Currently it is implemented only for x86-CPUs supporting the FMA extension. While it would
be possible to implement it in software or simulate fma(<single>,<single>,<single>) using
double operations, it makes no sense in my opinion to do so.
git-svn-id: trunk@27564 -
+ x86 targets now directly call procedure variables located in references.
- a_call_ref method removed from tcg and thlcgobj.
git-svn-id: trunk@26666 -
- Don't call make_simple_ref if operation will be optimized away or replaced with another one, which calls make_simple_ref itself.
git-svn-id: trunk@26515 -
Although basic things work (no regressions in test suite, also with TEST_OPT=-O2), there are some secondary issues/TODOs:
- Exception frame around PASCALMAIN is not properly removed in DLLs
- No stack traces yet
- Stack overallocated in finalizer procedures, their entry/exit code needs cleanup
- Signals unit is probably completely broken.
git-svn-id: trunk@26225 -
While it is possible to use 16-bit LEA on the i8086, it's probably not worth
doing it in the code generator, because of the unpleasant register
restrictions that 16-bit references have.
git-svn-id: trunk@25989 -
* This should be applicable to i8086 targets as well (but needs testing), by copying changes from i386/cgcpu.pas to i8086/cgcpu.pas and removing "$ifdef i386" lines in cgx86.pas.
git-svn-id: trunk@25224 -
differentiate between 64 bit and 128 bit (R_SUBMMWHOLE) SSE vector regs,
and support spilling/assembling for R_SUBQ SSE registers (8 bytes)
(mantis #23962)
We currently never use the full 128 bit of an SSE register, and
spilling for those hasn't been implemented yet either (R_SUBMMWHOLE
SSE regs are spilled into a 4-byte temp currently -> can overwrite data)
git-svn-id: trunk@23700 -
* If possible, use reference base instead of index, this yields shorter instructions.
* Added comment about offset limits for rip-relative addressing.
- Removed code related to taking threadvar address on win32, it is incorrect because Windows TLS is not directly accessible via segment registers (fs:0x2c points to array of pointers to TLS storages of each module, so at least double indirection is needed).
git-svn-id: trunk@23342 -
symbols on darwin/ppc and darwin/i386, and also for common symbols on
darwin/ppc, as they're not required for other kinds of symbols on those
platforms (gcc doesn't use them either with -O1 and higher; it does use
them always for darwin/ppc64 except for local symbols, and hence so do we)
git-svn-id: trunk@22271 -
after all, although for some reason the problems only showed up with the
external assembler on x86-64 platforms other than Darwin. Now we never use
GOT entries on x86-64 anymore for local symbols, regardless of their nature,
and instead use plain RIP-relative addressing for them.
git-svn-id: trunk@21924 -
fist(t)pq instructions (mantis #22331)
* also internalerror in case the internal fpu stack position counter
becomes larger than 7, so we don't have to wait for the entire
byte to overflow before we detect a problem
git-svn-id: trunk@21734 -
pass on whether the symbol refers to code or data (important for
AIX weak symbols; note that AIX does not support undefined weak
external symbols, and FPC does not yet support weakly defining
non-external symbols, so this functionality isn't really used yet)
git-svn-id: trunk@20802 -
* Insert seh_endprologue directive after the last prologue instruction, not before it
* Omit seh_stackalloc for zero bytes
* (For now) ignore SEH directives in NASM and MASM writers, instead of failing with internal error.
git-svn-id: trunk@19246 -
* For now placed actual processing of unwind info under {$ifdef TEST_WIN64_UNWIND}, because in the current state it doesn't add much value.
git-svn-id: trunk@19200 -
relocations for addr_pic_no_got references (do not go via the GOT)
* fixed code generator so that it generates addr_pic_no_got references
also for platforms that use the internal assembler (this will break linking
of some tests that check static data > 2GB on Linux/x86-64, but they were
already broken before as well because we don't use the "huge" x86-64
execution model)
* fixed x86-64 Linux syscall code to not expect GOT-based relocations for
the system unit's errno variable (it's an implementation-only variable)
git-svn-id: trunk@16717 -
later: the same as i386/darwin, except
a) uses the non-fragile Objective-C ABI/runtime
b) does not require stubs for direct calls/jumps (not required for
i386/darwin under 10.6 and later either, but still generated
there for backwards compatibility)
c) only the same packages are enabled as for ARM/Darwin
d) MacOSAll is compiled specifically for the iPhoneSimulator SDK
This target also defines the symbol "darwin" apart from the target
name "iphonesim" for source code compatibility reasons.
git-svn-id: trunk@16065 -
- sort of reverted r14134, which is no longer required after the above
change (new_section() inserts the alignment itself)
* made the tai_section.create() constructor private so it cannot be
called directly anymore
git-svn-id: trunk@15482 -
used to store a record function result in an MM/FPUREGISTER to memory
(the register size will be OS_F32/OS_F64, while the memory size will
be OS_32/OS_64 -> convert the latter to the former)
* extended tcalext6 to also test this scenario
git-svn-id: trunk@15385 -
because on linux/x86_64
a) the resulting code crashes when using the internal assembler
b) refuses to link when using the external assembler
(and hence I don't know whether it's safe to do on other non-darwin
platforms either) -- fixes mantis #15889
git-svn-id: trunk@14971 -
