cg386mat.pas 45 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052
  1. {
  2. $Id$
  3. Copyright (c) 1993-98 by Florian Klaempfl
  4. Generate i386 assembler for math nodes
  5. This program is free software; you can redistribute it and/or modify
  6. it under the terms of the GNU General Public License as published by
  7. the Free Software Foundation; either version 2 of the License, or
  8. (at your option) any later version.
  9. This program is distributed in the hope that it will be useful,
  10. but WITHOUT ANY WARRANTY; without even the implied warranty of
  11. MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
  12. GNU General Public License for more details.
  13. You should have received a copy of the GNU General Public License
  14. along with this program; if not, write to the Free Software
  15. Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  16. ****************************************************************************
  17. }
  18. unit cg386mat;
  19. interface
  20. uses
  21. tree;
  22. procedure secondmoddiv(var p : ptree);
  23. procedure secondshlshr(var p : ptree);
  24. procedure secondumminus(var p : ptree);
  25. procedure secondnot(var p : ptree);
  26. implementation
  27. uses
  28. globtype,systems,
  29. cobjects,verbose,globals,
  30. symconst,symtable,aasm,types,
  31. hcodegen,temp_gen,pass_2,
  32. cpubase,cpuasm,
  33. {$ifdef dummy}
  34. end { this overcomes the annoying highlighting problem in my TP IDE,
  35. the IDE assumes i386asm start a asm block (FK) }
  36. {$endif}
  37. cgai386,tgeni386;
  38. {*****************************************************************************
  39. SecondModDiv
  40. *****************************************************************************}
  41. procedure secondmoddiv(var p : ptree);
  42. var
  43. hreg1 : tregister;
  44. shrdiv, andmod, pushed,popeax,popedx : boolean;
  45. power : longint;
  46. hl : pasmlabel;
  47. hloc : tlocation;
  48. pushedreg : tpushed;
  49. typename,opname : string[6];
  50. begin
  51. shrdiv := false;
  52. andmod := false;
  53. secondpass(p^.left);
  54. set_location(p^.location,p^.left^.location);
  55. pushed:=maybe_push(p^.right^.registers32,p,is_64bitint(p^.left^.resulttype));
  56. secondpass(p^.right);
  57. if pushed then restore(p,is_64bitint(p^.left^.resulttype));
  58. if is_64bitint(p^.resulttype) then
  59. begin
  60. { save p^.lcoation, because we change it now }
  61. set_location(hloc,p^.location);
  62. release_qword_loc(p^.location);
  63. release_qword_loc(p^.right^.location);
  64. p^.location.registerlow:=getexplicitregister32(R_EAX);
  65. p^.location.registerhigh:=getexplicitregister32(R_EDX);
  66. pushusedregisters(pushedreg,$ff
  67. and not($80 shr byte(p^.location.registerlow))
  68. and not($80 shr byte(p^.location.registerhigh)));
  69. { the left operand is in hloc, because the
  70. location of left is p^.location but p^.location
  71. is already destroyed
  72. }
  73. emit_pushq_loc(hloc);
  74. clear_location(hloc);
  75. emit_pushq_loc(p^.right^.location);
  76. if porddef(p^.resulttype)^.typ=u64bit then
  77. typename:='QWORD'
  78. else
  79. typename:='INT64';
  80. if p^.treetype=divn then
  81. opname:='DIV_'
  82. else
  83. opname:='MOD_';
  84. emitcall('FPC_'+opname+typename);
  85. emit_reg_reg(A_MOV,S_L,R_EAX,p^.location.registerlow);
  86. emit_reg_reg(A_MOV,S_L,R_EDX,p^.location.registerhigh);
  87. popusedregisters(pushedreg);
  88. p^.location.loc:=LOC_REGISTER;
  89. end
  90. else
  91. begin
  92. { put numerator in register }
  93. if p^.left^.location.loc<>LOC_REGISTER then
  94. begin
  95. if p^.left^.location.loc=LOC_CREGISTER then
  96. begin
  97. hreg1:=getregister32;
  98. emit_reg_reg(A_MOV,S_L,p^.left^.location.register,hreg1);
  99. end
  100. else
  101. begin
  102. del_reference(p^.left^.location.reference);
  103. hreg1:=getregister32;
  104. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.left^.location.reference),
  105. hreg1)));
  106. end;
  107. clear_location(p^.left^.location);
  108. p^.left^.location.loc:=LOC_REGISTER;
  109. p^.left^.location.register:=hreg1;
  110. end
  111. else hreg1:=p^.left^.location.register;
  112. if (p^.treetype=divn) and (p^.right^.treetype=ordconstn) and
  113. ispowerof2(p^.right^.value,power) then
  114. Begin
  115. shrdiv := true;
  116. {for signed numbers, the numerator must be adjusted before the
  117. shift instruction, but not wih unsigned numbers! Otherwise,
  118. "Cardinal($ffffffff) div 16" overflows! (JM)}
  119. If is_signed(p^.left^.resulttype) Then
  120. Begin
  121. exprasmlist^.concat(new(pai386,op_reg_reg(A_OR,S_L,hreg1,hreg1)));
  122. getlabel(hl);
  123. emitjmp(C_NS,hl);
  124. if power=1 then
  125. exprasmlist^.concat(new(pai386,op_reg(A_INC,S_L,hreg1)))
  126. else
  127. exprasmlist^.concat(new(pai386,op_const_reg(A_ADD,S_L,p^.right^.value-1,hreg1)));
  128. emitlab(hl);
  129. exprasmlist^.concat(new(pai386,op_const_reg(A_SAR,S_L,power,hreg1)));
  130. End
  131. Else
  132. exprasmlist^.concat(new(pai386,op_const_reg(A_SHR,S_L,power,hreg1)));
  133. End
  134. else
  135. if (p^.treetype=modn) and (p^.right^.treetype=ordconstn) and
  136. ispowerof2(p^.right^.value,power) and Not(is_signed(p^.left^.resulttype)) Then
  137. {is there a similar trick for MOD'ing signed numbers? (JM)}
  138. Begin
  139. exprasmlist^.concat(new(pai386,op_const_reg(A_AND,S_L,p^.right^.value-1,hreg1)));
  140. andmod := true;
  141. End
  142. else
  143. begin
  144. { bring denominator to EDI }
  145. { EDI is always free, it's }
  146. { only used for temporary }
  147. { purposes }
  148. if (p^.right^.location.loc<>LOC_REGISTER) and
  149. (p^.right^.location.loc<>LOC_CREGISTER) then
  150. begin
  151. del_reference(p^.right^.location.reference);
  152. p^.left^.location.loc:=LOC_REGISTER;
  153. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.right^.location.reference),R_EDI)));
  154. end
  155. else
  156. begin
  157. emit_reg_reg(A_MOV,S_L,p^.right^.location.register,R_EDI);
  158. ungetregister32(p^.right^.location.register);
  159. end;
  160. popedx:=false;
  161. popeax:=false;
  162. if hreg1=R_EDX then
  163. begin
  164. if not(R_EAX in unused) then
  165. begin
  166. exprasmlist^.concat(new(pai386,op_reg(A_PUSH,S_L,R_EAX)));
  167. popeax:=true;
  168. end;
  169. emit_reg_reg(A_MOV,S_L,R_EDX,R_EAX);
  170. end
  171. else
  172. begin
  173. if not(R_EDX in unused) then
  174. begin
  175. exprasmlist^.concat(new(pai386,op_reg(A_PUSH,S_L,R_EDX)));
  176. popedx:=true;
  177. end;
  178. if hreg1<>R_EAX then
  179. begin
  180. if not(R_EAX in unused) then
  181. begin
  182. exprasmlist^.concat(new(pai386,op_reg(A_PUSH,S_L,R_EAX)));
  183. popeax:=true;
  184. end;
  185. emit_reg_reg(A_MOV,S_L,hreg1,R_EAX);
  186. end;
  187. end;
  188. { sign extension depends on the left type }
  189. if porddef(p^.left^.resulttype)^.typ=u32bit then
  190. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,R_EDX,R_EDX)))
  191. else
  192. exprasmlist^.concat(new(pai386,op_none(A_CDQ,S_NO)));
  193. { division depends on the right type }
  194. if porddef(p^.right^.resulttype)^.typ=u32bit then
  195. exprasmlist^.concat(new(pai386,op_reg(A_DIV,S_L,R_EDI)))
  196. else
  197. exprasmlist^.concat(new(pai386,op_reg(A_IDIV,S_L,R_EDI)));
  198. if p^.treetype=divn then
  199. begin
  200. { if result register is busy then copy }
  201. if popeax then
  202. begin
  203. if hreg1=R_EAX then
  204. internalerror(112);
  205. emit_reg_reg(A_MOV,S_L,R_EAX,hreg1)
  206. end
  207. else
  208. if hreg1<>R_EAX then
  209. Begin
  210. ungetregister32(hreg1);
  211. hreg1 := getexplicitregister32(R_EAX);
  212. { I don't think it's possible that now hreg1 <> R_EAX
  213. since popeax is false, but for all certainty I do
  214. support that situation (JM)}
  215. if hreg1 <> R_EAX then
  216. emit_reg_reg(A_MOV,S_L,R_EAX,hreg1);
  217. end;
  218. end
  219. else
  220. {if we did the mod by an "and", the result is in hreg1 and
  221. EDX certainly hasn't been pushed (JM)}
  222. if not(andmod) Then
  223. if popedx then
  224. {the mod was done by an (i)div (so the result is now in
  225. edx), but edx was occupied prior to the division, so
  226. move the result into a safe place (JM)}
  227. emit_reg_reg(A_MOV,S_L,R_EDX,hreg1)
  228. else
  229. Begin
  230. {Get rid of the unnecessary hreg1 if possible (same as with
  231. EAX in divn) (JM)}
  232. ungetregister32(hreg1);
  233. hreg1 := getexplicitregister32(R_EDX);
  234. if hreg1 <> R_EDX then
  235. emit_reg_reg(A_MOV,S_L,R_EDX,hreg1);;
  236. End;
  237. if popeax then
  238. exprasmlist^.concat(new(pai386,op_reg(A_POP,S_L,R_EAX)));
  239. if popedx then
  240. exprasmlist^.concat(new(pai386,op_reg(A_POP,S_L,R_EDX)));
  241. end;
  242. If not(andmod or shrdiv) then
  243. {andmod and shrdiv only use hreg1 (which is already in usedinproc,
  244. since it was acquired with getregister), the others also use both
  245. EAX and EDX (JM)}
  246. Begin
  247. usedinproc:=usedinproc or ($80 shr byte(R_EAX));
  248. usedinproc:=usedinproc or ($80 shr byte(R_EDX));
  249. End;
  250. clear_location(p^.location);
  251. p^.location.loc:=LOC_REGISTER;
  252. p^.location.register:=hreg1;
  253. end;
  254. end;
  255. {*****************************************************************************
  256. SecondShlShr
  257. *****************************************************************************}
  258. procedure secondshlshr(var p : ptree);
  259. var
  260. hregister1,hregister2,hregister3,
  261. hregisterhigh,hregisterlow : tregister;
  262. pushed,popecx : boolean;
  263. op : tasmop;
  264. hr : preference;
  265. l1,l2,l3 : pasmlabel;
  266. begin
  267. popecx:=false;
  268. secondpass(p^.left);
  269. pushed:=maybe_push(p^.right^.registers32,p,is_64bitint(p^.left^.resulttype));
  270. secondpass(p^.right);
  271. if pushed then
  272. restore(p,is_64bitint(p^.left^.resulttype));
  273. if is_64bitint(p^.left^.resulttype) then
  274. begin
  275. { load left operator in a register }
  276. if p^.left^.location.loc<>LOC_REGISTER then
  277. begin
  278. if p^.left^.location.loc=LOC_CREGISTER then
  279. begin
  280. hregisterlow:=getregister32;
  281. hregisterhigh:=getregister32;
  282. emit_reg_reg(A_MOV,S_L,p^.left^.location.registerlow,
  283. hregisterlow);
  284. emit_reg_reg(A_MOV,S_L,p^.left^.location.registerhigh,
  285. hregisterlow);
  286. end
  287. else
  288. begin
  289. del_reference(p^.left^.location.reference);
  290. hregisterlow:=getregister32;
  291. hregisterhigh:=getregister32;
  292. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.left^.location.reference),
  293. hregisterlow)));
  294. hr:=newreference(p^.left^.location.reference);
  295. inc(hr^.offset,4);
  296. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,hr,
  297. hregisterhigh)));
  298. end;
  299. end
  300. else
  301. begin
  302. hregisterlow:=p^.left^.location.registerlow;
  303. hregisterhigh:=p^.left^.location.registerhigh;
  304. end;
  305. { shifting by a constant directly coded: }
  306. if (p^.right^.treetype=ordconstn) then
  307. begin
  308. { shrd/shl works only for values <=31 !! }
  309. if p^.right^.value>31 then
  310. begin
  311. if p^.treetype=shln then
  312. begin
  313. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,hregisterhigh,
  314. hregisterhigh)));
  315. exprasmlist^.concat(new(pai386,op_const_reg(A_SHL,S_L,p^.right^.value and 31,
  316. hregisterlow)));
  317. end
  318. else
  319. begin
  320. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,hregisterlow,
  321. hregisterlow)));
  322. exprasmlist^.concat(new(pai386,op_const_reg(A_SHR,S_L,p^.right^.value and 31,
  323. hregisterhigh)));
  324. end;
  325. p^.location.registerhigh:=hregisterlow;
  326. p^.location.registerlow:=hregisterhigh;
  327. end
  328. else
  329. begin
  330. if p^.treetype=shln then
  331. begin
  332. exprasmlist^.concat(new(pai386,op_const_reg_reg(A_SHLD,S_L,p^.right^.value and 31,
  333. hregisterlow,hregisterhigh)));
  334. exprasmlist^.concat(new(pai386,op_const_reg(A_SHL,S_L,p^.right^.value and 31,
  335. hregisterlow)));
  336. end
  337. else
  338. begin
  339. exprasmlist^.concat(new(pai386,op_const_reg_reg(A_SHRD,S_L,p^.right^.value and 31,
  340. hregisterhigh,hregisterlow)));
  341. exprasmlist^.concat(new(pai386,op_const_reg(A_SHR,S_L,p^.right^.value and 31,
  342. hregisterhigh)));
  343. end;
  344. p^.location.registerlow:=hregisterlow;
  345. p^.location.registerhigh:=hregisterhigh;
  346. end;
  347. p^.location.loc:=LOC_REGISTER;
  348. end
  349. else
  350. begin
  351. { load right operators in a register }
  352. if p^.right^.location.loc<>LOC_REGISTER then
  353. begin
  354. if p^.right^.location.loc=LOC_CREGISTER then
  355. begin
  356. hregister2:=getexplicitregister32(R_ECX);
  357. emit_reg_reg(A_MOV,S_L,p^.right^.location.register,
  358. hregister2);
  359. end
  360. else
  361. begin
  362. del_reference(p^.right^.location.reference);
  363. hregister2:=getexplicitregister32(R_ECX);
  364. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.right^.location.reference),
  365. hregister2)));
  366. end;
  367. end
  368. else
  369. hregister2:=p^.right^.location.register;
  370. { left operator is already in a register }
  371. { hence are both in a register }
  372. { is it in the case ECX ? }
  373. if (hregisterlow=R_ECX) then
  374. begin
  375. { then only swap }
  376. emit_reg_reg(A_XCHG,S_L,hregisterlow,hregister2);
  377. hregister3:=hregisterlow;
  378. hregisterlow:=hregister2;
  379. hregister2:=hregister3;
  380. end
  381. else if (hregisterhigh=R_ECX) then
  382. begin
  383. { then only swap }
  384. emit_reg_reg(A_XCHG,S_L,hregisterhigh,hregister2);
  385. hregister3:=hregisterhigh;
  386. hregisterhigh:=hregister2;
  387. hregister2:=hregister3;
  388. end
  389. { if second operator not in ECX ? }
  390. else if (hregister2<>R_ECX) then
  391. begin
  392. { ECX occupied then push it }
  393. if not (R_ECX in unused) then
  394. begin
  395. popecx:=true;
  396. exprasmlist^.concat(new(pai386,op_reg(A_PUSH,S_L,R_ECX)));
  397. end;
  398. emit_reg_reg(A_MOV,S_L,hregister2,R_ECX);
  399. end;
  400. ungetregister32(hregister2);
  401. { the damned shift instructions work only til a count of 32 }
  402. { so we've to do some tricks here }
  403. if p^.treetype=shln then
  404. begin
  405. getlabel(l1);
  406. getlabel(l2);
  407. getlabel(l3);
  408. exprasmlist^.concat(new(pai386,op_const_reg(A_CMP,S_L,64,R_ECX)));
  409. emitjmp(C_L,l1);
  410. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,hregisterlow,hregisterlow)));
  411. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,hregisterhigh,hregisterhigh)));
  412. emitjmp(C_None,l3);
  413. emitlab(l1);
  414. exprasmlist^.concat(new(pai386,op_const_reg(A_CMP,S_L,32,R_ECX)));
  415. emitjmp(C_L,l2);
  416. exprasmlist^.concat(new(pai386,op_const_reg(A_SUB,S_L,32,R_ECX)));
  417. exprasmlist^.concat(new(pai386,op_reg_reg(A_SHL,S_L,R_CL,
  418. hregisterlow)));
  419. exprasmlist^.concat(new(pai386,op_reg_reg(A_MOV,S_L,hregisterlow,hregisterhigh)));
  420. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,hregisterlow,hregisterlow)));
  421. emitjmp(C_None,l3);
  422. emitlab(l2);
  423. exprasmlist^.concat(new(pai386,op_reg_reg_reg(A_SHLD,S_L,R_CL,
  424. hregisterlow,hregisterhigh)));
  425. exprasmlist^.concat(new(pai386,op_reg_reg(A_SHL,S_L,R_CL,
  426. hregisterlow)));
  427. emitlab(l3);
  428. end
  429. else
  430. begin
  431. getlabel(l1);
  432. getlabel(l2);
  433. getlabel(l3);
  434. exprasmlist^.concat(new(pai386,op_const_reg(A_CMP,S_L,64,R_ECX)));
  435. emitjmp(C_L,l1);
  436. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,hregisterlow,hregisterlow)));
  437. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,hregisterhigh,hregisterhigh)));
  438. emitjmp(C_None,l3);
  439. emitlab(l1);
  440. exprasmlist^.concat(new(pai386,op_const_reg(A_CMP,S_L,32,R_ECX)));
  441. emitjmp(C_L,l2);
  442. exprasmlist^.concat(new(pai386,op_const_reg(A_SUB,S_L,32,R_ECX)));
  443. exprasmlist^.concat(new(pai386,op_reg_reg(A_SHR,S_L,R_CL,
  444. hregisterhigh)));
  445. exprasmlist^.concat(new(pai386,op_reg_reg(A_MOV,S_L,hregisterhigh,hregisterlow)));
  446. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,hregisterhigh,hregisterhigh)));
  447. emitjmp(C_None,l3);
  448. emitlab(l2);
  449. exprasmlist^.concat(new(pai386,op_reg_reg_reg(A_SHRD,S_L,R_CL,
  450. hregisterhigh,hregisterlow)));
  451. exprasmlist^.concat(new(pai386,op_reg_reg(A_SHR,S_L,R_CL,
  452. hregisterhigh)));
  453. emitlab(l3);
  454. end;
  455. { maybe put ECX back }
  456. if popecx then
  457. exprasmlist^.concat(new(pai386,op_reg(A_POP,S_L,R_ECX)));
  458. p^.location.registerlow:=hregisterlow;
  459. p^.location.registerhigh:=hregisterhigh;
  460. end;
  461. end
  462. else
  463. begin
  464. { load left operators in a register }
  465. if p^.left^.location.loc<>LOC_REGISTER then
  466. begin
  467. if p^.left^.location.loc=LOC_CREGISTER then
  468. begin
  469. hregister1:=getregister32;
  470. emit_reg_reg(A_MOV,S_L,p^.left^.location.register,
  471. hregister1);
  472. end
  473. else
  474. begin
  475. del_reference(p^.left^.location.reference);
  476. hregister1:=getregister32;
  477. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.left^.location.reference),
  478. hregister1)));
  479. end;
  480. end
  481. else
  482. hregister1:=p^.left^.location.register;
  483. { determine operator }
  484. if p^.treetype=shln then
  485. op:=A_SHL
  486. else
  487. op:=A_SHR;
  488. { shifting by a constant directly coded: }
  489. if (p^.right^.treetype=ordconstn) then
  490. begin
  491. { l shl 32 should 0 imho, but neither TP nor Delphi do it in this way (FK)
  492. if p^.right^.value<=31 then
  493. }
  494. exprasmlist^.concat(new(pai386,op_const_reg(op,S_L,p^.right^.value and 31,
  495. hregister1)));
  496. {
  497. else
  498. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,hregister1,
  499. hregister1)));
  500. }
  501. p^.location.loc:=LOC_REGISTER;
  502. p^.location.register:=hregister1;
  503. end
  504. else
  505. begin
  506. { load right operators in a register }
  507. if p^.right^.location.loc<>LOC_REGISTER then
  508. begin
  509. if p^.right^.location.loc=LOC_CREGISTER then
  510. begin
  511. hregister2:=getexplicitregister32(R_ECX);
  512. emit_reg_reg(A_MOV,S_L,p^.right^.location.register,
  513. hregister2);
  514. end
  515. else
  516. begin
  517. del_reference(p^.right^.location.reference);
  518. hregister2:=getexplicitregister32(R_ECX);
  519. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.right^.location.reference),
  520. hregister2)));
  521. end;
  522. end
  523. else
  524. hregister2:=p^.right^.location.register;
  525. { left operator is already in a register }
  526. { hence are both in a register }
  527. { is it in the case ECX ? }
  528. if (hregister1=R_ECX) then
  529. begin
  530. { then only swap }
  531. emit_reg_reg(A_XCHG,S_L,hregister1,hregister2);
  532. hregister3:=hregister1;
  533. hregister1:=hregister2;
  534. hregister2:=hregister3;
  535. end
  536. { if second operator not in ECX ? }
  537. else if (hregister2<>R_ECX) then
  538. begin
  539. { ECX occupied then push it }
  540. if not (R_ECX in unused) then
  541. begin
  542. popecx:=true;
  543. exprasmlist^.concat(new(pai386,op_reg(A_PUSH,S_L,R_ECX)));
  544. end;
  545. emit_reg_reg(A_MOV,S_L,hregister2,R_ECX);
  546. end;
  547. ungetregister32(hregister2);
  548. { right operand is in ECX }
  549. emit_reg_reg(op,S_L,R_CL,hregister1);
  550. { maybe ECX back }
  551. if popecx then
  552. exprasmlist^.concat(new(pai386,op_reg(A_POP,S_L,R_ECX)));
  553. p^.location.register:=hregister1;
  554. end;
  555. end;
  556. end;
  557. {*****************************************************************************
  558. SecondUmMinus
  559. *****************************************************************************}
  560. procedure secondumminus(var p : ptree);
  561. {$ifdef SUPPORT_MMX}
  562. procedure do_mmx_neg;
  563. var
  564. op : tasmop;
  565. begin
  566. p^.location.loc:=LOC_MMXREGISTER;
  567. if cs_mmx_saturation in aktlocalswitches then
  568. case mmx_type(p^.resulttype) of
  569. mmxs8bit:
  570. op:=A_PSUBSB;
  571. mmxu8bit:
  572. op:=A_PSUBUSB;
  573. mmxs16bit,mmxfixed16:
  574. op:=A_PSUBSW;
  575. mmxu16bit:
  576. op:=A_PSUBUSW;
  577. end
  578. else
  579. case mmx_type(p^.resulttype) of
  580. mmxs8bit,mmxu8bit:
  581. op:=A_PSUBB;
  582. mmxs16bit,mmxu16bit,mmxfixed16:
  583. op:=A_PSUBW;
  584. mmxs32bit,mmxu32bit:
  585. op:=A_PSUBD;
  586. end;
  587. emit_reg_reg(op,S_NO,p^.location.register,R_MM7);
  588. emit_reg_reg(A_MOVQ,S_NO,R_MM7,p^.location.register);
  589. end;
  590. {$endif}
  591. var
  592. hr : preference;
  593. begin
  594. if is_64bitint(p^.left^.resulttype) then
  595. begin
  596. secondpass(p^.left);
  597. clear_location(p^.location);
  598. p^.location.loc:=LOC_REGISTER;
  599. case p^.left^.location.loc of
  600. LOC_REGISTER :
  601. begin
  602. p^.location.registerlow:=p^.left^.location.registerlow;
  603. p^.location.registerhigh:=p^.left^.location.registerhigh;
  604. end;
  605. LOC_CREGISTER :
  606. begin
  607. p^.location.registerlow:=getregister32;
  608. p^.location.registerhigh:=getregister32;
  609. emit_reg_reg(A_MOV,S_L,p^.left^.location.registerlow,p^.location.registerlow);
  610. emit_reg_reg(A_MOV,S_L,p^.left^.location.registerhigh,p^.location.registerhigh);
  611. end;
  612. LOC_REFERENCE,LOC_MEM :
  613. begin
  614. del_reference(p^.left^.location.reference);
  615. p^.location.registerlow:=getregister32;
  616. p^.location.registerhigh:=getregister32;
  617. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,
  618. newreference(p^.left^.location.reference),p^.location.registerlow)));
  619. hr:=newreference(p^.left^.location.reference);
  620. inc(hr^.offset,4);
  621. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,
  622. hr,p^.location.registerhigh)));
  623. end;
  624. end;
  625. exprasmlist^.concat(new(pai386,op_reg(A_NEG,S_L,p^.location.registerlow)));
  626. exprasmlist^.concat(new(pai386,op_const_reg(A_ADC,S_L,0,p^.location.registerhigh)));
  627. exprasmlist^.concat(new(pai386,op_reg(A_NEG,S_L,p^.location.registerhigh)));
  628. end
  629. else
  630. begin
  631. secondpass(p^.left);
  632. p^.location.loc:=LOC_REGISTER;
  633. case p^.left^.location.loc of
  634. LOC_REGISTER:
  635. begin
  636. p^.location.register:=p^.left^.location.register;
  637. exprasmlist^.concat(new(pai386,op_reg(A_NEG,S_L,p^.location.register)));
  638. end;
  639. LOC_CREGISTER:
  640. begin
  641. p^.location.register:=getregister32;
  642. emit_reg_reg(A_MOV,S_L,p^.location.register,
  643. p^.location.register);
  644. exprasmlist^.concat(new(pai386,op_reg(A_NEG,S_L,p^.location.register)));
  645. end;
  646. {$ifdef SUPPORT_MMX}
  647. LOC_MMXREGISTER:
  648. begin
  649. set_location(p^.location,p^.left^.location);
  650. emit_reg_reg(A_PXOR,S_NO,R_MM7,R_MM7);
  651. do_mmx_neg;
  652. end;
  653. LOC_CMMXREGISTER:
  654. begin
  655. p^.location.register:=getregistermmx;
  656. emit_reg_reg(A_PXOR,S_NO,R_MM7,R_MM7);
  657. emit_reg_reg(A_MOVQ,S_NO,p^.left^.location.register,
  658. p^.location.register);
  659. do_mmx_neg;
  660. end;
  661. {$endif SUPPORT_MMX}
  662. LOC_REFERENCE,LOC_MEM:
  663. begin
  664. del_reference(p^.left^.location.reference);
  665. if (p^.left^.resulttype^.deftype=floatdef) and
  666. (pfloatdef(p^.left^.resulttype)^.typ<>f32bit) then
  667. begin
  668. p^.location.loc:=LOC_FPU;
  669. floatload(pfloatdef(p^.left^.resulttype)^.typ,
  670. p^.left^.location.reference);
  671. exprasmlist^.concat(new(pai386,op_none(A_FCHS,S_NO)));
  672. end
  673. {$ifdef SUPPORT_MMX}
  674. else if (cs_mmx in aktlocalswitches) and is_mmx_able_array(p^.left^.resulttype) then
  675. begin
  676. p^.location.register:=getregistermmx;
  677. emit_reg_reg(A_PXOR,S_NO,R_MM7,R_MM7);
  678. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOVQ,S_NO,
  679. newreference(p^.left^.location.reference),
  680. p^.location.register)));
  681. do_mmx_neg;
  682. end
  683. {$endif SUPPORT_MMX}
  684. else
  685. begin
  686. p^.location.register:=getregister32;
  687. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,
  688. newreference(p^.left^.location.reference),
  689. p^.location.register)));
  690. exprasmlist^.concat(new(pai386,op_reg(A_NEG,S_L,p^.location.register)));
  691. end;
  692. end;
  693. LOC_FPU:
  694. begin
  695. p^.location.loc:=LOC_FPU;
  696. exprasmlist^.concat(new(pai386,op_none(A_FCHS,S_NO)));
  697. end;
  698. end;
  699. end;
  700. { Here was a problem... }
  701. { Operand to be negated always }
  702. { seems to be converted to signed }
  703. { 32-bit before doing neg!! }
  704. { So this is useless... }
  705. { emitoverflowcheck(p);}
  706. end;
  707. {*****************************************************************************
  708. SecondNot
  709. *****************************************************************************}
  710. procedure secondnot(var p : ptree);
  711. const
  712. flagsinvers : array[F_E..F_BE] of tresflags =
  713. (F_NE,F_E,F_LE,F_GE,F_L,F_G,F_NC,F_C,
  714. F_A,F_AE,F_B,F_BE);
  715. var
  716. hl : pasmlabel;
  717. opsize : topsize;
  718. hr : preference;
  719. begin
  720. if is_boolean(p^.resulttype) then
  721. begin
  722. opsize:=def_opsize(p^.resulttype);
  723. case p^.left^.location.loc of
  724. LOC_JUMP :
  725. begin
  726. hl:=truelabel;
  727. truelabel:=falselabel;
  728. falselabel:=hl;
  729. secondpass(p^.left);
  730. maketojumpbool(p^.left);
  731. hl:=truelabel;
  732. truelabel:=falselabel;
  733. falselabel:=hl;
  734. end;
  735. LOC_FLAGS :
  736. begin
  737. secondpass(p^.left);
  738. p^.location.resflags:=flagsinvers[p^.left^.location.resflags];
  739. end;
  740. LOC_REGISTER :
  741. begin
  742. secondpass(p^.left);
  743. {p^.location.register:=p^.left^.location.register;
  744. exprasmlist^.concat(new(pai386,op_const_reg(A_XOR,opsize,1,p^.location.register)));}
  745. p^.location.loc:=LOC_FLAGS;
  746. p^.location.resflags:=F_E;
  747. exprasmlist^.concat(new(pai386,op_reg_reg(A_TEST,opsize,
  748. p^.left^.location.register,p^.left^.location.register)));
  749. ungetregister(p^.left^.location.register);
  750. end;
  751. LOC_CREGISTER :
  752. begin
  753. secondpass(p^.left);
  754. clear_location(p^.location);
  755. p^.location.loc:=LOC_REGISTER;
  756. p^.location.register:=def_getreg(p^.resulttype);
  757. emit_reg_reg(A_MOV,opsize,p^.left^.location.register,p^.location.register);
  758. exprasmlist^.concat(new(pai386,op_reg_reg(A_TEST,opsize,p^.location.register,p^.location.register)));
  759. ungetregister(p^.location.register);
  760. p^.location.loc:=LOC_FLAGS;
  761. p^.location.resflags:=F_E;
  762. end;
  763. LOC_REFERENCE,
  764. LOC_MEM :
  765. begin
  766. secondpass(p^.left);
  767. clear_location(p^.location);
  768. p^.location.loc:=LOC_REGISTER;
  769. del_reference(p^.left^.location.reference);
  770. { this was placed before del_ref => internaalerror(10) }
  771. p^.location.register:=def_getreg(p^.resulttype);
  772. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,opsize,
  773. newreference(p^.left^.location.reference),p^.location.register)));
  774. exprasmlist^.concat(new(pai386,op_reg_reg(A_TEST,opsize,p^.location.register,p^.location.register)));
  775. ungetregister(p^.location.register);
  776. p^.location.loc:=LOC_FLAGS;
  777. p^.location.resflags:=F_E;
  778. end;
  779. end;
  780. end
  781. {$ifdef SUPPORT_MMX}
  782. else
  783. if (cs_mmx in aktlocalswitches) and is_mmx_able_array(p^.left^.resulttype) then
  784. begin
  785. secondpass(p^.left);
  786. p^.location.loc:=LOC_MMXREGISTER;
  787. { prepare EDI }
  788. exprasmlist^.concat(new(pai386,op_const_reg(A_MOV,S_L,$ffffffff,R_EDI)));
  789. { load operand }
  790. case p^.left^.location.loc of
  791. LOC_MMXREGISTER:
  792. set_location(p^.location,p^.left^.location);
  793. LOC_CMMXREGISTER:
  794. begin
  795. p^.location.register:=getregistermmx;
  796. emit_reg_reg(A_MOVQ,S_NO,p^.left^.location.register,p^.location.register);
  797. end;
  798. LOC_REFERENCE,LOC_MEM:
  799. begin
  800. del_reference(p^.left^.location.reference);
  801. p^.location.register:=getregistermmx;
  802. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOVQ,S_NO,
  803. newreference(p^.left^.location.reference),p^.location.register)));
  804. end;
  805. end;
  806. { load mask }
  807. emit_reg_reg(A_MOV,S_D,R_EDI,R_MM7);
  808. { lower 32 bit }
  809. emit_reg_reg(A_PXOR,S_D,R_MM7,p^.location.register);
  810. { shift mask }
  811. exprasmlist^.concat(new(pai386,op_const_reg(A_PSLLQ,S_NO,32,R_MM7)));
  812. { higher 32 bit }
  813. emit_reg_reg(A_PXOR,S_D,R_MM7,p^.location.register);
  814. end
  815. {$endif SUPPORT_MMX}
  816. else if is_64bitint(p^.left^.resulttype) then
  817. begin
  818. secondpass(p^.left);
  819. clear_location(p^.location);
  820. p^.location.loc:=LOC_REGISTER;
  821. case p^.left^.location.loc of
  822. LOC_REGISTER :
  823. begin
  824. p^.location.registerlow:=p^.left^.location.registerlow;
  825. p^.location.registerhigh:=p^.left^.location.registerhigh;
  826. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.registerlow)));
  827. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.registerhigh)));
  828. end;
  829. LOC_CREGISTER :
  830. begin
  831. p^.location.registerlow:=getregister32;
  832. p^.location.registerhigh:=getregister32;
  833. emit_reg_reg(A_MOV,S_L,p^.left^.location.registerlow,p^.location.registerlow);
  834. emit_reg_reg(A_MOV,S_L,p^.left^.location.registerhigh,p^.location.registerhigh);
  835. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.registerlow)));
  836. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.registerhigh)));
  837. end;
  838. LOC_REFERENCE,LOC_MEM :
  839. begin
  840. del_reference(p^.left^.location.reference);
  841. p^.location.registerlow:=getregister32;
  842. p^.location.registerhigh:=getregister32;
  843. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,
  844. newreference(p^.left^.location.reference),p^.location.registerlow)));
  845. hr:=newreference(p^.left^.location.reference);
  846. inc(hr^.offset,4);
  847. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,
  848. hr,p^.location.registerhigh)));
  849. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.registerlow)));
  850. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.registerhigh)));
  851. end;
  852. end;
  853. end
  854. else
  855. begin
  856. secondpass(p^.left);
  857. clear_location(p^.location);
  858. p^.location.loc:=LOC_REGISTER;
  859. case p^.left^.location.loc of
  860. LOC_REGISTER :
  861. begin
  862. p^.location.register:=p^.left^.location.register;
  863. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.register)));
  864. end;
  865. LOC_CREGISTER :
  866. begin
  867. p^.location.register:=getregister32;
  868. emit_reg_reg(A_MOV,S_L,p^.left^.location.register,p^.location.register);
  869. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.register)));
  870. end;
  871. LOC_REFERENCE,LOC_MEM :
  872. begin
  873. del_reference(p^.left^.location.reference);
  874. p^.location.register:=getregister32;
  875. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,
  876. newreference(p^.left^.location.reference),p^.location.register)));
  877. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.register)));
  878. end;
  879. end;
  880. end;
  881. end;
  882. end.
  883. {
  884. $Log$
  885. Revision 1.29 1999-08-04 00:22:51 florian
  886. * renamed i386asm and i386base to cpuasm and cpubase
  887. Revision 1.28 1999/08/03 22:02:45 peter
  888. * moved bitmask constants to sets
  889. * some other type/const renamings
  890. Revision 1.27 1999/06/28 22:29:14 florian
  891. * qword division fixed
  892. + code for qword/int64 type casting added:
  893. range checking isn't implemented yet
  894. Revision 1.26 1999/06/02 10:11:44 florian
  895. * make cycle fixed i.e. compilation with 0.99.10
  896. * some fixes for qword
  897. * start of register calling conventions
  898. Revision 1.25 1999/05/27 19:44:16 peter
  899. * removed oldasm
  900. * plabel -> pasmlabel
  901. * -a switches to source writing automaticly
  902. * assembler readers OOPed
  903. * asmsymbol automaticly external
  904. * jumptables and other label fixes for asm readers
  905. Revision 1.24 1999/05/25 20:36:13 florian
  906. * some bugs in the qword code generation fixed
  907. Revision 1.23 1999/05/08 20:41:08 jonas
  908. + positive number MOD power of 2 now done with AND instruction
  909. * fix to division of positive numbers by power of 2
  910. * the result of a MOD is left in EDX if possible
  911. Revision 1.22 1999/05/01 13:24:11 peter
  912. * merged nasm compiler
  913. * old asm moved to oldasm/
  914. Revision 1.21 1999/04/16 13:42:27 jonas
  915. * more regalloc fixes (still not complete)
  916. Revision 1.20 1999/02/22 02:15:13 peter
  917. * updates for ag386bin
  918. Revision 1.19 1999/02/04 10:49:40 florian
  919. + range checking for ansi- and widestrings
  920. * made it compilable with TP
  921. Revision 1.18 1999/02/03 10:11:12 pierre
  922. * fix for bug0211 for i386
  923. Revision 1.17 1999/01/21 22:10:41 peter
  924. * fixed array of const
  925. * generic platform independent high() support
  926. Revision 1.16 1999/01/19 10:51:32 pierre
  927. * fix to bug0183 in secondnot
  928. Revision 1.15 1998/12/11 16:50:22 florian
  929. + typed const int64 and qword
  930. + unary minus-operator q1:=-q2;
  931. + not-operator
  932. Revision 1.14 1998/12/11 16:10:07 florian
  933. + shifting for 64 bit ints added
  934. * bug in getexplicitregister32 fixed: usableregs wasn't decremented !!
  935. Revision 1.13 1998/12/11 00:02:52 peter
  936. + globtype,tokens,version unit splitted from globals
  937. Revision 1.12 1998/11/26 21:45:29 jonas
  938. - removed A_CLTD opcode (use A_CDQ instead)
  939. * changed cbw, cwde and cwd to cbtw, cwtl and cwtd in att_op2str array
  940. * in daopt386: adapted AsmInstr array to reflect changes + fixed line too long
  941. Revision 1.11 1998/11/05 14:26:02 peter
  942. * fixed shlshr which would push ecx when not needed
  943. Revision 1.10 1998/10/20 13:12:38 peter
  944. * fixed 'not not boolean', the location was not set to register
  945. Revision 1.9 1998/10/20 08:06:42 pierre
  946. * several memory corruptions due to double freemem solved
  947. => never use p^.loc.location:=p^.left^.loc.location;
  948. + finally I added now by default
  949. that ra386dir translates global and unit symbols
  950. + added a first field in tsymtable and
  951. a nextsym field in tsym
  952. (this allows to obtain ordered type info for
  953. records and objects in gdb !)
  954. Revision 1.8 1998/10/09 08:56:24 pierre
  955. * several memory leaks fixed
  956. Revision 1.7 1998/09/17 09:42:17 peter
  957. + pass_2 for cg386
  958. * Message() -> CGMessage() for pass_1/pass_2
  959. Revision 1.6 1998/09/09 14:37:37 florian
  960. * mod/div for cardinal type fixed
  961. Revision 1.5 1998/08/23 16:07:20 florian
  962. * internalerror with mod/div fixed
  963. Revision 1.4 1998/08/18 09:24:38 pierre
  964. * small warning position bug fixed
  965. * support_mmx switches splitting was missing
  966. * rhide error and warning output corrected
  967. Revision 1.3 1998/06/05 17:44:12 peter
  968. * splitted cgi386
  969. Revision 1.2 1998/06/02 17:02:59 pierre
  970. * with node corrected for objects
  971. * small bugs for SUPPORT_MMX fixed
  972. Revision 1.1 1998/06/01 16:50:18 peter
  973. + boolean -> ord conversion
  974. * fixed ord -> boolean conversion
  975. }