Skip to content

Latest commit

 

History

History
248 lines (221 loc) · 10.7 KB

NEWS.md

File metadata and controls

248 lines (221 loc) · 10.7 KB

libFirm 1.22.1 (2016-01-07)

  • make: Fix cmake/make build
  • ia32: New just in time compilation mode which compiles into a memory buffer
  • amd64: Support PIC with PLT for ELF
  • ia32: Add ia32-get_ip={pop,thunk}
  • ia32: Generate mov $~1, %r; rol x, %r for ~(1 << x)
  • ia32: Generate mov $~0x80000000, %r; ror x, %r for ~(0x80000000 >> x)
  • be: Stub support for the asm constraint modifier %
  • amd64, ia32: Support the asm constraint e
  • be: Support the asm modifier c in all backends
  • ia32: Support JIT compilation
  • be: Improve permutation moving for copy coalescing
  • ir: Improve handling of negative overflow in float to int tarval conversion
  • amd64: Improve matching of immediates during instruction selection
  • amd64: Add peephole optimization mov $0, %r -> xorl %r, %r
  • amd64: Add peephole optimization lea c(%r), %r -> add $c, %r
  • amd64: Add peephole optimization lea (%r1, %r2), %r1 -> add %r2, %r1 and the commutated case
  • ia32: Add peephole optimization lea c(, %i, 2), %d -> lea c(%i, %i), %d
  • ia32: Add peephole optimization lea (%b), %d -> mov %b, %d
  • ia32: Add peephole optimization testl $0x0000XX00, %eRx -> testb $0xXX, %Rh
  • ia32: Generate slightly better code to load the floating-point constants -0.0 and -1.0
  • amd64: Reduce number of stack adjustments
  • api: Set the length of an array type solely when creating the array type
  • api: Set whether a function type is variadic and its calling convention and additional properties solely when creating the function type
  • api: Automatically infer the mode when creating Add, And, Div, Eor, Minus, Mod, Mul, Mulh, Mux, Not, Or, Shl, Shr, Shrs and Sub nodes
  • api: Remove the notion of atomic entities, use the initializer access functions instead
  • api: Remove visibility from types
  • api: Remove the type flag tf_variable_size, test the array size for 0 instead
  • api: Remove plist, use pdeq instead
  • api: Remove get_{class,segment,struct,union}_{ident,name}(), use get_compound_{ident,name}() instead
  • ir: Improve IR graph verifier
  • arm: Improve address mode use in instruction selection
  • be, ir: Improve preservation of debug info during transformations
  • amd64: Improve use of lea in instruction selection
  • sparc: Support the asm constaint K
  • sparc: Support computed goto
  • arm: Support computed goto
  • opt: Simplify computed goto with known destination to unconditional branch
  • opt: Handle constant folding for the builtins clz, ctz, ffs, parity and popcount
  • arm: Avoid redundant sign/zero extension right after a load
  • amd64, arm: Avoid redundant sign/zero extension right before a store
  • amd64: Avoid redundant sign/zero extension for truncation
  • amd64: Mark floating point constants as candidates for rematerialization
  • amd64: Improve block schedule by removing empty basic blocks
  • amd64: Use cltd/cqto for sign extension before a division
  • amd64: Use store with immediate for call arguments
  • opt: Add local optimization a * b [-1 <= b <= 0] -> -(a & b)
  • sparc: More strict checking of modifiers of placeholders in asm templates
  • arm: Support inline asm with with constraints I, J, K, L, M, Q, g, i, l, m, n and r as well as modifiers B, C and c
  • mips: Add new backend
  • be: By default permutation lowering now uses copy instead swap
  • amd64: Improve code generation for v & 0xFF/0xFFFF/0xFFFFFFFF
  • amd64, ia32: Always use potentially smaller address mode c(x,x,1) instead of c(,x,2)
  • opt: Add local optimization a >>s b >>u (n - 1) -> a >>u (n - 1)
  • ia32: Improve code generation for 64 bit multiplication if the lower half of a factor is 0 or 1
  • amd64: Add peephole optimization cmp $0, %r -> test %r, %r
  • amd64, ia32: Support the asm operand modifiers A, P, X and p
  • be: Accept # and % as register name prefix for gcc compatibility
  • arm, mips, sparc: Handle clobbers in asm
  • riscv: Add new backend
  • be: Improve modelling of register demand
  • amd64, ia32: Support all address modes in inline asm
  • ir, be: Support asm goto; even allow output constraints, which gcc and clang do not
  • Bugfixes

libFirm 1.22.0 (2015-12-31)

  • Improved PIC support, linux/elf is now supported
  • Many additional local optimization rules (instcombine)
  • Inline assembly support for sparc/amd64
  • Improved constant bit analysis
  • Improved load/store optimization, featuring compound value optimizations
  • Improved reassociation pass
  • Improved amd64 (aka x86_64) backend, it is now able to bootstrap libfirm
  • Improved arm backend (but still experimental)
  • Improved inliner (can inline compound types)
  • Compiletime performance improvements
  • Cleanups and API simplifications
  • Switch to C99 and corresponding code cleanup and simplification
  • More details in dumped IR graphs
  • Streamlined IR
  • Improved IR graph verifiers
  • Further unify backend infrastructure
  • Support for arrays as parameter and return types
  • Updated documentation
  • Perform precise 80bit spilling for x87 floating point
  • Improved permutation decomposition for copy coalescing
  • Added bitwise 'don't care' analysis and use it for optimizing occult constants
  • Support builtin 'saturating increment' (amd64, ia32, sparc)
  • Improved generated code for division by constant
  • Support builtin 'compare and swap' (amd64, ia32, sparc)
  • Support alias entities
  • Improved scalar conversion optimization
  • Added store combine optimization
  • Overhauled representation and handling of infinite loops
  • Improved use of flag results generated as side effect of instructions (ia32)
  • Improved calculation of execution frequency estimation for large graphs
  • Improved instruction selection (amd64, arm, ia32, sparc)
  • Improved x87 floating point code generation
  • Support early clobbers in inline assembly
  • Support x87 constraints in inline assembly (ia32)
  • Support __attribute__((visibility("hidden"))) and __attribute__((visible("protected")))
  • Optionally use CMake as build system
  • Support make install with the usual variables (INSTALLPREFIX, ...)
  • Bugfixes

libFirm 1.21.0 (2012-11-16)

  • Improvements of x86 backend (esp. x87 floatingpoint code)
  • Improvements to sparc backend (better use of delay slots)
  • Improved local optimization rules (esp. conversions)
  • Make compiler more deterministic
  • Bugfixes

libFirm 1.20.0 (2011-12-07)

  • Further improvemens to sparc backend (SPEC2000 works with softfloat)
  • Tuning of x86 backend
  • Software floatingpoint lowerer
  • Fixed firm profiling
  • New pass management allowing to specify pre-/postconditions for passes
  • Remove dependency on liblpp, add support for gurobi ILP solver
  • Experimental dwarf debugging support
  • Code cleanups, refactoring
  • Restructured API documentation
  • Bugfixes (we did a lot of csmith testing)

libFirm 1.19.1 (2011-05-17)

  • Fix some set_XXX functions not being exported in the shared library

libFirm 1.19.0 (2011-03-15)

  • Includes "SSA-Based Register Allocation with PBQP"
  • Improved Sparc backend
  • New (optimistic) fixpoint based value-range propagation/bit analysis
  • Code cleanup and refactoring
  • Bugfixes

libFirm 1.18.1 (2010-05-05)

  • Fix bug where stackframe was not always setup for -fno-omit-frame-pointer
  • bugfixes in Asm handling

libFirm 1.18.0 (2010-04-15)

  • Includes "Preference Guided Register Assignment" algorithm
  • Experimental Value Range Propagation
  • Loop Inversion and experimental Loop Unrolling code
  • Simplified construction interface. Most node constructors don't need graph/block arguments anymore.
  • Reworked type interface. Type names are optional now. Support for additional linkage types that among others support C++ 'linkonce' semantics now.
  • Small changes in constructors and node getters/setters (mostly adding 'const' to some getters)
  • code cleanup, smaller improvements in API specification
  • bugfixes

libFirm 1.17.0 (2009-05-15)

  • bugfixes
  • advanced load/store optimization which hoists loads out of loops
  • Internal restruturing: Alot of node structures are automatically generated from a specification file now.
  • Add support for multiple calling conventions
  • New experimental support for reading and writing programgraphs to disk
  • Support and optimizations for trampolines
  • fix PIC support

libFirm 1.16.0 (2009-01-28)

  • bugfixes
  • support for builtin nodes

libFirm 1.15.0 (2008-12-01)

  • bugfixes

libFirm 1.14.0 (2008-11-22)

  • Implementation of Clicks Combined Analysis/Optimizations
  • New switch lowering code
  • support for global asm statements
  • improved asm support
  • PIC support for Mac OS X
  • New register pressure minimizing scheduler
  • Improvements to spill algorithm
  • fix endless loop problems
  • further improve inlining heuristics
  • improve peephole optimizations for x86
  • bugfixes

libFirm 1.13.0 (2008-07-31)

  • VanDrunen's GVN-PRE fixed
  • operator strength reduce fixed and improved
  • fixed 64bit code generation for some rare compare cases
  • better tailrecursion optimization: handles x * func() and x + func()
  • improved inliner: better heuristics for inlining, can now inline recursive calls
  • improved spiller
  • lowering of CopyB nodes
  • better memory disambiguator
  • float->64bit conversion fixed for x87
  • removed old verbosity level based debugging: all modules use the new debug facility
  • Improved Confirm based optimization and conditional evaluation (using Confirm nodes)
  • BugFixes: tail recursion, load/store optimization, lowering of structure return, conditional evaluation, removal of unused methods
  • reduced numer of indirections for backend operation
  • ia32 Backend: supports more CPU architectures
  • ARM Backend: fixed frame access
  • support for special segments (like constructors, destructors)

libFirm 1.12.1 (2008-02-18)

  • bugfixes for new style initializers with bitfield types
  • make lowerer look at constant initializers too

libFirm 1.12.0 (2008-02-14)

  • dependency on libcore and libobstack dropped
  • there's an alternative easier to use way to construct compound initializers
  • bugfixes
  • improved support for exceptions
  • speed improvements
  • optimization of known libc functions

libFirm 1.11.0 (2008-11-05)

  • Lots of bugfixes
  • Compilation speed improved
  • Completely improved and rewritten handling of x86 address mode
  • Optimized Mul -> Lea,Shift,Add transformation
  • 64bit operations fixed and improved
  • More local optimizations
  • New backend peephole optimizations
  • Explicit status flag modeling (only for x86 for now)
  • Improvements of Load/Store optimization and alias analysis
  • All C benchmarks from Spec CINT2000 work now (with our edg frontend)