an identifier.
This also means that we don't need to carry around the InternlName field,
which helps simplify the code.
Test Plan: This will be tested on the clang side.
Reviewers: rnk
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5445
llvm-svn: 218270
We generate broadcast instructions on CPUs with AVX2 to load some constant splat vectors.
This patch should preserve all existing behavior with regular optimization levels,
but also use splats whenever possible when optimizing for *size* on any CPU with AVX or AVX2.
The tradeoff is up to 5 extra instruction bytes for the broadcast instruction to save
at least 8 bytes (up to 31 bytes) of constant pool data.
Differential Revision: http://reviews.llvm.org/D5347
llvm-svn: 218263
Summary:
Update segmented-stacks*.ll tests with x32 target case and make
corresponding changes to make them pass.
Test Plan: tests updated with x32 target
Reviewers: nadav, rafael, dschuff
Subscribers: llvm-commits, zinovy.nis
Differential Revision: http://reviews.llvm.org/D5245
llvm-svn: 218247
The PSHUFB mask decode routine used to assert if the mask index was out of
range (<0 or greater than the size of the vector). The problem is, we can
legitimately have a PSHUFB with a large index using intrinsics. The
instruction only uses the least significant 4 bits. This change removes the
assert and masks the index to match the instruction behaviour.
llvm-svn: 218242
The implementation of the callback in clang's Sema will return an
internal name for labels.
Test Plan: Will be tested in clang.
Reviewers: rnk
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D4587
llvm-svn: 218229
a more sane approach to AVX2 support.
Fundamentally, there is no useful way to lower integer vectors in AVX.
None. We always end up with a VINSERTF128 in the end, so we might as
well eagerly switch to the floating point domain and do everything
there. This cleans up lots of weird and unlikely to be correct
differences between integer and floating point shuffles when we only
have AVX1.
The other nice consequence is that by doing things this way we will make
it much easier to write the integer lowering routines as we won't need
to duplicate the logic to check for AVX vs. AVX2 in each one -- if we
actually try to lower a 256-bit vector as an integer vector, we have
AVX2 and can rely on it. I think this will make the code much simpler
and more comprehensible.
Currently, I've disabled *all* support for AVX2 so that we always fall
back to AVX. This keeps everything working rather than asserting. That
will go away with the subsequent series of patches that provide
a baseline AVX2 implementation.
Please note, I'm going to implement AVX2 *without access to hardware*.
That means I cannot correctness test this path. I will be relying on
those with access to AVX2 hardware to do correctness testing and fix
bugs here, but as a courtesy I'm trying to sketch out the framework for
the new-style vector shuffle lowering in the context of the AVX2 ISA.
llvm-svn: 218228
input v8f32 shuffles which are not 128-bit lane crossing but have
different shuffle patterns in the low and high lanes. This removes most
of the extract/insert traffic that was unnecessary and is particularly
good at lowering cases where only one of the two lanes is shuffled at
all.
I've also added a collection of test cases with undef lanes because this
lowering is somewhat more sensitive to undef lanes than others.
llvm-svn: 218226
lowering when it can use a symmetric SHUFPS across both 128-bit lanes.
This required making the SHUFPS lowering tolerant of other vector types,
and adjusting our canonicalization to canonicalize harder.
This is the last of the clever uses of symmetry I've thought of for
v8f32. The rest of the tricks I'm aware of here are to work around
assymetry in the mask.
llvm-svn: 218216
a generic vector shuffle mask into a helper that isn't specific to the
other things that influence which choice is made or the specific types
used with the instruction.
No functionality changed.
llvm-svn: 218215
of a single element into a zero vector for v4f64 and v4i64 in AVX.
Ironically, there is less to see here because xor+blend is so crazy fast
that we can't really beat that to zero the high 128-bit lane.
llvm-svn: 218214
UNPCKHPS with AVX vectors by recognizing those patterns when they are
repeated for both 128-bit lanes.
With this, we now generate the exact same (really nice) code for
Quentin's avx_test_case.ll which was the most significant regression
reported for the new shuffle lowering. In fact, I'm out of specific test
cases for AVX lowering, the rest were AVX2 I think. However, there are
a bunch of pretty obvious remaining things to improve with AVX...
llvm-svn: 218213
important bits of cleverness: to detect and lower repeated shuffle
patterns between the two 128-bit lanes with a single instruction.
This patch just teaches it how to lower single-input shuffles that fit
this model using VPERMILPS. =] There is more that needs to happen here.
llvm-svn: 218211
v8f32 shuffles in the new vector shuffle lowering code.
This is very cheap to do and makes it much more clear that anything more
expensive but overlapping with this lowering should be selected
afterward (for example using AVX2's VPERMPS). However, no functionality
changed here as without this code we would fall through to create no-op
shuffles of each input and a blend. =]
llvm-svn: 218209
VBLENDPD over using VSHUFPD. While the 256-bit variant of VBLENDPD slows
down to the same speed as VSHUFPD on Sandy Bridge CPUs, it has twice the
reciprocal throughput on Ivy Bridge CPUs much like it does everywhere
for 128-bits. There isn't a downside, so just eagerly use this
instruction when it suffices.
llvm-svn: 218208
awkward conditions. The readability improvement of this will be even
more important as I generalize it to handle more types.
No functionality changed.
llvm-svn: 218205
128-bit lane crossings, not 'half' crossings. This came up in code
review ages ago, but I hadn't really addresesd it. Also added some
documentation for the helper.
No functionality changed.
llvm-svn: 218203
actual support for complex AVX shuffling tricks. We can do independent
blends of the low and high 128-bit lanes of an avx vector, so shuffle
the inputs into place and then do the blend at 256 bits. This will in
many cases remove one blend instruction.
The next step is to permute the low and high halves in-place rather than
extracting them and re-inserting them.
llvm-svn: 218202
single-input shuffles with doubles. This allows them to fold memory
operands into the shuffle, etc. This is just the analog to the v4f32
case in my prior commit.
llvm-svn: 218193
instruction for single-vector floating point shuffles. This in turn
allows the shuffles to fold a load into the instruction which is one of
the common regressions hit with the new shuffle lowering.
llvm-svn: 218190
tricky case of single-element insertion into the zero lane of a zero
vector.
We can't just use the same pattern here as we do in every other vector
type because the general insertion logic can handle insertion into the
non-zero lane of the vector. However, in SSE4.1 with v4f32 vectors we
have INSERTPS that is a much better choice than the generic one for such
lowerings. But INSERTPS can do lots of other lowerings as well so
factoring its logic into the general insertion logic doesn't work very
well. We also can't just extract the core common part of the general
insertion logic that is faster (forming VZEXT_MOVL synthetic nodes that
lower to MOVSS when they can) because VZEXT_MOVL is often *faster* than
a blend while INSERTPS is slower! So instead we do a restrictive
condition on attempting to use the generic insertion logic to narrow it
to those cases where VZEXT_MOVL won't need a shuffle afterward and thus
will do better than INSERTPS. Then we try blending. Then we go back to
INSERTPS.
This still doesn't generate perfect code for some silly reasons that can
be fixed by tweaking the td files for lowering VZEXT_MOVL to use
XORPS+BLENDPS when available rather than XORPS+MOVSS when the input ends
up in a register rather than a load from memory -- BLENDPSrr has twice
the reciprocal throughput of MOVSSrr. Don't you love this ISA?
llvm-svn: 218177
analysis used elsewhere. This removes the last duplicate of this logic.
Also simplify the code here quite a bit. No functionality changed.
llvm-svn: 218176
floating point types and use it for both v2f64 and v2i64 single-element
insertion lowering.
This fixes the last non-AVX performance regression test case I've gotten
of for the new vector shuffle lowering. There is obvious analogous
lowering for v4f32 that I'll add in a follow-up patch (because with
INSERTPS, v4f32 requires special treatment). After that, its AVX stuff.
llvm-svn: 218175
vector lanes can be modeled as zero with a call to the new function that
computes a bit-vector representing that information.
No functionality changed here, but will allow doing more clever things
with the zero-test.
llvm-svn: 218174
I just tried reproducing some of the optimization failures in README.txt in the
X86 backend, and many of them could not be reproduced. In general the entire
file appears quite bit-rotted, whatever interesting parts remain should be
moved to bugzilla, and the rest deleted. I did not spend the time to do that,
so I just deleted the few I tried reproducing which are obsolete, to save some
time to whoever will find the courage to do it.
llvm-svn: 218170
lowering to support both anyext and zext and to custom lower for many
different microarchitectures.
Using this allows us to get *exactly* the right code for zext and anyext
shuffles in all the vector sizes. For v16i8, the improvement is *huge*.
The new SSE2 test case added I refused to add before this because it was
sooooo muny instructions.
llvm-svn: 218143
to undef lanes as well as defined widenable lanes. This dramatically
improves the lowering we use for undef-shuffles in a zext-ish pattern
for SSE2.
llvm-svn: 218115
shuffles that are zext-ing.
Not a lot to see here; the undef lane variant is better handled with
pshufd, but this improves the actual zext pattern.
llvm-svn: 218112
to the new vector shuffle lowering code.
This allows us to emit PMOVZX variants consistently for patterns where
it is a viable lowering. This instruction is both fast and allows us to
fold loads into it. This only hooks the new lowering up for i16 and i8
element widths, mostly so I could manage the change to the tests. I'll
add the i32 one next, although it is significantly less interesting.
One thing to note is that we already had some tests for these patterns
but those tests had far less horrible instructions. The problem is that
those tests weren't checking the strict start and end of the instruction
sequence. =[ As a consequence something changed in the lowering making
us generate *TERRIBLE* code for these patterns in SSE2 through SSSE3.
I've consolidated all of the tests and spelled out the madness that we
currently emit for these shuffles. I'm going to try to figure out what
has gone wrong here.
llvm-svn: 218102
There is no purpose in using it for single-input shuffles as
pshufd is just as fast and doesn't tie the two operands. This removes
a substantial amount of wrong-domain blend operations in SSSE3 mode. It
also completes the usage of PALIGNR for integer shuffles and addresses
one of the test cases Quentin hit with the new vector shuffle lowering.
There is still the question of whether and when to use this for floating
point shuffles. It is faster than shufps or shufpd but in the integer
domain. I don't yet really have a good heuristic here for when to use
this instruction for floating point vectors.
llvm-svn: 218038
PALIGNR. This just adds it to the v8i16 and v16i8 lowering steps where
it is completely unmatched. It also introduces the logic for detecting
rotation shuffle masks even in the presence of single input or blend
masks and arbitrarily undef lanes.
I've added fairly comprehensive tests for the matching logic in v8i16
because the tests at that size are much easier to write and manage.
I've not checked the SSE2 code generated for these tests because the
code is *horrible*. It is absolute madness. Testing it will just make
the test brittle without giving any interesting improvements in the
correctness confidence.
llvm-svn: 218013
This required a new hook called hasLoadLinkedStoreConditional to know whether
to expand atomics to LL/SC (ARM, AArch64, in a future patch Power) or to
CmpXchg (X86).
Apart from that, the new code in AtomicExpandPass is mostly moved from
X86AtomicExpandPass. The main result of this patch is to get rid of that
pass, which had lots of code duplicated with AtomicExpandPass.
llvm-svn: 217928
the blend that is matched by this are "used" in any sense, and so any
build_vector or other nodes feeding these will already drop other lanes.
llvm-svn: 217855
matching. This design just fundamentally didn't work because ADDSUB is
available prior to any legal lowerings of BLENDI nodes. Instead, we have
a dedicated ADDSUB synthetic ISD node which is pattern matched trivially
into the instructions. These nodes are then recognized by both the
existing and a trivial new lowering combine in the backend. Removing
these patterns required adding 2 missing shuffle masks to the DAG
combine, without which tests would have failed. Added the masks and
a helpful assert as well to catch if anything ever goes wrong here.
llvm-svn: 217851
that we don't use VSELECT and directly emit an addsub synthetic node.
Also remove a stale comment referencing VSELECT.
The test case is updated to use 'core2' which only has SSE3, not SSE4.1,
and it still passes. Previously it would not because we lacked
sufficient blend support to legalize the VSELECT.
llvm-svn: 217849
ADDSUBPD nodes out of blends of adds and subs.
This allows us to actually form these instructions with SSE3 rather than
only forming them when we had both SSE3 for the ADDSUB instructions and
SSE4.1 for the blend instructions. ;] Kind-of important.
I've adjusted the CPU requirements on one of the tests to demonstrate
this kicking in nicely for an SSE3 cpu configuration.
llvm-svn: 217848
introducing a synthetic X86 ISD node representing this generic
operation.
The relevant patterns for mapping these nodes into the concrete
instructions are also added, and a gnarly bit of C++ code in the
target-specific DAG combiner is replaced with simple code emitting this
primitive.
The next step is to generically combine blends of adds and subs into
this node so that we can drop the reliance on an SSE4.1 ISD node
(BLENDI) when matching an SSE3 feature (ADDSUB).
llvm-svn: 217819
Peephole optimization was folding MOVSDrm, which is a zero-extending double
precision floating point load, into ADDPDrr, which is a SIMD add of two packed
double precision floating point values.
(before)
%vreg21<def> = MOVSDrm <fi#0>, 1, %noreg, 0, %noreg; mem:LD8[%7](align=16)(tbaa=<badref>) VR128:%vreg21
%vreg23<def,tied1> = ADDPDrr %vreg20<tied0>, %vreg21; VR128:%vreg23,%vreg20,%vreg21
(after)
%vreg23<def,tied1> = ADDPDrm %vreg20<tied0>, <fi#0>, 1, %noreg, 0, %noreg; mem:LD8[%7](align=16)(tbaa=<badref>) VR128:%vreg23,%vreg20
X86InstrInfo::foldMemoryOperandImpl already had the logic that prevented this
from happening. However the check wasn't being conducted for loads from stack
objects. This commit factors out the logic into a new function and uses it for
checking loads from stack slots are not zero-extending loads.
rdar://problem/18236850
llvm-svn: 217799
when SSE4.1 is available.
This removes a ton of domain crossing from blend code paths that were
ending up in the floating point code path.
This is just the tip of the iceberg though. The real switch is for
integer blend lowering to more actively rely on this instruction being
available so we don't hit shufps at all any longer. =] That will come in
a follow-up patch.
Another place where we need better support is for using PBLENDVB when
doing so avoids the need to have two complementary PSHUFB masks.
llvm-svn: 217767
instructions from the relevant shuffle patterns.
This is the last tweak I'm aware of to generate essentially perfect
v4f32 and v2f64 shuffles with the new vector shuffle lowering up through
SSE4.1. I'm sure I've missed some and it'd be nice to check since v4f32
is amenable to exhaustive exploration, but this is all of the tricks I'm
aware of.
With AVX there is a new trick to use the VPERMILPS instruction, that's
coming up in a subsequent patch.
llvm-svn: 217761
instructions when it finds an appropriate pattern.
These are lovely instructions, and its a shame to not use them. =] They
are fast, and can hand loads folded into their operands, etc.
I've also plumbed the comment shuffle decoding through the various
layers so that the test cases are printed nicely.
llvm-svn: 217758
AVX is available, and generally tidy up things surrounding UNPCK
formation.
Originally, I was thinking that the only advantage of PSHUFD over UNPCK
instruction variants was its free copy, and otherwise we should use the
shorter encoding UNPCK instructions. This isn't right though, there is
a larger advantage of being able to fold a load into the operand of
a PSHUFD. For UNPCK, the operand *must* be in a register so it can be
the second input.
This removes the UNPCK formation in the target-specific DAG combine for
v4i32 shuffles. It also lifts the v8 and v16 cases out of the
AVX-specific check as they are potentially replacing multiple
instructions with a single instruction and so should always be valuable.
The floating point checks are simplified accordingly.
This also adjusts the formation of PSHUFD instructions to attempt to
match the shuffle mask to one which would fit an UNPCK instruction
variant. This was originally motivated to allow it to match the UNPCK
instructions in the combiner, but clearly won't now.
Eventually, we should add a MachineCombiner pass that can form UNPCK
instructions post-RA when the operand is known to be in a register and
thus there is no loss.
llvm-svn: 217755
'punpckhwd' instructions when suitable rather than falling back to the
generic algorithm.
While we could canonicalize to these patterns late in the process, that
wouldn't help when the freedom to use them is only visible during
initial lowering when undef lanes are well understood. This, it turns
out, is very important for matching the shuffle patterns that are used
to lower sign extension. Fixes a small but relevant regression in
gcc-loops with the new lowering.
When I changed this I noticed that several 'pshufd' lowerings became
unpck variants. This is bad because it removes the ability to freely
copy in the same instruction. I've adjusted the widening test to handle
undef lanes correctly and now those will correctly continue to use
'pshufd' to lower. However, this caused a bunch of churn in the test
cases. No functional change, just churn.
Both of these changes are part of addressing a general weakness in the
new lowering -- it doesn't sufficiently leverage undef lanes. I've at
least a couple of patches that will help there at least in an academic
sense.
llvm-svn: 217752
Chandler Carruth