This patch adds support for vrndi_f32() and vrndiq_f32()
intrinsics in AArch32 mode and for vrndns_f32() intrinsic in
AArch64 mode.
Differential Revision: https://reviews.llvm.org/D48829
llvm-svn: 337690
As documented here: https://software.intel.com/en-us/node/682969 and
https://software.intel.com/en-us/node/523346. cpu_dispatch multiversioning
is an ICC feature that provides for function multiversioning.
This feature is implemented with two attributes: First, cpu_specific,
which specifies the individual function versions. Second, cpu_dispatch,
which specifies the location of the resolver function and the list of
resolvable functions.
This is valuable since it provides a mechanism where the resolver's TU
can be specified in one location, and the individual implementions
each in their own translation units.
The goal of this patch is to be source-compatible with ICC, so this
implementation diverges from the ICC implementation in a few ways:
1- Linux x86/64 only: This implementation uses ifuncs in order to
properly dispatch functions. This is is a valuable performance benefit
over the ICC implementation. A future patch will be provided to enable
this feature on Windows, but it will obviously more closely fit ICC's
implementation.
2- CPU Identification functions: ICC uses a set of custom functions to identify
the feature list of the host processor. This patch uses the cpu_supports
functionality in order to better align with 'target' multiversioning.
1- cpu_dispatch function def/decl: ICC's cpu_dispatch requires that the function
marked cpu_dispatch be an empty definition. This patch supports that as well,
however declarations are also permitted, since the linker will solve the
issue of multiple emissions.
Differential Revision: https://reviews.llvm.org/D47474
llvm-svn: 337552
The codegen for this builtin was initially implemented to match GCC.
However, due to interest from users GCC changed behaviour to account for the
big endian bias of the instruction and correct it. This patch brings the
handling inline with GCC.
Fixes https://bugs.llvm.org/show_bug.cgi?id=38192
Differential Revision: https://reviews.llvm.org/D49424
llvm-svn: 337449
This will convert the i8 mask argument to <8 x i1> and extract an i1 and then emit a select instruction. This replaces the '(__U & 1)" and ternary operator used in some of intrinsics. The old sequence was lowered to a scalar and and compare. The new sequence uses an i1 vector that will interoperate better with other mask intrinsics.
This removes the need to handle div_ss/sd specially in CGBuiltin.cpp. A follow up patch will add the GCCBuiltin name back in llvm and remove the custom handling.
I made some adjustments to legacy move_ss/sd intrinsics which we reused here to do a simpler extract and insert instead of 2 extracts and two inserts or a shuffle.
llvm-svn: 336622
This is part of an ongoing attempt at making 512 bit vectors illegal in the X86 backend type legalizer due to CPU frequency penalties associated with wide vectors on Skylake Server CPUs. We want the loop vectorizer to be able to emit IR containing wide vectors as intermediate operations in vectorized code and allow these wide vectors to be legalized to 256 bits by the X86 backend even though we are targetting a CPU that supports 512 bit vectors. This is similar to what happens with an AVX2 CPU, the vectorizer can emit wide vectors and the backend will split them. We want this splitting behavior, but still be able to use new Skylake instructions that work on 256-bit vectors and support things like masking and gather/scatter.
Of course if the user uses explicit vector code in their source code we need to not split those operations. Especially if they have used any of the 512-bit vector intrinsics from immintrin.h. And we need to make it so that merely using the intrinsics produces the expected code in order to be backwards compatible.
To support this goal, this patch adds a new IR function attribute "min-legal-vector-width" that can indicate the need for a minimum vector width to be legal in the backend. We need to ensure this attribute is set to the largest vector width needed by any intrinsics from immintrin.h that the function uses. The inliner will be reponsible for merging this attribute when a function is inlined. We may also need a way to limit inlining in the future as well, but we can discuss that in the future.
To make things more complicated, there are two different ways intrinsics are implemented in immintrin.h. Either as an always_inline function containing calls to builtins(can be target specific or target independent) or vector extension code. Or as a macro wrapper around a taget specific builtin. I believe I've removed all cases where the macro was around a target independent builtin.
To support the always_inline function case this patch adds attribute((min_vector_width(128))) that can be used to tag these functions with their vector width. All x86 intrinsic functions that operate on vectors have been tagged with this attribute.
To support the macro case, all x86 specific builtins have also been tagged with the vector width that they require. Use of any builtin with this property will implicitly increase the min_vector_width of the function that calls it. I've done this as a new property in the attribute string for the builtin rather than basing it on the type string so that we can opt into it on a per builtin basis and avoid any impact to target independent builtins.
There will be future work to support vectors passed as function arguments and supporting inline assembly. And whatever else we can find that isn't covered by this patch.
Special thanks to Chandler who suggested this direction and reviewed a preview version of this patch. And thanks to Eric Christopher who has had many conversations with me about this issue.
Differential Revision: https://reviews.llvm.org/D48617
llvm-svn: 336583
This case occurs in the intrinsic headers so we should avoid emitting the mask in those cases.
Factor the code into a helper function to make this easy.
llvm-svn: 336472
Shufflevector is easier to generate and matches what the backend pattern matches without relying on constant selects being turned into shuffles.
While I was there I also made the IR regular expressions a little stricter to ensure operand order on the shuffle.
llvm-svn: 336388
This patch removes on optimization used with the TRUE/FALSE
predicates, as was suggested in https://reviews.llvm.org/D45616
for r335339.
The optimization was buggy, since r335339 used it also
for *_mask builtins, without actually applying the mask -- the
mask argument was just ignored.
Reviewers: craig.topper, uriel.k, RKSimon, andrew.w.kaylor, spatel, scanon, efriedma
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D48715
llvm-svn: 336355
Additional IR is emitted to convert between scalar and vXi1 type to match the expected software inferface for the builtin that clang exposes.
llvm-svn: 335564
D48464 contains changes that will loosen some of the range checks in SemaChecking to a DefaultError warning that can be disabled.
This patch adds explicit masking to avoid using the upper bits of immediates to gracefully handle the warning being disabled.
llvm-svn: 335308
Summary: All *_sqrt_round_s[s|d] intrinsics should execute a square root on
zeroth element from B (Ops[1]) and insert in to A (Ops[0]), not the other way around.
Reviewers: itaraban, craig.topper
Reviewed By: craig.topper
Subscribers: craig.topper, cfe-commits
Differential Revision: https://reviews.llvm.org/D48288
llvm-svn: 334964
The previous names took the shift amount in bits to match gcc and required a multiply by 8 in the header. This creates a misleading error message when we check the range of the immediate to the builtin since the allowed range also got multiplied by 8.
This commit changes the builtins to use a byte shift amount to match the underlying instruction and the Intel intrinsic.
Fixes the remaining issue from PR37795.
llvm-svn: 334773
Summary: These intrinsics result in hint instructions. They are provided here for MSVC ARM64 compatibility.
Reviewers: mstorsjo, compnerd, javed.absar
Reviewed By: mstorsjo
Subscribers: kristof.beyls, chrib, cfe-commits
Differential Revision: https://reviews.llvm.org/D48132
llvm-svn: 334639
Summary: We've had these target independent intrinsics for at least a year and a half. Looks like they do exactly what we need here and the backend already supports them.
Reviewers: RKSimon, delena, spatel, GBuella
Reviewed By: RKSimon
Subscribers: cfe-commits, llvm-commits
Differential Revision: https://reviews.llvm.org/D47693
llvm-svn: 334366
We currently support them only in AArch64. The NEON Reference,
however, says they are 'ARMv7, ARMv8' intrinsics.
Differential Revision: https://reviews.llvm.org/D47446
llvm-svn: 334362
I'd like to make the select builtins require an avx512f, avx512bw, or avx512vl fature to match what is normally required to get masking. Truncate is special in that there are instructions with a 128/256-bit masked result even without avx512vl.
By using special buitlins we can emit a select without using the 128/256-bit select builtins.
llvm-svn: 334331
I'm looking into making the select builtins require avx512f, avx512bw, or avx512vl since masking operations generally require those features.
The extract builtins are funny because the 512-bit versions return a 128 or 256 bit vector with masking even when avx512vl is not supported.
llvm-svn: 334330
Test changes are due to differences in how we generate undef elements now. We also changed the types used for extractf128_si256/insertf128_si256 to match the signature of the builtin that previously existed which this patch resurrects. This also matches gcc.
llvm-svn: 334261
Adds support for these intrinsics, which are ARM and ARM64 only:
_interlockedbittestandreset_acq
_interlockedbittestandreset_rel
_interlockedbittestandreset_nf
_interlockedbittestandset_acq
_interlockedbittestandset_rel
_interlockedbittestandset_nf
Refactor the bittest intrinsic handling to decompose each intrinsic into
its action, its width, and its atomicity.
llvm-svn: 334239
We still emit shufflevector instructions we just do it from CGBuiltin.cpp now. This ensures the intrinsics that use this are only available on CPUs that support the feature.
I also added range checking to the immediate, but only checked it is 8 bits or smaller. We should maybe be stricter since we never use all 8 bits, but gcc doesn't seem to do that.
llvm-svn: 334237
We still lower them to native shuffle IR, but we do it in CGBuiltin.cpp now. This allows us to check the target feature and ensure the immediate fits in 8 bits.
This also improves our -O0 codegen slightly because we're able to see the zeroinitializer in the shuffle. It looks like it got lost behind a store+load previously.
llvm-svn: 334208
Summary:
We recently switch to using a selects in the intrinsics header files for FMA instructions. But the 512-bit versions support flavors with rounding mode which must be an Integer Constant Expression. This has forced those intrinsics to be implemented as macros. As it stands now the mask and mask3 intrinsics evaluate one of their macro arguments twice. If that argument itself is another intrinsic macro, we can end up over expanding macros. Or if its something we can CSE later it would show up multiple times when it shouldn't.
I tried adding __extension__ around the macro and making it an expression statement and declaring a local variable. But whatever name you choose for the local variable can never be used as the name of an input to the macro in user code. If that happens you would end up with the same name on the LHS and RHS of an assignment after expansion. We might be safe if we use __ in front of the variable names because those names are reserved and user code shouldn't use that, but I wasn't sure I wanted to make that claim.
The other option which I've chosen here, is to add back _mask, _maskz, and _mask3 flavors of the builtin which we will expand in CGBuiltin.cpp to replicate the argument as needed and insert any fneg needed on the third operand to make a subtract. The _maskz isn't truly necessary if we have an unmasked version or if we use the masked version with a -1 mask and wrap a select around it. But I've chosen to make things more uniform.
I separated out the scalar builtin handling to avoid too many things going on in EmitX86FMAExpr. It was different enough due to the extract and insert that the minor duplication of the CreateCall was probably worth it.
Reviewers: tkrupa, RKSimon, spatel, GBuella
Reviewed By: tkrupa
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D47724
llvm-svn: 334159
Factor out the common setjmp call emission code.
Based on a patch by Chris January
Differential Revision: https://reviews.llvm.org/D47784
llvm-svn: 334112
I tested these locally on an x86 machine by disabling the inline asm
codepath and confirming that it does the same bitflips as we do with the
inline asm.
Addresses code review feedback.
llvm-svn: 334059
Previously we were just using extended vector operations in the header file.
This unfortunately allowed non-constant indices to be used with the intrinsics. This is incompatible with gcc, icc, and MSVC. It also introduces a different performance characteristic because non-constant index gets lowered to a vector store and an element sized load.
By adding the builtins we can check for the index to be a constant and ensure its in range of the vector element count.
User code still has the option to use extended vector operations themselves if they need non-constant indexing.
llvm-svn: 334057
This builtin takes an index as its second operand, but the codegen hardcodes an index of 0 and doesn't use the operand. The only use of the builtin in the header file passes 0 to the operand so this works for that usage. But its more correct to use the real operand.
llvm-svn: 334054
We need to implement _interlockedbittestandset as a builtin for
windows.h, so we might as well do the whole family. It reduces code
duplication anyway.
Fixes PR33188, a long standing bug in our bittest implementation
encountered by Chakra.
llvm-svn: 333978
