paweld/fpc
1
0
Fork 0
mirror of https://gitlab.com/freepascal.org/fpc/source.git synced 2025-07-16 08:26:04 +02:00
Code Issues Packages Projects Releases Wiki Activity
58,234 Commits 100 Branches 76 Tags 267 MiB
4ba19f5418
Commit Graph

415 Commits

Author SHA1 Message Date
Jonas Maebe
0afbe85aab * various memory reference alignment fixes 
git-svn-id: trunk@34544 -
 2016-09-20 21:43:19 +00:00
Jonas Maebe
aa1be3276f - removed default value of _typ parameter of TAsmData.(Weak)RefAsmSymbol(): 
it was AT_NONE, which is invalid and should never be used
  * explicitly pass the correct value for all calls to those methods elsewhere
    in the compiler

git-svn-id: trunk@34250 -
 2016-08-05 07:09:16 +00:00
Jonas Maebe
1cb8c0d00c * specify the def of assembler level symbols defined via 
tasmdata.DefineAsmSymbol() and all routines that call it
   o will be used to automatically generate AB_INDIRECT sybols when
     necessary

git-svn-id: trunk@34164 -
 2016-07-20 20:52:59 +00:00
pierre
fb50d0860b Use BX register instead of AX to reload DS at proc entry if register convention is used 
git-svn-id: trunk@34024 -
 2016-06-29 20:48:51 +00:00
nickysn
c78f406d99 + implemented proper stack checking for the i8086 
git-svn-id: trunk@33787 -
 2016-05-24 23:57:47 +00:00
florian
d2b74be153 * always use vmov variants of instructions if avx is enabled 
git-svn-id: trunk@33554 -
 2016-04-24 20:03:15 +00:00
florian
a742df9035 * reverse merged r33524 as it is not safe as test results showed 
--- Reverse-merging r33524 into '.':
U    compiler\i386\popt386.pas
U    compiler\x86\cgx86.pas
--- Recording mergeinfo for reverse merge of r33524 into '.':
 U   .

git-svn-id: trunk@33527 -
 2016-04-17 11:33:29 +00:00
florian
f576b0c01b * make use of xor reg,reg by generating it directly instead of hoping for the peephole 
optimizer which cannot do this properly due to missing information about flags. By doing
  so the size of the compiler executable gets reduced by ~1 %

git-svn-id: trunk@33524 -
 2016-04-15 19:27:22 +00:00
nickysn
c4343c1a2d * don't call make_simple_ref in tcgx86.g_concatcopy on i8086, because that 
screws up the handling of segmented references badly and breaks the large,
  compact and huge memory models completely

git-svn-id: trunk@33314 -
 2016-03-22 13:56:10 +00:00
yury
3c6d4f1ee2 * Removed unused vars. 
git-svn-id: trunk@33308 -
 2016-03-21 09:21:24 +00:00
svenbarth
f297b00f5b Extend the x86 targets by the ability to handle indirect symbols. 
x86/cgx86.pas, tcgx86:
  + new method make_direct_ref() which is used to convert an indirect reference into a direct one (uses the boolean field in_make_direct_ref to avoid recursive calls)
  * make_simple_ref: call make_direct_ref() before anything else
  * a_loadaddr_ref_ref: call make_direct_ref() (the loading could probably be folded into the loadaddr method, but for now that is sufficent)
i386/cgcpu.pas, tcg386:
  * a_loadaddr_ref_cgpara: call make_direct_ref(); the same remark as for a_loadaddr_ref_ref() applies here

git-svn-id: trunk@33280 -
 2016-03-18 21:45:41 +00:00
svenbarth
77ede2ac9f i386/cgcpu.pas, tcg386: 
* a_load_ref_cgpara: call make_simple_ref() before calling the base a_load_ref_cgpara()
x86/cgx86.pas, tcgx86:
  * a_loadfpu_ref_reg, a_loadfpu_reg_ref, g_concatcopy: call make_simple_ref() on the passed references

git-svn-id: trunk@33277 -
 2016-03-18 21:22:04 +00:00
sergei
1b965e6766 * x86: Don't save/restore integer registers that are volatile per calling convention of current procedure. It implies that nothing will be saved for procedures with OLDFPCCALL, FAR16 and PASCAL and calling conventions. OLDFPCCALL restores behavior that was in effect before r25224. 
git-svn-id: trunk@32542 -
 2015-11-27 05:48:49 +00:00
nickysn
b26e10d53c + generate proper win16 prologue/epilogue for exported routines (we don't yet 
actually export them from the .exe or .dll, but they can now be used as
  callbacks)
+ support win16 smart callbacks via the $K switch

git-svn-id: trunk@31835 -
 2015-09-26 18:04:59 +00:00
yury
e04a3b4ac6 * Removed unused vars for x86 compiler. 
git-svn-id: trunk@31743 -
 2015-09-17 15:20:15 +00:00
florian
40a5f97239 - cclasses unit is not used 
git-svn-id: trunk@31672 -
 2015-09-13 21:07:21 +00:00
nickysn
bdfd23cc2e + new target switch "FarProcsPushOddBP", which causes the i8086 compiler to push 
odd values of BP if the procedure is far. Enabled this by default for Win16.

git-svn-id: trunk@31569 -
 2015-09-07 19:05:18 +00:00
nickysn
98ec7a2211 * eliminated the dual initialization of DS in interrupt procedures in the huge 
memory model

git-svn-id: trunk@31519 -
 2015-09-04 21:37:07 +00:00
nickysn
0fdc62e0f7 + initialize DS with the current unit's data segment in the function entry code 
generated in the huge memory model

git-svn-id: trunk@31500 -
 2015-09-03 21:44:16 +00:00
Jonas Maebe
67b8aceaee * synchronized with privatetrunk till r30095 
git-svn-id: branches/hlcgllvm@30101 -
 2015-03-05 20:32:15 +00:00
Jonas Maebe
2ab7f5c35d * moved x86-specific requirements from the generic bsr/bsf code to the 
x86 code generator (register size constraints)

git-svn-id: trunk@29984 -
 2015-02-23 22:57:18 +00:00
Jonas Maebe
fa328436ab * regenerated for iphonesim/x86-64 
git-svn-id: trunk@29974 -
 2015-02-23 22:56:41 +00:00
Jonas Maebe
79a06b1514 + iphonesim/x86_64 target (64 bit iOS simulator) 
git-svn-id: trunk@29970 -
 2015-02-23 22:56:09 +00:00
nickysn
2d3ab816b0 * Tcgx86.g_concatcopy use: copy_move (mov sequence) instead of copy_string (rep 
movsb) when copying a 4 byte block on i8086, when optimizing for size. 4 movs
  are usually shorter than the instructions needed to setup rep movsb and also
  greatly reduce register pressure. This fixes i8086 snapshot building with -Os,
  Mantis #27387

git-svn-id: trunk@29639 -
 2015-02-07 13:19:34 +00:00
florian
35a04ad095 * modified patch of Paul W to replace leave by mov/pop, resolves #26455 
git-svn-id: trunk@29633 -
 2015-02-05 20:51:12 +00:00
Jonas Maebe
b745dcc64c * moved g_external_wrapper() to the hlcg, and also g_intf_wrapper() because 
for some platforms it depends on that routine

git-svn-id: branches/hlcgllvm@28492 -
 2014-08-19 20:22:54 +00:00
sergei
ae627a4ba8 * tcgx86.a_op_const_reg_reg: optimize trivial cases before considering more complex ones. 
git-svn-id: trunk@27892 -
 2014-06-08 00:08:56 +00:00
nickysn
daf71e6d88 * fixed multiplication by 64-bit constants on x86_64, mantis #26230 
git-svn-id: trunk@27822 -
 2014-05-28 22:52:08 +00:00
nickysn
6e7dd647fd * Tcgx86.g_concatcopy: optimization for the case where source.segment is 
specified, but equal to DS in the current memory model

git-svn-id: trunk@27720 -
 2014-05-02 13:07:54 +00:00
nickysn
1eb24f9d98 + remove the segment prefix in make_simple_ref on i8086, in case it is the same 
as the default segment for the reference (SS for BP-relative, DS for all
  others)

git-svn-id: trunk@27687 -
 2014-04-29 23:11:10 +00:00
sergei
07e90aaa24 + Implemented IEEE 754-compliant checking for unordered results of floating-point compares on x86 targets. Mantis #9362. 
git-svn-id: trunk@27581 -
 2014-04-14 12:36:11 +00:00
florian
18233f2c0c * fix compilation of i8086 compiler 
git-svn-id: trunk@27566 -
 2014-04-13 19:28:59 +00:00
florian
d88d644925 + support for FMA intrinsic: if there is no hardware support, the compiler throws an error. 
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 -
 2014-04-13 19:21:54 +00:00
nickysn
afcb8c3f10 * initialize DS from CS in the interrupt procedure entry code in tiny model, 
because dgroup emits relocations, which results in a linker warning and broken
  code when the target format is .com file.

git-svn-id: trunk@27517 -
 2014-04-10 20:02:35 +00:00
nickysn
7a873f495d * use the stack segment for SP-relative addresses in Tcgx86.g_concatcopy in 
the memory models with SS<>DS (necessary, because make_simple_ref isn't called
  in the copy_string case)

git-svn-id: trunk@27479 -
 2014-04-05 12:02:04 +00:00
nickysn
ee20ecfcf2 * tcgx86.make_simple_ref: in i8086 memory models with DS<>SS, add an SS: segment 
override when converting SP-relative references to use a general purpose
  register

git-svn-id: trunk@27473 -
 2014-04-05 08:56:40 +00:00
nickysn
b36a98ce51 + Tcgx86.g_concatcopy: added support for memory models with DS<>SS 
git-svn-id: trunk@27461 -
 2014-04-04 01:13:02 +00:00
nickysn
e9e4cc8bcb * Tcgx86.g_concatcopy: use boolean variables saved_ds and saved_es to keep track 
whether these segment registers were saved on the stack or not

git-svn-id: trunk@27460 -
 2014-04-04 00:10:36 +00:00
nickysn
096bba90da * fixed Tcgx86.g_concatcopy not using the correct push instruction size when 
pushing the source segment

git-svn-id: trunk@27389 -
 2014-03-30 18:09:15 +00:00
pierre
044891fd94 Avoid range check error in a_op_const_reg_reg 
git-svn-id: trunk@27060 -
 2014-03-09 19:49:29 +00:00
nickysn
229ebf33a8 + support getting the address of far references on i8086 in 
tcgx86.a_loadaddr_ref_reg

git-svn-id: trunk@26932 -
 2014-03-02 22:21:16 +00:00
florian
37b34edbc6 * do not reuse registers when calculating addresses in pic mode 
git-svn-id: trunk@26743 -
 2014-02-09 21:12:07 +00:00
sergei
9c1f917e3a * a_call_ref functionality cannot be implemented efficiently at code generator level, because references need specific preparations at earlier points. Moved this support to tcgcallnode and its x86 descendants, and got rid of all ifdef's around. 
+ x86 targets now directly call procedure variables located in references.
- a_call_ref method removed from tcg and thlcgobj.

git-svn-id: trunk@26666 -
 2014-02-03 13:28:56 +00:00
sergei
0d3f36eebf - Remove references to global variable 'cg' from methods of tcg and some of its descendants. 
git-svn-id: trunk@26665 -
 2014-02-03 12:27:48 +00:00
florian
28a8dc42c6 * fix tcgx86.a_op_reg_reg on x86-64 and make use of it 
git-svn-id: trunk@26636 -
 2014-01-30 20:24:41 +00:00
florian
7d7bf1d877 + make use of SHLX/SHRX on CPUs supporting BMI2 
git-svn-id: trunk@26635 -
 2014-01-30 20:24:39 +00:00
florian
48ae2d215a + concatcopy variants using sse and avx, only activated if optimization for size is done because at least on an i7-4770 it has shown no benefit 
git-svn-id: trunk@26588 -
 2014-01-26 12:37:54 +00:00
nickysn
85dd9e5789 + added a size parameter to optimize_op_const and do a sign extension of the 'a' parameter up from the specified size, so that things like (i and $ffffffff) get optimized away the same way as (i and -1) 
git-svn-id: trunk@26561 -
 2014-01-22 15:00:34 +00:00
sergei
6b1f021fcf - Removed remaining optimizations that duplicate functionality of tcg.optimize_op_const. 
- 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 -
 2014-01-19 01:19:01 +00:00
nickysn
1f7b14f38e * fixed the conversion of powerof2 mul/imuls to lea in tcgx86.a_op_const_reg_reg 
git-svn-id: trunk@26509 -
 2014-01-18 17:58:31 +00:00
nickysn
cc1faad3d2 + make use of the 3-operand form of IMUL on i386 and x86_64 
git-svn-id: trunk@26508 -
 2014-01-18 17:04:57 +00:00
nickysn
5d75bf4f92 * always use IMUL instead of MUL on i386, when doing a 32->32 multiplication 
without overflow checking

git-svn-id: trunk@26498 -
 2014-01-18 02:56:51 +00:00
nickysn
820b667e1b - rm the OP_IMUL to SHL optimization from tcgx86.a_op_const_reg and tcgx86.a_op_const_ref, because it's already done in optimize_op_const, which is called in the beginning of both methods. 
git-svn-id: trunk@26491 -
 2014-01-17 18:34:13 +00:00
florian
e210d5f30e + cpu_capabilites for x86_64 and i386 
* take advantage of bmi2 instruction rorx

git-svn-id: trunk@26482 -
 2014-01-16 21:47:28 +00:00
sergei
f2096de53a * tcg386.g_proc_exit: instead of recalculating stack size, use current_procinfo.final_localsize which was calculated in g_proc_entry. 
* tcgx86.g_proc_entry: don't over-allocate stack in SEH finalizer procedures.

git-svn-id: trunk@26251 -
 2013-12-19 10:32:32 +00:00
sergei
179586f589 + SEH support for Win32. Enable by cycling with OPT=-dTEST_WIN32_SEH. 
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 -
 2013-12-12 12:43:46 +00:00
florian
4d5119bf1c * fixes several issues which cause warnings by the dfa code when using it to detect uninitialized variables 
git-svn-id: trunk@26161 -
 2013-12-01 17:02:08 +00:00
nickysn
7ae59a62c5 * emit a FWAIT after a floating point store, when targeting the 8087 and 80287 
git-svn-id: trunk@26109 -
 2013-11-19 23:11:26 +00:00
nickysn
dea947f87b * fixed tcgx86.a_op_const_ref for shl/shr/sar/rol/ror on i8086 and x86_64 
git-svn-id: trunk@26018 -
 2013-11-10 20:39:47 +00:00
nickysn
d0f988f7a8 * don't use LEA in the code generator on i8086, as this breaks it quite badly. 
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 -
 2013-11-07 17:00:58 +00:00
florian
7fc1d70e49 * do reg2=reg1*const by lea if possible 
git-svn-id: trunk@25908 -
 2013-11-01 19:01:36 +00:00
florian
3c58cad96b * make use of lea if possible 
git-svn-id: trunk@25898 -
 2013-11-01 19:01:11 +00:00
nickysn
67fdcb02f2 * in tcgx86.make_simple_ref, on the i8086, emit 'mov es, reg', instead of 
'push reg/pop es', since that is a lot faster on 8088..80386

git-svn-id: trunk@25700 -
 2013-10-06 19:52:38 +00:00
nickysn
6be9627e74 * use all kinds of copy and not just copy_string in Tcgx86.g_concatcopy, when 
segments are specified on the i8086

git-svn-id: trunk@25698 -
 2013-10-06 19:04:13 +00:00
nickysn
bd0585274e * tsettings.enablecld converted to a targetswitch ts_cld 
git-svn-id: trunk@25592 -
 2013-09-28 13:41:58 +00:00
nickysn
52fcc0a407 + added a new x86-specific compiler option 'enablecld', which controls whether 
the compiler should emit a CLD before using the x86 string instructions.

git-svn-id: trunk@25590 -
 2013-09-28 11:54:02 +00:00
Jonas Maebe
3b9b6a0cd7 * explicitly define the asmsymbols for Darwin call stubs as local 
git-svn-id: trunk@25264 -
 2013-08-16 10:07:46 +00:00
sergei
e1550e1aed * Restore registers in correct order (for push/pop it *does* matter), fixes crashes caused by r25224. 
git-svn-id: trunk@25228 -
 2013-08-08 05:35:51 +00:00
sergei
5c4abfa75a + Implemented saving/restoring registers using PUSH/POP instructions on i386 targets with non-fixed stack. Provides some advantage in code size (e.g. the compiler becomes about 20KBytes smaller). 
* 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 -
 2013-08-07 09:59:07 +00:00
sergei
a21a20d559 + i8086-specific version of g_stackpointer_alloc. 
git-svn-id: trunk@25219 -
 2013-08-06 08:11:39 +00:00
sergei
1f72143eef * In preparation for fixing Mantis #24791: Copy related code from tcgx86.g_proc_entry to tcgx86_64.g_proc_entry, and remove win64-specific stuff from tcgx86 version. 
git-svn-id: trunk@25218 -
 2013-08-06 07:34:50 +00:00
florian
507edb16de * use add/sub instead of inc/dec on modern CPUs when optimizing for speed 
git-svn-id: trunk@25057 -
 2013-07-07 19:50:13 +00:00
nickysn
7e70a5f763 * i8086 fix after r25010: i8086 doesn't support lea sp,[sp-XXX] 
git-svn-id: trunk@25012 -
 2013-06-29 12:21:15 +00:00
florian
94cf650d9a * use lea to adjust stack pointer, this is equal or faster on all modern CPUs than add/sub 
git-svn-id: trunk@25010 -
 2013-06-28 17:06:57 +00:00
florian
466eb4c684 * use MOVAPS/MOVAPD for reg -> reg moves to reduce data dependencies 
git-svn-id: trunk@24967 -
 2013-06-24 20:55:04 +00:00
florian
e81d2d1f3b * basic avx support for floating point operations (use -Cfavx to activate) 
git-svn-id: trunk@24896 -
 2013-06-14 20:03:01 +00:00
nickysn
62316da236 * a_call_reg moved to a_call_reg_near 
git-svn-id: trunk@24855 -
 2013-06-09 22:09:39 +00:00
nickysn
604b7c9deb + added cg.a_call_ref_near and a_call_ref_far 
git-svn-id: trunk@24853 -
 2013-06-09 20:22:47 +00:00
nickysn
ec497ae11e + the regular tcgx86.a_call_name and .a_call_name_static moved to .a_call_name_near and .a_call_name_static_near, because a_call_name and a_call_name_static are going to be overriden in tcg8086 
git-svn-id: trunk@24827 -
 2013-06-09 10:42:39 +00:00
nickysn
ec76b2cf40 * refactored the segment-in-integer-register reference support, so it's handled in make_simple_ref, instead of all over the place in the code generator 
git-svn-id: trunk@24752 -
 2013-06-01 22:18:09 +00:00
nickysn
c9c882c30a * always use 16-bit pushes for pushing/popping the segment registers on i8086 in Tcgx86.g_concatcopy 
git-svn-id: trunk@24735 -
 2013-06-01 17:49:49 +00:00
nickysn
e243b6b869 * never use the 66h prefix when pushing/popping segment registers on i8086, as the stack is aligned on 2 bytes there 
git-svn-id: trunk@24729 -
 2013-06-01 16:03:14 +00:00
nickysn
c2e3fb5918 + emit proper interrupt procedure entry/exit code on i8086 
git-svn-id: trunk@24728 -
 2013-06-01 15:50:11 +00:00
nickysn
0a8e008a0d + the ES register made volatile on i8086 as per the 16-bit x86 calling conventions 
git-svn-id: trunk@24462 -
 2013-05-07 14:27:21 +00:00
nickysn
e4f3dd1264 * do not emit shl/shr with a const different than 1 on 8086 
git-svn-id: trunk@24382 -
 2013-04-29 23:02:09 +00:00
nickysn
fbc41991f9 * i8086 fixes in tcgx86.a_op_reg_reg for 16-bit OP_SHR,OP_SHL,OP_SAR,OP_ROL,OP_ROR 
git-svn-id: branches/i8086@23881 -
 2013-03-17 00:05:05 +00:00
florian
edd4c1ce4b * proper fix to avoid sp relative adressing as r23867 tries 
git-svn-id: branches/i8086@23868 -
 2013-03-16 16:51:31 +00:00
florian
b321f867c6 + trgintcpu.add_cpu_interferences for i8086 so references are properly build 
* empty make_simple_ref code path for i8086

git-svn-id: branches/i8086@23782 -
 2013-03-10 18:51:42 +00:00
nickysn
95b15d15bd another nested ifdef cleaned up 
git-svn-id: branches/i8086@23741 -
 2013-03-09 11:35:42 +00:00
nickysn
981f0a5c6c nested ifdefs converted to series of elseif + some other ifdefs cleaned up 
git-svn-id: branches/i8086@23740 -
 2013-03-09 11:25:25 +00:00
nickysn
5f7d432ff9 i8086 fixes in Tcgx86.g_concatcopy and TCGSize2OpSize 
git-svn-id: branches/i8086@23720 -
 2013-03-08 00:13:02 +00:00
Jonas Maebe
5d628b29bb * set the subregsize of OS_M64 SSE registers to R_SUBQ so we can 
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 -
 2013-03-06 12:42:46 +00:00
yury
e13e2e1697 * Turn off special handling of interrupt calling convention for i386-android. 
git-svn-id: branches/targetandroid@23492 -
 2013-01-22 14:43:42 +00:00
yury
0960ee2034 * Add missing checks for android target. 
git-svn-id: branches/targetandroid@23477 -
 2013-01-21 12:08:25 +00:00
sergei
7530fb3352 * tcgx86.make_simple_ref: handle references to absolute addresses (these are quite rare on x86_64, but still worth to handle correctly). 
* 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 -
 2013-01-07 16:00:12 +00:00
florian
d7e0f07aab + UseAVX check 
git-svn-id: trunk@22642 -
 2012-10-14 14:04:31 +00:00
florian
283ff05127 * merged avx support in inline assembler developed by Torsten Grundke 
git-svn-id: trunk@22568 -
 2012-10-06 19:47:18 +00:00
florian
8818b58e5d * Bsf/Bsr on x86 handle now correctly 0 as argument, resolves #22783 
* generic Bsf implementations handle now correctly 0 as argument
* test extended

git-svn-id: trunk@22327 -
 2012-09-05 13:51:45 +00:00
Jonas Maebe
b1dc518ac4 * removed systems_need_16_byte_stack_alignment and use target_info.stackalign instead 
git-svn-id: trunk@22279 -
 2012-09-02 14:32:21 +00:00
Jonas Maebe
952ac1ffa6 * only use indirect symbol loads for (weak)external and private_external 
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 -
 2012-08-30 21:30:38 +00:00