Home - News ( United States | United Kingdom | Italy | Germany ) - Football scores

Showing content from https://releases.llvm.org/10.0.0/docs/ReleaseNotes.html below:

LLVM 10.0.0 Release Notes — LLVM 10 documentation

This document contains the release notes for the LLVM Compiler Infrastructure, release 10.0.0. Here we describe the status of LLVM, including major improvements from the previous release, improvements in various subprojects of LLVM, and some of the current users of the code. All LLVM releases may be downloaded from the LLVM releases web site.

For more information about LLVM, including information about the latest release, please check out the main LLVM web site. If you have questions or comments, the LLVM Developer’s Mailing List is a good place to send them.

• The ISD::FP_ROUND_INREG opcode and related code was removed from SelectionDAG.
• Enabled MemorySSA as a loop dependency. Since r370957 (D58311 [MemorySSA & LoopPassManager] Enable MemorySSA as loop dependency. Update tests.), the MemorySSA analysis is being preserved and used by a series of loop passes. The most significant use is in LICM, where the instruction hoisting and sinking relies on aliasing information provided by MemorySSA vs previously creating an AliasSetTracker. The LICM step of promoting variables to scalars still relies on the creation of an AliasSetTracker, but its use is reduced to only be enabled for loops with a small number of overall memory instructions. This choice was motivated by experimental results showing compile and run time benefits or replacing the AliasSetTracker usage with MemorySSA without any performance penalties. The fact that MemorySSA is now preserved by and available in a series of loop passes, also opens up opportunities for its use in those respective passes.
• The BasicBlockPass, BBPassManager and all their uses were deleted in this revision.
• The LLVM_BUILD_LLVM_DYLIB and LLVM_LINK_LLVM_DYLIB CMake options are no longer available on Windows.
• As per LLVM Language Reference Manual, getelementptr inbounds can not change the null status of a pointer, meaning it can not produce non-null pointer given null base pointer, and likewise given non-null base pointer it can not produce null pointer; if it does, the result is a poison value. Since r369789 (D66608 [InstCombine] icmp eq/ne (gep inbounds P, Idx..), null -> icmp eq/ne P, null) LLVM uses that for transformations. If the original source violates these requirements this may result in code being miscompiled. If you are using Clang front-end, Undefined Behaviour Sanitizer -fsanitize=pointer-overflow check will now catch such cases.
• Windows Control Flow Guard: the -cfguard option now emits CFG checks on indirect function calls. The previous behavior is still available with the -cfguard-nochecks option. Note that this feature should always be used with optimizations enabled.
• Callbacks have been added to CommandLine Options. These can be used to validate or selectively enable other options.
• The function attributes no-frame-pointer-elim and no-frame-pointer-elim-non-leaf have been replaced by frame-pointer, which has 3 values: none, non-leaf, and all. The values mean what functions should retain frame pointers.
• The inter-procedural analysis and optimization capabilities in the Attributor framework and pass have been substantially advanced (initial commit D59918, LLVM-Dev talk). In this release, 19 different attributes are inferred, including 12 LLVM IR attributes and 7 “abstract” attributes, such as liveness. The Attributor is still under heavy development and disabled by default; to enable an early run pass -mllvm -attributor-disable=false to an invocation of clang.
• New matrix math intrinsics have been added to LLVM (see LLVM Language Reference Manual), together with the LowerMatrixIntrinsics pass. The pass lowers matrix intrinsics to a set of efficient vector instructions. The lowering pass is off by default and can be enabled by passing -mllvm -enable-matrix to an invocation of clang.

• Unnamed function arguments now get printed with their automatically generated name (e.g. “i32 %0”) in definitions. This may require front-ends to update their tests; if so there is a script utils/add_argument_names.py that correctly converted 80-90% of Clang tests. Some manual work will almost certainly still be needed.
• A new freeze instruction is added. The freeze instruction is used to stop IR-level propagation of undef and poison values. Currently its support is preliminary; a freeze-equivalent operation for SelDag/MIR needs to be added.

• Added support for Cortex-A65, Cortex-A65AE, Neoverse E1 and Neoverse N1 cores.
• With a few more bugs fixed in the LLVM 10 release, clang-cl can now target Windows-on-ARM well, demonstrated by building complex pieces of software such as Chromium and the Electron framework.
• Support for -fpatchable-function-entry was added.

• Optimized ARMv8.1-M code generation, including generating Low Overhead Loops.
• Added auto-vectorization for the ARMv8.1-M MVE vector extension.
• Support was added for inline asm constraints s,j,x,N,O.

• Improved support for octeon and added support for octeon+ MIPS-family CPU.
• min, max, umin, umax atomics now supported on MIPS targets.
• Now PC-relative relocations are generated for .eh_frame sections when possible. That allows to link MIPS binaries without having to pass the -Wl,-z,notext option.
• Fix evaluating J-format branch (j, jal, …) targets when the instruction is not in the first 256 MB region.
• Fixed jal, sc, scs, ll, lld, la, lw, sw instructions expanding. Now they accept more types of expression as arguments, correctly handle load/store for XGOT model, expand using less instructions or registers.
• Initial MIPS support has been added to llvm-exegesis.
• Generates _mcount calls using proper MIPS ABI.
• Improved support of GlobalISel instruction selection framework. This feature is still in experimental state for MIPS targets though.

Optimization:

• Improved register pressure estimates in the loop vectorizer based on type
• Improved the PowerPC cost model for the vectorizer
• Enabled vectorization of math routines on PowerPC using MASSV (Mathematical Acceleration SubSystem) library

compiler-rt:

• Added/improved conversion functions from IBM long double to 128-bit integers

Codegen:

• Optimized memory access instructions in loops (pertaining to update-form instructions and address computation)
• Added options to disable hoisting instructions to hotter blocks based on statically or profile-based block hotness estimates
• Code generation improvements (particularly with floating point and vector code as well as handling condition registers)
• Various infrastructural improvements, code refactoring, and bug fixes
• Optimized handling of control flow based on multiple comparison of same values

Tools:

• llvm-readobj supports displaying file header, section headers, symbol table and relocation entries for XCOFF object files
• llvm-objdump supports disassembling physical sections for XCOFF object files

• Added support for the -march=z15 and -mtune=z15 command line options (as aliases to the existing -march=arch13 and -mtune=arch13 options).
• Added support for the -march=native command line option.
• Added support for the -mfentry, -mnop-mcount, and -mrecord-mcount command line options.
• Added support for the GHC calling convention.
• Miscellaneous codegen enhancements, in particular to enable better reuse of condition code values and improved use of conditional move instructions.

• Less-than-128-bit vector types, v2i32, v4i16, v2i16, v8i8, v4i8, and v2i8, are now stored in the lower bits of an xmm register and the upper bits are undefined. Previously the elements were spread apart with undefined bits in between them.
• v32i8 and v64i8 vectors with AVX512F enabled, but AVX512BW disabled will now be passed in ZMM registers for calls and returns. Previously they were passed in two YMM registers. Old behavior can be enabled by passing -x86-enable-old-knl-abi.
• -mprefer-vector-width=256 is now the default behavior skylake-avx512 and later Intel CPUs. This tries to limit the use of 512-bit registers which can cause a decrease in CPU frequency on these CPUs. This can be re-enabled by passing -mprefer-vector-width=512 to clang or passing -mattr=-prefer-256-bit to llc.
• Deprecated the mpx feature flag for the Intel MPX instructions. There were no intrinsics for this feature. This change only this effects the results returned by getHostCPUFeatures on CPUs that implement the MPX instructions.
• The feature flag fast-partial-ymm-or-zmm-write which previously disabled vzeroupper insertion has been removed. It has been replaced with a vzeroupper feature flag which has the opposite polarity. So -vzeroupper has the same effect as +fast-partial-ymm-or-zmm-write.

• __attribute__((used)) no longer implies that a symbol is exported, for consistency with other targets.
• Multivalue function signatures are now supported in WebAssembly object files
• The new atomic.fence instruction is now supported
• Thread-Local Storage (TLS) is now supported.
• SIMD support is significantly expanded.

New Features:

• The Machine Outliner is now supported, but not enabled by default.
• Shrink-wrapping is now supported.
• The Machine Scheduler has been enabled and scheduler descriptions for the Rocket micro-architecture have been added, covering both 32- and 64-bit Rocket cores.
• This release lays the groundwork for enabling LTO in a future LLVM release. In particular, LLVM now uses a new target-abi module metadata item to represent the chosen RISC-V psABI variant. Frontends should add this module flag to prevent ABI lowering problems when LTO is enabled in a future LLVM release.
• Support has been added for assembling RVC HINT instructions.
• Added code lowering for half-precision floats.
• The fscsr and frcsr (fssr, frsr) obsolete aliases have been added to the assembler for use in legacy code.
• The stack can now be realigned even when there are variable-sized objects in the same frame.
• fastcc is now supported. This is a more efficient, unstandardised, calling convention for calls to private leaf functions in the same IR module.
• llvm-objdump now supports -M no-aliases and -M numeric for altering the dumped assembly. These match the behaviour of GNU objdump, respectively disabling instruction aliases and printing the numeric register names rather than the ABI register names.

Improvements:

• Trap and Debugtrap now lower to RISC-V-specific trap instructions.
• LLVM IR Inline assembly now supports using ABI register names and using floating point registers in constraints.
• Stack Pointer adjustments have been changed to better match RISC-V’s immediates.
• ra (x1) can now be used as a callee-saved register.
• The assembler now suggests spelling corrections for unknown assembly mnemonics.
• Stack offsets of greater than 32-bits are now accepted on RV64.
• Variadic functions can now be tail-call optimised, as long as they do not use stack memory for passing arguments.
• Code generation has been changed for 32-bit arithmetic operations on RV64 to reduce sign-extensions.

Bug Fixes:

• There was an issue with register preservation after calls in interrupt handlers, where some registers were marked as preserved even though they were not being preserved by the call. This has been corrected, and now only callee-saved registers are live over a function call in an interrupt handler (just like calls in regular functions).
• Atomic instructions now only accept GPRs (plus an offset) in memory operands.
• Fixed some issues with evaluation of relocations and fixups.
• The error messages around missing RISC-V extensions in the assembler have been improved.
• The error messages around unsupported relocations have been improved.
• Non-PIC code no longer forces Local Exec TLS.
• There have been some small changes to the code generation for atomic operations.
• RISC-V no longer emits incorrect CFI directives in function prologues and epilogues.
• RV64 no longer clears the upper bits when returning complex types from libcalls using the LP64 psABI.

Compiler-RT:

• RISC-V (both 64-bit and 32-bit) is now supported by compiler-rt, allowing crtbegin and crtend to be built.
• The Sanitizers now support 64-bit RISC-V on Linux.

• C DebugInfo API LLVMDIBuilderCreateTypedef is updated to include an extra argument AlignInBits, to facilitate / propagate specified Alignment information present in a typedef to Debug information in LLVM IR.

• Go DebugInfo API CreateTypedef is updated to include an extra argument AlignInBits, to facilitate / propagate specified Alignment information present in a typedef to Debug information in LLVM IR.

• Improved support for building with MinGW
• Initial support for debugging Windows ARM and ARM64 binaries
• Improved error messages in the expression evaluator.
• Tab completions for command options now also provide a description for each option.
• Fixed that printing structs/classes with the expression command sometimes did not print the members/contents of the class.
• Improved support for using classes with bit-field members in the expression evaluator.
• Greatly improved support for DWARF v5.

Zig is a system programming language intended to be an alternative to C. It provides high level features such as generics, compile time function execution, and partial evaluation, while exposing low level LLVM IR features such as aliases and intrinsics. Zig uses Clang to provide automatic import of .h symbols, including inline functions and simple macros. Zig uses LLD combined with lazily building compiler-rt to provide out-of-the-box cross-compiling for all supported targets.

A wide variety of additional information is available on the LLVM web page, in particular in the documentation section. The web page also contains versions of the API documentation which is up-to-date with the Subversion version of the source code. You can access versions of these documents specific to this release by going into the llvm/docs/ directory in the LLVM tree.

If you have any questions or comments about LLVM, please feel free to contact us via the mailing lists.

RetroSearch is an open source project built by @garambo | Open a GitHub Issue

Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo

HTML: 3.2 | Encoding: UTF-8 | Version: 0.7.3