+ add a comment to the destructor of tsymtable so that others (and maybe me a well) will hopefully see that the destructor of that class is not completely trivial!

* fixed some issues to get correct values of section_count, works better than before but it is still not perfect for unknown reasons

* forgotten part of 0039750

+ based on a patch by Laksen, optimize (v>=x) and (v<=y) into (v-x)<(y-x)

* support aitconst_*bit_unaligned in the intel assembler writer, resolves 0034222

* relax formula for max. nodes to inline

cleanup: remove ARMHF-related comments from RISC-V only codepath. no functional change.

* addendum for Mantis 0034333: also complain about number of arguments with the maximum number of arguments for Copy ( = 3 ) is exceeded

* is_range_test returns false if "lower" bound is greater than "upper" bound

* fix building with LLVM=1, thanks to Karl-Michael Schindler for the hint

+ patch by J. Gareth Moreton to support BMI2 instructions
+ extended avx test generator with the newly added BMI2 instructions

* Fixed overflow error when calling align() with corner case values such as align($7FFFFFFC, 4) for the longint variant.

Fix riscv64 compiler compilation with -dEXTDEBUG

Fix compilation with -dEXTDEBUG

Add -Tlinux and -Tembedded help lines for riscv32 and riscv64 compilers

* Fixed writing of method's RTTI for aarch64 by explicitly writing the $self parameter first.

* Reverted 0039947. According to Sven, the $self parameters is not required to be first parameter in RTTI. Appropriate tests will be fixed.

* Fixed overflow error for the qword version of align() when i=0.

* Undef FPC_HAS_TYPE_EXTENDED for x86_64-android.

* Reverted 0039961.

* avoid range check errors

* remove .NOTPARALLEL: from compiler Makefile, accelerates make fullcycle with -j a lot
* few cleanups

* typo

* fullinstall builds (and installs) utils now only once

* use the built compiler to install files, so the version is set correctly

* make install shall not make assumptions about FPC, if one was provided

+ make use of the BFI instruction in tcgaarch64.a_load_ref_reg_unaligned

* cleanup debug code

+ implement assembler optimization Str/LdrAdd/Sub2Str/Ldr Postindex done

* avoid to create a stack frame on aarch64 if possible

* take care of the fact that x and w registers overlap when carrying out the Str/LdrAdd/Sub2Str/Ldr Postindex optimization

* Optimized generic implementations of tcg.a_op_const_ref() and tcg.a_op_reg_ref() for PIC and CPUs which need 2 or more instructions to load a symbol's value by reference.
First the symbol's address is taken and then it is used to load the value and also store the result. It eliminates extra calculation of the symbol's address needed to store the result.
x86 and other CPUs which can handle refs ops directly or more efficiently already have overloaded versions of these methods and are not affected by this change.

* SetExecutionWeight(): Set execution weight also for nodes with special handling. Minor code cleanup/improvements.
* Restore previous execution weight before exiting secondpass().
* It all improves register allocation for variables inside loops in some cases.

* version update

* replaced "is" test with faster field value test

+ introduce tai_align_abstract.create_max which limits the maximal number of bytes being inserted for alignment

- x86 align does not use a register anymore, code removed

+ new fields for more precise control of alignment
+ support of new alignment control values for GNU AS writer
+ support of new alignment control values for binary writer
* better alignment settings on i386-win32
* improved inserted align directives for if nodes

* skip align directives after unconditional jumps

+ factored out TX86AsmOptimizer.PrePeepholeOptIMUL, used now by x86-64 and i386
* generalized and simplified the code

Disable three operand overload of MatchOpType for avr and jvm targets, which only have 2 operands max

* Removed 'inline' from public methods of TFPObjectList and TFPHashObjectList since they use private fields and can not be inlined when called from other units.

* Reverted 0040181 because the "not inlined" notes are issued not due to access to private fields, but due to string conversion. It need to be fixed by overloaded versions of some methods.

* Removed unused local vars.

* Commented out unused "size" local var.

* Removed unused local vars.

- get rid of MaxOps, it is redundant with max_operands
* MatchOpType with three operands is only available of max_operands>2

Fix 3 operand version of MatchOpType

* Moved inline methods to the beginning in order to allow them to be actually inlined in this unit.

* Rearranged the code to place implementation of inline methods before their usage within this unit.

+ Added overloaded version of stringdup() for ansistring to allow stringdup to be inlined.

* If a function can't be inline (contains an assembler block, etc) remove the po_inline flag from its procoptions.

* Removed inline for methods which are too big.

* Removed inline for functions which access private variables and can't be inlined from outside.

* Removed inline for Unix2AmigaPath since it is not necessary and can't be inlined anyway due to ansistring conversion.

* Addition to 0040225: Forgot to remove inline in the implementation.

* po_inline can't be removed from procoptions when the interface part is already parsed. It changes the interface's CRC and causes other units to recompile.

+ Added the pio_inline_not_possible flag to timplprocoption. It is set if a function can't be inline (contains an assembler block, etc). When calling such functions, no note "Call to subroutine "?" marked as inline is not inlined" is emitted.

* Changed verbosity of ""???" not yet supported inside inline procedure/function" to note to be consistent with note "Call to subroutine "?" marked as inline is not inlined".

+ initial work for tls-based threadvar support on arm-linux

+ check if initialization/finalization code uses implicitly thread vars

+ tsettings.tlsmodel

+ i386-linux support for tls-based threadvars

* define renamed into tls_threadvars

* compilation on x86-64 and i8086 fixed

* non-linux arm targets fixed

+ support for tlsm_general on i386-linux

* redundant code removed, work is meanwhile done by aasmcpu.optimize_ref

* i8086 compiler fixed, resolves 0034552

Avoid overflow error in riscv code generator

* do not re-used gottpoff symbols as they are PC-relative

Fix riscv32 compilation error introduced in last commit

Use same parameter type for update_reference_offset override

* better alignment on 386-linux

* three operand shl for shifting by 2, 4 or 8 can be simulated by lea

* set nf_write flag properly for vec. nodes of types with implicit dereferencing

* better alignment of while loop heads

* fixed adjustments of lower/upper bounds in range test optimization in case
of strictly smaller/greater comparisons (mantis 0034385)

* moved finalization of code generator temps to a node, so it can be getcopy'd
(needed for LLVM, where we need two copies of the finally code: one in case
an exception occurs, and one in case none is raised)
o also first finalize parameters and only then code generator temps, since
in theory the former could create more of the latter

+ new spilling heuristics which takes care of the usage weigth of a register,
not yet activated, can be selected by passing SPILLING_NEW when compiling the compiler

* finalizetemps nodes should never be considered equal and have side effects

* fixes arm

* don't generate tempsfinalize nodes in except filter procdefs, as that would
finalize the parent's temps twice (fix for 0040345, mantis 0034574)

+ MovOpMov2MovOp optimization

* building on i386-linux fixed

* disable MovOpMov2MovOp optimization on x86-64 for now as it seems to break

* pass volatility information to llvm.memcpy

* thlcgobj.g_unreachable() dummy implementation + llvm override

* fixed dynamic array constants for llvm

- reverted 0039188, as this hack removes the type correctness of the
instructions, which breaks the llvm code generator. Put such hacks
in CPU-specific overrides

+ customint torddef type to create arbitraty bit-width integers
o use this to handle non-power-of-two-sized parameters for llvm
o no general support in the parser/code generator, so don't expose

* fixed dwarf debug info after 0040398 (mantis 0034622)

* systematically capitalize first word of errormsg/sentence, patch by J. Gareth Moreton, mantis 34618

Regenerate msgid and msgtxt include files after commit 40422

m68k: reinstate the inc/dec tempregister type hack again in an isolated way (it was removed in 0040385, as it was breaking llvm target)

+ abi tag for i386
* arm abi tag moved into pascal file

- disable cs_opt_use_load_modify_store because for llvm because it is
apparently buggy as far as type information is concerned, and it also
results in tons of spilling

* ensure the generic version of tcgshlshrnode gets used for llvm

+ volatile() expression that marks an expression as volatile
* disable matching volatile references in the assembler optimisers, so they
can't be removed (more conservative than needed, but better than removing
too many)
o the CSE optimiser will ignore them by default, because they're an unknown
inline node for it
* also removed no longer used fpc_in_move_x and fpc_in_fillchar_x inline node
identifiers from rtl/inc/innr.inc, and placed fpc_in_unaligned_x at the
right place

* only free the temp holding the function result after we've loaded it into
the function return register (so we won't load from an already freed
temp once we emit lifetime info for temps on LLVM)

* emit temp lifetime information for llvm
o also allow freeing temps in the middle of a routine for llvm (we won't
reuse them, but it results in better lifetime information)

* add nocapture and dereferenceable/dereferenceable_or_null decorators to
parameters for llvm where appropriate

* fixed bug that caused the "sret" argument attribute to never be used (and
a guaranteed internalerror on AArch64/llvm)

* fixed paraloc def when changing the size to OS_32

* fixed shifval for big endian (bits instead of bytes)