I believe these all get canonicalized to vzext_movl. The only case where that wasn't true was when the load was loadi32 and the load was an extload aligned to 32 bits. But that was fixed in r364207.
Differential Revision: https://reviews.llvm.org/D63701
llvm-svn: 364337
We currently have some isel patterns for treating vzmovl+load the same as vzload, but that shrinks the load which we shouldn't do if the load is volatile.
Rather than adding isel checks for volatile. This patch removes the patterns and teachs DAG combine to merge them into vzload when its legal to do so.
Differential Revision: https://reviews.llvm.org/D63665
llvm-svn: 364333
Rename masked_load/masked_store to masked_ld/masked_st to discourage
their direct use. We need to check truncating/extending and
compressing/expanding before using them. This revealed that
our scalar masked load/store patterns were misusing these.
With those out of the way, renamed masked_load_unaligned and
masked_store_unaligned to remove the "_unaligned". We didn't
check the alignment anyway so the name was somewhat misleading.
Make the aligned versions inherit from masked_load/store instead
from a separate identical version. Merge the 3 different alignments
PatFrags into a single version that uses the VT from the SDNode to
determine the size that the alignment needs to match.
llvm-svn: 364150
128/256 bit scalar_to_vectors are canonicalized to (insert_subvector undef, (scalar_to_vector), 0). We have isel patterns that try to match this pattern being used by a vzmovl to use a 128-bit instruction and a subreg_to_reg.
This patch detects the insert_subvector undef portion of this and pulls it through the vzmovl, creating a narrower vzmovl and an insert_subvector allzeroes. We can then match the insertsubvector into a subreg_to_reg operation by itself. Then we can fall back on existing (vzmovl (scalar_to_vector)) patterns.
Note, while the scalar_to_vector case is the motivating case I didn't restrict to just that case. I'm also wondering about shrinking any 256/512 vzmovl to an extract_subvector+vzmovl+insert_subvector(allzeros) but I fear that would have bad implications to shuffle combining.
I also think there is more canonicalization we can do with vzmovl with loads or scalar_to_vector with loads to create vzload.
Differential Revision: https://reviews.llvm.org/D63512
llvm-svn: 364095
We already use vmovq for v2i64/v2f64 vzmovl. But we were using a
blendpd+xorpd for v4i64/v4f64/v8i64/v8f64 under opt speed. Or
movsd+xorpd under optsize.
I think the blend with 0 or movss/d is only needed for
vXi32 where we don't have an instruction that can move 32
bits from one xmm to another while zeroing upper bits.
movq is no worse than blendpd on any known CPUs.
llvm-svn: 364079
Summary:
llvm.x86.sse.stmxcsr only writes to memory.
llvm.x86.sse.ldmxcsr only reads from memory, and might generate an FPE.
Reviewers: craig.topper, RKSimon
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62896
llvm-svn: 363773
First step toward addressing the vector-reduce-mul-widen.ll regression in D63281 - we should replace ANY_EXTEND/ANY_EXTEND_VECTOR_INREG in X86ISelDAGToDAG to avoid having to add duplicate patterns when treating any extensions as legal.
In future patches this will also allow us to keep any extension nodes around a lot longer in the DAG, which should mean that we can keep better track of undef elements that otherwise become zeros that we think we have to keep......
Differential Revision: https://reviews.llvm.org/D63326
llvm-svn: 363655
The isel patterns for these use a bitcast and load/store, but
DAG combine should have canonicalized those away.
For the purposes of the memory folding table these opcodes can be
replaced by the MOVSSrm_alt/MOVSDrm_alt and MOVSSmr/MOVSDmr opcodes.
llvm-svn: 363644
Rename the old versions that use FR32/FR64 to MOVSSrm_alt/MOVSDrm_alt.
Use the new versions in patterns that previously used a COPY_TO_REGCLASS
to VR128. These patterns expect the upper bits to be zero. The
current set up appears to work, but I'm not sure we should be
enforcing upper bits being zero through a COPY_TO_REGCLASS.
I wanted to flip the arrangement and use a COPY_TO_REGCLASS to
FR32/FR64 for the patterns that need an f32/f64 result, but that
complicated fastisel and globalisel.
I've been doing some experiments with reducing some isel patterns
and ended up in a situation where I had a
(SUBREG_TO_REG (COPY_TO_RECLASS (VMOVSSrm), VR128)) and our
post-isel peephole was unable to avoid using an instruction for
the SUBREG_TO_REG due to the COPY_TO_REGCLASS. Having a VR128
instruction removes the COPY_TO_REGCLASS that was breaking this.
llvm-svn: 363643
Summary:
We can only use the memory form of cvtss2sd under optsize due to a partial register update. So previously we were emitting 2 instructions for extload when optimizing for speed. Also due to a late optimization in preprocessiseldag we had to handle (fpextend (loadf32)) under optsize.
This patch forces extload to expand so that it will always be in the (fpextend (loadf32)) form during isel. And when optimizing for speed we can just let each of those pieces select an instruction independently.
Reviewers: spatel, RKSimon
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62710
llvm-svn: 362919
Previously we did the equivalent operation in isel patterns with
COPY_TO_REGCLASS operations to transition. By inserting
scalar_to_vetors and extract_vector_elts before isel we can
allow each piece to be selected individually and accomplish the
same final result.
I ideally we'd use vector operations earlier in lowering/combine,
but that looks to be more difficult.
The scalar-fp-to-i64.ll changes are because we have a pattern for
using movlpd for store+extract_vector_elt. While an f64 store
uses movsd. The encoding sizes are the same.
llvm-svn: 362914
We already need to have patterns for X86ISD::RNDSCALE to support software intrinsics. But we currently have 5 sets of patterns for the 5 rounding operations. For of these 6 patterns we have to support 3 vectors widths, 2 element sizes, sse/vex/evex encodings, load folding, and broadcast load folding. This results in a fair amount of bytes in the isel table.
This patch adds code to PreProcessIselDAG to morph the fceil/ffloor/ftrunc/fnearbyint/frint to X86ISD::RNDSCALE. This way we can remove everything, but the intrinsic pattern while still allowing the operations to be considered Legal for DAGCombine and Legalization. This shrinks the DAGISel by somewhere between 9K and 10K.
There is one complication to this, the STRICT versions of these nodes are currently mutated to their none strict equivalents at isel time when the node is visited. This won't be true in the future since that loses the chain ordering information. For now I've also added support for the non-STRICT nodes to Select so we can change the STRICT versions there after they've been mutated to their non-STRICT versions. We'll probably need a STRICT version of RNDSCALE or something to handle this in the future. Which will take us back to needing 2 sets of patterns for strict and non-strict, but that's still better than the 11 or 12 sets of patterns we'd need.
We can probably do something similar for scalar, but I haven't looked at it yet.
Differential Revision: https://reviews.llvm.org/D62757
llvm-svn: 362535
These patterns can incorrectly narrow a volatile load from 128-bits to 64-bits.
Similar to PR42079.
Switch to using (v4i32 (bitcast (v2i64 (scalar_to_vector (loadi64))))) as the
load pattern used in the instructions.
This probably still has issues in 32-bit mode where loadi64 isn't legal. Maybe
we should use VZMOVL for widened loads even when we don't need the upper bits
as zeroes?
llvm-svn: 362203
We were turning roundss/sd/ps/pd intrinsics with immediates of 1 or 2 into
llvm.floor/ceil. The llvm.ceil/floor intrinsics are supposed to correspond
to the libm functions. For the libm functions we need to disable the
precision exception so the llvm.floor/ceil functions should always map to
encodings 0x9 and 0xA.
We had a mix of isel patterns where some used 0x9 and 0xA and others used
0x1 and 0x2. We need to be consistent and always use 0x9 and 0xA.
Since we have no way in isel of knowing where the llvm.ceil/floor came
from, we can't map X86 specific intrinsics with encodings 1 or 2 to it.
We could map 0x9 and 0xA to llvm.ceil/floor instead, but I'd really like
to see a use case and optimization advantage first.
I've left the backend test cases to show the blend we now emit without
the extra isel patterns. But I've removed the InstCombine tests completely.
llvm-svn: 361425
We require d/q suffixes on the memory form of these instructions to disambiguate the memory size.
We don't require it on the register forms, but need to support parsing both with and without it.
Previously we always printed the d/q suffix on the register forms, but it's redundant and
inconsistent with gcc and objdump.
After this patch we should support the d/q for parsing, but not print it when its unneeded.
llvm-svn: 360085
Reverts "[X86] Remove (V)MOV64toSDrr/m and (V)MOVDI2SSrr/m. Use 128-bit result MOVD/MOVQ and COPY_TO_REGCLASS instead"
Reverts "[TargetLowering][AMDGPU][X86] Improve SimplifyDemandedBits bitcast handling"
Eric Christopher and Jorge Gorbe Moya reported some issues with these patches to me off list.
Removing the CodeGenOnly instructions has changed how fneg is handled during fast-isel with sse/sse2. We're now emitting fsub -0.0, x instead
moving to the integer domain(in a GPR), xoring the sign bit, and then moving back to xmm. This is because the fast isel table no longer
contains an entry for (f32/f64 bitcast (i32/i64)) so the target independent fneg code fails. The use of fsub changes the behavior of nan with
respect to -O2 codegen which will always use a pxor. NOTE: We still have a difference with double with -m32 since the move to GPR doesn't work
there. I'll file a separate PR for that and add test cases.
Since removing the CodeGenOnly instructions was fixing PR41619, I'm reverting r358887 which exposed that PR. Though I wouldn't be surprised
if that bug can still be hit independent of that.
This should hopefully get Google back to green. I'll work with Simon and other X86 folks to figure out how to move forward again.
llvm-svn: 360066
The 128/256 bit version of these instructions require an 'x' or 'y' suffix to
disambiguate the memory form in att syntax.
We were allowing the same suffix in intel syntax, but it appears gas does not
do that.
gas does allow the 'x' and 'y' suffix on register and broadcast forms even
though its not needed. We were allowing it on unmasked register form, but not on
masked versions or on masked or unmasked broadcast form.
While there fix some test coverage holes so they can be extended with the 'x'
and 'y' suffix tests.
llvm-svn: 359418
Summary:
The register form of these instructions are CodeGenOnly instructions that cover
GR32->FR32 and GR64->FR64 bitcasts. There is a similar set of instructions for
the opposite bitcast. Due to the patterns using bitcasts these instructions get
marked as "bitcast" machine instructions as well. The peephole pass is able to
look through these as well as other copies to try to avoid register bank copies.
Because FR32/FR64/VR128 are all coalescable to each other we can end up in a
situation where a GR32->FR32->VR128->FR64->GR64 sequence can be reduced to
GR32->GR64 which the copyPhysReg code can't handle.
To prevent this, this patch removes one set of the 'bitcast' instructions. So
now we can only go GR32->VR128->FR32 or GR64->VR128->FR64. The instruction that
converts from GR32/GR64->VR128 has no special significance to the peephole pass
and won't be looked through.
I guess the other option would be to add support to copyPhysReg to just promote
the GR32->GR64 to a GR64->GR64 copy. The upper bits were basically undefined
anyway. But removing the CodeGenOnly instruction in favor of one that won't be
optimized seemed safer.
I deleted the peephole test because it couldn't be made to work with the bitcast
instructions removed.
The load version of the instructions were unnecessary as the pattern that selects
them contains a bitcasted load which should never happen.
Fixes PR41619.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61223
llvm-svn: 359392
Summary: If we have SSE2 we can use a MOVQ to store 64-bits and avoid falling back to a cmpxchg8b loop. If its a seq_cst store we need to insert an mfence after the store.
Reviewers: spatel, RKSimon, reames, jfb, efriedma
Reviewed By: RKSimon
Subscribers: hiraditya, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60546
llvm-svn: 359368
Many of our instructions have both a _Int form used by intrinsics and a form
used by other IR constructs. In the EVEX space the _Int versions usually cover
all the capabilities include broadcasting and rounding. While the other version
only covers simple register/register or register/load forms. For this reason
in EVEX, the non intrinsic form is usually marked isCodeGenOnly=1.
In the VEX encoding space we were less consistent, but usually the _Int version
was the isCodeGenOnly version.
This commit makes the VEX instructions match the EVEX instructions. This was
done by manually studying the AsmMatcher table so its possible I missed some
cases, but we should be closer now.
I'm thinking about using the isCodeGenOnly bit to simplify the EVEX2VEX
tablegen code that disambiguates the _Int and non _Int versions. Currently it
checks register class sizes and Record the memory operands come from. I have
some other changes I was looking into for D59266 that may break the memory check.
I had to make a few scheduler hacks to keep the _Int versions from being treated
differently than the non _Int version.
Differential Revision: https://reviews.llvm.org/D60441
llvm-svn: 358138
Years ago I moved this to an InstAlias using VR128H/VR128L. But now that we support {vex3} pseudo prefix, we need to block the optimization when it is set to match gas behavior.
llvm-svn: 358046
Scalar VEX/EVEX instructions don't use the L bit and don't look at it for decoding either.
So we should ignore it in our disassembler.
The missing instructions here were found by grepping the raw tablegen class definitions in
the tablegen debug output.
llvm-svn: 358040
The instruction's document this as W0 for the VEX encoding. But there's a
footnote mentioning that VEX.W is ignored in 64-bit mode. And the main VEX
encoding description says the VEX.W bit is ignored for instructions that are
equivalent to a legacy SSE instruction that uses REX.W to select a GPR which
would apply here.
By making this match EVEX we can remove a special case of allowing EVEX2VEX to
turn an EVEX.WIG instruction into VEX.W0.
llvm-svn: 357971
This custom inserter existed so we could do a weird thing where we pretended that the instructions support
a full address mode instead of taking a pointer in EAX/RAX. I think was largely so we could be pointer
size agnostic in the isel pattern.
To make this work we would then put the address into an LEA into EAX/RAX in front of the instruction after
isel. But the LEA is overkill when we just have a base pointer. So we end up using the LEA as a slower MOV
instruction.
With this change we now just do custom selection during isel instead and just assign the incoming address
of the intrinsic into EAX/RAX based on its size. After the intrinsic is selected, we can let isel take
care of selecting an LEA or other operation to do any address computation needed in this basic block.
I've also split the instruction into a 32-bit mode version and a 64-bit mode version so the implicit
use is properly sized based on the pointer. Without this we get comments in the assembly output about
killing eax and defing rax or vice versa depending on whether we define the instruction to use EAX/RAX.
llvm-svn: 357652
Includes a fix to emit a CheckOpcode for build_vector when immAllZerosV/immAllOnesV is used as a pattern root. This means it can't be used to look through bitcasts when used as a root, but that's probably ok. This extra CheckOpcode will ensure that the first match in the isel table will be a SwitchOpcode which is needed by the caching optimization in the ISel Matcher.
Original commit message:
Previously we had build_vector PatFrags that called ISD::isBuildVectorAllZeros/Ones. Internally the ISD::isBuildVectorAllZeros/Ones look through bitcasts, but we aren't able to take advantage of that in isel. Instead of we have to canonicalize the types of the all zeros/ones build_vectors and insert bitcasts. Then we have to pattern match those exact bitcasts.
By emitting specific matchers for these 2 nodes, we can make isel look through any bitcasts without needing to explicitly match them. We should also be able to remove the canonicalization to vXi32 from lowering, but I've left that for a follow up.
This removes something like 40,000 bytes from the X86 isel table.
Differential Revision: https://reviews.llvm.org/D58595
llvm-svn: 355784
This caused the first matcher in the isel table for many targets to Opc_Scope instead of Opc_SwitchOpcode. This leads to a significant increase in isel match failures.
llvm-svn: 355433
We have quite a few cases of using FP instructions for integer operations when only AVX1 is available. Then we switch to integer instructions with AVX2. In a lot of these cases execution domain fixing will take care of turning FP instructions into integer if its profitable.
With this patch we just keep on using the FP instructions even with AVX2. I've only handled some cases that don't require messing with patterns that are defined in the instruction definition. Those will require more subtle multiclass work possibly involving null_frag, hasSideEffects = 0, etc.
Differential Revision: https://reviews.llvm.org/D58470
llvm-svn: 355361
We were using VPBLENDW for v2i64 and VBLENDPD for v4i64. VPBLENDD has better throughput than VPBLENDW on some CPUs so it makes sense to use it when possible. VBLENDPD will probably become VBLENDD during execution domain fixing, but we might as well use integer in isel while we can.
This should work around some issues with the domain fixing pass prefering PBLENDW when we start with PBLENDW. There may still be some v8i16 cases that could use PBLENDD.
llvm-svn: 355281
Previously we had build_vector PatFrags that called ISD::isBuildVectorAllZeros/Ones. Internally the ISD::isBuildVectorAllZeros/Ones look through bitcasts, but we aren't able to take advantage of that in isel. Instead of we have to canonicalize the types of the all zeros/ones build_vectors and insert bitcasts. Then we have to pattern match those exact bitcasts.
By emitting specific matchers for these 2 nodes, we can make isel look through any bitcasts without needing to explicitly match them. We should also be able to remove the canonicalization to vXi32 from lowering, but I've left that for a follow up.
This removes something like 40,000 bytes from the X86 isel table.
Differential Revision: https://reviews.llvm.org/D58595
llvm-svn: 355224
These allows use to use the same set of isel patterns for sra/srl/shl which are undefined for out of range shifts and intrinsic shifts which aren't undefined.
Doing this late allows DAG combine to have every opportunity to optimize the sra/srl/shl nodes.
This removes about 7000 bytes from the isel table and simplies the td files.
llvm-svn: 355071
Summary:
Previously we used BLENDPS/BLENDPD but that puts the blend in the FP domain. Under optsize, the two address instruction pass can cause blendps/blendpd to commute to blendps/blendpd. But we probably shouldn't do that if the original type was a integer. So use pblendw instead.
Reviewers: spatel, RKSimon
Reviewed By: RKSimon
Subscribers: jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D58574
llvm-svn: 354755
r354363 caused https://crbug.com/934963#c1, which has a plain C reduced
test case.
I also had to revert some dependent changes:
- r354648
- r354647
- r354640
- r354511
llvm-svn: 354713
This avoids depending on the peephole pass to do load folding.
Also adds some load folding for some insert_subvector patterns that use blend.
All of this was found by temporarily adding TB_NO_FORWARD to the blend immediate entries in the load folding tables.
I've added -disable-peephole to some of the affected tests from that experiment to ensure we're testing isel patterns.
llvm-svn: 354511
D42042 introduced the ability for the ExecutionDomainFixPass to more easily change between BLENDPD/BLENDPS/PBLENDW as the domains required.
With this ability, we can avoid most bitcasts/scaling in the DAG that was occurring with X86ISD::BLENDI lowering/combining, blend with the vXi32/vXi64 vectors directly and use isel patterns to lower to the float vector equivalent vectors.
This helps the shuffle combining and SimplifyDemandedVectorElts be more aggressive as we lose track of fewer UNDEF elements than when we go up/down through bitcasts.
I've introduced a basic blend(bitcast(x),bitcast(y)) -> bitcast(blend(x,y)) fold, there are more generalizations I can do there (e.g. widening/scaling and handling the tricky v16i16 repeated mask case).
The vector-reduce-smin/smax regressions will be fixed in a future improvement to SimplifyDemandedBits to peek through bitcasts and support X86ISD::BLENDV.
Reapplied after reversion at rL353699 - AVX2 isel fix was applied at rL354358, additional test at rL354360/rL354361
Differential Revision: https://reviews.llvm.org/D57888
llvm-svn: 354363
This was the cause of the regression in D57888 - the commuted load pattern wasn't hidden by the predicate so once we enabled v4i32 blends on SSE41+ targets then isel was incorrectly matched against AVX2+ instructions.
llvm-svn: 354358
The VBROADCAST combines and SimplifyDemandedVectorElts improvements mean that we now more consistently use shorter (128-bit) X86vzload input operands.
Follow up to D58053
llvm-svn: 354346
D42042 introduced the ability for the ExecutionDomainFixPass to more easily change between BLENDPD/BLENDPS/PBLENDW as the domains required.
With this ability, we can avoid most bitcasts/scaling in the DAG that was occurring with X86ISD::BLENDI lowering/combining, blend with the vXi32/vXi64 vectors directly and use isel patterns to lower to the float vector equivalent vectors.
This helps the shuffle combining and SimplifyDemandedVectorElts be more aggressive as we lose track of fewer UNDEF elements than when we go up/down through bitcasts.
I've introduced a basic blend(bitcast(x),bitcast(y)) -> bitcast(blend(x,y)) fold, there are more generalizations I can do there (e.g. widening/scaling and handling the tricky v16i16 repeated mask case).
The vector-reduce-smin/smax regressions will be fixed in a future improvement to SimplifyDemandedBits to peek through bitcasts and support X86ISD::BLENDV.
Differential Revision: https://reviews.llvm.org/D57888
llvm-svn: 353610
Account for bypass delays when computing the latency of scalar int-to-float
conversions.
On Jaguar we need to account for an extra 6cy latency (see AMD fam16h SOG).
This patch also fixes the number of micropcodes for the register-memory variants
of scalar int-to-float conversions.
Differential Revision: https://reviews.llvm.org/D57148
llvm-svn: 352518
Craig Topper