Only perform this if we are shuffling lower and upper lane elements across the lanes (otherwise splitting to lower xmm shuffles would be better).
This is a regression if we shuffle build_vectors due to getVectorShuffle canonicalizing 'blend of splat' build vectors, for now I've set this not to shuffle build_vector nodes at all to avoid this.
The primary motivation of this change is to bring the code more closely in sync behavior wise with the assembler's version of nop emission. I'd like to eventually factor them into one, but that's hard to do when one has features the other doesn't.
The longest encodeable nop on x86 is 15 bytes, but many processors - for instance all intel chips - can't decode the 15 byte form efficiently. On those processors, it's better to use either a 10 byte or 11 byte sequence depending.
Add initial support for lowering v4f64 shuffles to SHUFPD(VPERM2F128(V1, V2), VPERM2F128(V1, V2)), eventually this could be used for v8f32 (and maybe v8f64/v16f32) but I'm being conservative for the initial implementation as only v4f64 can always succeed.
This currently is only called from lowerShuffleAsLanePermuteAndShuffle so only gets used for unary shuffles, and we limit this to cases where we use upper elements as otherwise concating 2 xmm shuffles is probably the better case.
Helps with poor shuffles mentioned in D66004.
ONE is currently softened to OGT | OLT. But the libcalls for OGT and OLT libcalls will trigger an exception for QNAN. At least for X86 with libgcc. UEQ on the other hand uses UO | OEQ. The UO and OEQ libcalls will not trigger an exception for QNAN.
This patch changes ONE to use the inverse of the UEQ lowering. So we now produce O & UNE. Technically the existing behavior was correct for a signalling ONE, but since I don't know how to generate one of those from clang that seemed like something we can deal with later as we would need to fix other predicates as well. Also removing spurious exceptions seemed better than missing an exception.
There are also problems with quiet OGT/OLT/OLE/OGE, but those are harder to fix.
Differential Revision: https://reviews.llvm.org/D72477
This system wasn't very well designed for multi-result nodes. As
a consequence they weren't consistently registered in the
LegalizedNodes map leading to nodes being revisited for different
results.
I've removed the "Result" variable from the main LegalizeOp method
and used a SDNode* instead. The result number from the incoming
Op SDValue is only used for deciding which result to return to the
caller. When LegalizeOp is called it should always register a
legalized result for all of its results. Future calls for any other
result should be pulled for the LegalizedNodes map.
Legal nodes will now register all of their results in the map
instead of just the one we were called for.
The Expand and Promote handling to use a vector of results similar
to LegalizeDAG. Each of the new results is then re-legalized and
logged in the LegalizedNodes map for all of the Results for the
node being legalized. None of the handles register their own
results now. And none call ReplaceAllUsesOfValueWith now.
Custom handling now always passes result number 0 to LowerOperation.
This matches what LegalizeDAG does. Since the introduction of
STRICT nodes, I've encountered several issues with X86's custom
handling being called with an SDValue pointing at the chain and
our custom handlers using that to get a VT instead of result 0.
This should prevent us from having any more of those issues. On
return we will update the LegalizedNodes map for all results so
we shouldn't call the custom handler again for each result number.
I want to push SDNode* further into the Expand and Promote
handlers, but I've left that for a follow to keep this patch size
down. I've created a dummy SDValue(Node, 0) to keep the handlers
working.
Differential Revision: https://reviews.llvm.org/D72224
Summary:
Avoid using the `nocf_check` attribute with Control Flow Guard. Instead, use a
new `"guard_nocf"` function attribute to indicate that checks should not be
added on indirect calls within that function. Add support for
`__declspec(guard(nocf))` following the same syntax as MSVC.
Reviewers: rnk, dmajor, pcc, hans, aaron.ballman
Reviewed By: aaron.ballman
Subscribers: aaron.ballman, tomrittervg, hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D72167
In D71841 we inverted the sense of the SDNode-level flag to ensure all nodes
default to potentially raising FP exceptions unless otherwise specified --
i.e. if we forget to propagate the flag somewhere, the effect is now only
lost performance, not incorrect code.
However, the related flag at the MI level still defaults to nodes not raising
FP exceptions unless otherwise specified. To be fully on the (conservatively)
safe side, we should invert that flag as well.
This patch does so by replacing MIFlag::FPExcept with MIFlag::NoFPExcept.
(Note that this does also introduce an incompatible change in the MIR format.)
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D72466
The ONE expansion calls OGT/OLT libcalls which will signal for QNAN.
The UEQ expansion uses unord and eq libcalls which won't signal.
We should probably use those libcalls for ONE with appropriate
logic.
Quiet OGT/OLT/OLE/OGE have similar issue, but not sure how to fix
those yet.
If we're doing a compare that only tests the sign bit and only the sign bit is demanded, we can just bypass the node. This removes one of the blend dependencies in our v2i64->v2f32 uint_to_fp codegen on pre-sse4.2 targets.
Differential Revision: https://reviews.llvm.org/D72356
If we are extracting a chunk of a vector that's a fraction of an
operand of the concatenated vector operand, we can extract directly
from one of those original operands.
This is another suggestion from PR42024:
https://bugs.llvm.org/show_bug.cgi?id=42024#c2
But I'm not sure yet if it will make any difference on those patterns.
It seems to help a few existing AVX512 tests though.
Differential Revision: https://reviews.llvm.org/D72361
For v4i64->v4f32 uint_to_fp on pre-avx targets where v4i64 isn't legal we create to v2i64->v2f32 uint_to_fp that need to be shuffled together. Our codegen for v2i64->v2f32 involves detecting if the number is larger than (2^31 - 1), if so we do a special divison by 2 so we can do a signed conversion which we need to scalarize, then do a multiply by 2 at the end if we divided earlier.
When v4i64 isn't legal we need to split the checking for a larger number and dividing by 2 into two v2i64 vectors. The scalar part can extract the 4 i64 values from those 4 splits. But we can reassemble the 4 scalar f32 results directly into a single v432 vector. Then we just need to combine the fixup indications from the 2 halves and we can do the final multiply by 2 fixup on all 4 values if needed at once using a single v4f32 blend and v4f32 fadd.
Differential Revision: https://reviews.llvm.org/D72368
As discussed heavily in the original review (D70157), there's a need for the compiler to be able to selective suppress padding (either nop or prefix) to respect assumptions about the meaning of labels and instructions in generated code.
Rather than wait for syntax to be finalized - which appears to be a very slow process - this patch focuses on the compiler use case and *only* worries about the integrated assembler. To my knowledge, this covers all cases mentioned to date for clang/JIT support.
For testing purposes, I wired it up so that if the integrated assembler was using autopadding for branch alignment (e.g. enabled at command line) then the textual assembly output would contain a comment for each location where padding was enabled or disabled. This seemed like the least painful choice overall.
Note that the result of this patch effective disables the jcc errata mitigation for many constructs (statepoints, implicit null checks, xray, etc...) which is non ideal. It is at least *correct* and should allow us to enable the mitigation for the compiler. Once that's done, and a few other items are worked through, we probably want to come back to this an explore a bundling based approach instead so that we can pad instructions while keeping labels in the right place.
Differential Revision: https://reviews.llvm.org/D72303
Summary:
This patch adds intrinsics and ISelDAG nodes for
signed and unsigned fixed-point division:
llvm.sdiv.fix.*
llvm.udiv.fix.*
These intrinsics perform scaled division on two
integers or vectors of integers. They are required
for the implementation of the Embedded-C fixed-point
arithmetic in Clang.
Patch by: ebevhan
Reviewers: bjope, leonardchan, efriedma, craig.topper
Reviewed By: craig.topper
Subscribers: Ka-Ka, ilya, hiraditya, jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D70007
Summary:
Remove the restrictions that preventing "asm goto" from returning non-void
values. The values returned by "asm goto" are only valid on the "fallthrough"
path.
Reviewers: jyknight, nickdesaulniers, hfinkel
Reviewed By: jyknight, nickdesaulniers
Subscribers: rsmith, hiraditya, llvm-commits, cfe-commits, craig.topper, rnk
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D69876
Now that we generate decent code for (v2i64 (setlt zero, X)) on pre-sse4.2 targets I think we can use this now.
Differential Revision: https://reviews.llvm.org/D72354
Similar to D72302 but for the canonical form for the opposite case. I've changed foldVectorXorShiftIntoCmp to form a target independent setcc node instead of PCMPGT now and enabled its for v2i64 on pre-SSE4.2 targets. The setcc should eventually get lowered to PCMPGT or the new v2i64 sequence.
Differential Revision: https://reviews.llvm.org/D72318
Without sse4.2 a v2i64 setlt needs to expand into a pcmpgtd, pcmpeqd, 3 shuffles, and 2 logic ops. But if we're only interested in the sign bit of the i64 elements, we can just use one pcmpgtd and shuffle the odd elements to the even elements.
Differential Revision: https://reviews.llvm.org/D72302
Don't overwrite existing target-cpu attributes.
I've often found the replacement behavior annoying, and this is
inconsistent with how the fast math command line flags interact with
the function attributes.
Does not yet change target-features, since I think that should behave
as a concatenation.
This is possibly a small part towards solving PR42024:
https://bugs.llvm.org/show_bug.cgi?id=42024
The vectorizer is creating shuffles of concat like this:
%63 = shufflevector <4 x i64> %x, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
%64 = shufflevector <8 x i64> %63, <8 x i64> undef, <8 x i32> <i32 0, i32 4, i32 1, i32 5, i32 2, i32 6, i32 3, i32 7>
That might be fixable in the vectorizers, but we're not allowed to fold that into a single shuffle in instcombine,
so we should have a backend backstop to convert that into the likely simpler form:
%64 = shufflevector <4 x i64> %x, <4 x i64> undef, <8 x i32> <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
Differential Revision: https://reviews.llvm.org/D72300
Summary:
It's not necessary to use an 'l'(ell) modifier when referencing a label.
Treat block addresses and MBB references as if the modifier is used
anyway. This prevents us from generating references to ficticious
labels.
Reviewers: jyknight, nickdesaulniers, hfinkel
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71849
This patch adds widening which really just scalarizes because we don't have a strategy for the extra elements we would need to pad with.
Differential Revision: https://reviews.llvm.org/D72193
Summary:
Based on Simon's D52965, but improved to handle strict fp and improve some of the shuffling.
Rather than use v2i1/v4i1 and let type legalization continue, just generate all the code with legal types and use an explicit shuffle.
I also added an explicit setcc to the v4i64 code to match the semantics of vselect which doesn't just use the sign bit. I'm also using a v4i64->v4i32 truncate instead of the shuffle in Simon's original code. With the setcc this will become a pack.
Future work can look into using X86ISD::BLENDV and a different shuffle that only moves the sign bit.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71956
Attempt to use combineLogicBlendIntoConditionalNegate for (select M, (sub 0, X), X) -> (sub (xor X, M), M)
We limit this to cases that can't easily replace the VSELECT with a shuffle (non-constant masks) or where a BLENDV is likely to occur (which tends to result in slower codegen).
The code here isn't great in all caess. Particularly v4f64->v4i32
on 64-bit AVX targets. But there is some improvement in some
configurations.
There's definitely some issues with computeNumSignBits with
X86ISD::STRICT_FCMP. As well as not being able to propagate sign
bits through merge_values nodes that get created during custom
legalization.
This patch attempts to peek through vectors based on the demanded bits/elt of a particular ISD::EXTRACT_VECTOR_ELT node, allowing us to avoid dependencies on ops that have no impact on the extract.
In particular this helps remove some unnecessary scalar->vector->scalar patterns.
The wasm shift patterns are annoying - @tlively has indicated that the wasm vector shift codegen are to be refactored in the near-term and isn't considered a major issue.
Reapplied after reversion at rL368660 due to PR42982 which was fixed at rGca7fdd41bda0.
Differential Revision: https://reviews.llvm.org/D65887
This represents a more realistic version of the code being tested.
The cmov converter doesn't look at the code after the loop so
it doesn't matter for what's being tested.
But as noted in this twitter thread https://twitter.com/trav_downs/status/1213311159413161987
gcc can turn the previous MaxIndex code into the MaxValue code. So
returning the index makes it a distinct case.
This uses an alternative implementation of this conversion derived
from our v2i32->v2f32 handling. We can zero extend the v2i32 to
v2i64, or it with the bit representation of 2.0^52 which will give
us 2.0^52 plus the 32-bit integer since double's mantissa is 52 bits.
Then we just need to subtract 2.0^52 as a double and let the floating
point unit normalize the remaining bits into a valid double.
This is less instructions then our previous code, but does require
a port 5 shuffle for the zero extend or unpack.
Differential Revision: https://reviews.llvm.org/D71945
The fold 'A - (A & (B - 1))' -> 'A & (0 - B)'
added in 8dab0a4a7d
is too specific. It should/can just be 'A - (A & B)' -> 'A & (~B)'
Even if we don't manage to fold `~` into B,
we have likely formed `ANDN` node.
Also, this way there's less similar-but-duplicate folds.
Name: X - (X & Y) -> X & (~Y)
%o = and i32 %X, %Y
%r = sub i32 %X, %o
=>
%n = xor i32 %Y, -1
%r = and i32 %X, %n
https://rise4fun.com/Alive/kOUl
See
https://bugs.llvm.org/show_bug.cgi?id=44448https://reviews.llvm.org/D71499
Simon Pilgrim