cg386mat.pas 24 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588
  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. cobjects,verbose,globals,systems,
  29. symtable,aasm,types,
  30. hcodegen,temp_gen,pass_2,
  31. i386,cgai386,tgeni386;
  32. {*****************************************************************************
  33. SecondModDiv
  34. *****************************************************************************}
  35. procedure secondmoddiv(var p : ptree);
  36. var
  37. hreg1 : tregister;
  38. pushed,popeax,popedx : boolean;
  39. power : longint;
  40. hl : plabel;
  41. begin
  42. secondpass(p^.left);
  43. set_location(p^.location,p^.left^.location);
  44. pushed:=maybe_push(p^.right^.registers32,p);
  45. secondpass(p^.right);
  46. if pushed then restore(p);
  47. { put numerator in register }
  48. if p^.left^.location.loc<>LOC_REGISTER then
  49. begin
  50. if p^.left^.location.loc=LOC_CREGISTER then
  51. begin
  52. hreg1:=getregister32;
  53. emit_reg_reg(A_MOV,S_L,p^.left^.location.register,hreg1);
  54. end
  55. else
  56. begin
  57. del_reference(p^.left^.location.reference);
  58. hreg1:=getregister32;
  59. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.left^.location.reference),
  60. hreg1)));
  61. end;
  62. p^.left^.location.loc:=LOC_REGISTER;
  63. p^.left^.location.register:=hreg1;
  64. end
  65. else hreg1:=p^.left^.location.register;
  66. if (p^.treetype=divn) and (p^.right^.treetype=ordconstn) and
  67. ispowerof2(p^.right^.value,power) then
  68. begin
  69. exprasmlist^.concat(new(pai386,op_reg_reg(A_OR,S_L,hreg1,hreg1)));
  70. getlabel(hl);
  71. emitl(A_JNS,hl);
  72. if power=1 then
  73. exprasmlist^.concat(new(pai386,op_reg(A_INC,S_L,hreg1)))
  74. else exprasmlist^.concat(new(pai386,op_const_reg(A_ADD,S_L,p^.right^.value-1,hreg1)));
  75. emitl(A_LABEL,hl);
  76. exprasmlist^.concat(new(pai386,op_const_reg(A_SAR,S_L,power,hreg1)));
  77. end
  78. else
  79. begin
  80. { bring denominator to EDI }
  81. { EDI is always free, it's }
  82. { only used for temporary }
  83. { purposes }
  84. if (p^.right^.location.loc<>LOC_REGISTER) and
  85. (p^.right^.location.loc<>LOC_CREGISTER) then
  86. begin
  87. del_reference(p^.right^.location.reference);
  88. p^.left^.location.loc:=LOC_REGISTER;
  89. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.right^.location.reference),R_EDI)));
  90. end
  91. else
  92. begin
  93. ungetregister32(p^.right^.location.register);
  94. emit_reg_reg(A_MOV,S_L,p^.right^.location.register,R_EDI);
  95. end;
  96. popedx:=false;
  97. popeax:=false;
  98. if hreg1=R_EDX then
  99. begin
  100. if not(R_EAX in unused) then
  101. begin
  102. exprasmlist^.concat(new(pai386,op_reg(A_PUSH,S_L,R_EAX)));
  103. popeax:=true;
  104. end;
  105. emit_reg_reg(A_MOV,S_L,R_EDX,R_EAX);
  106. end
  107. else
  108. begin
  109. if not(R_EDX in unused) then
  110. begin
  111. exprasmlist^.concat(new(pai386,op_reg(A_PUSH,S_L,R_EDX)));
  112. popedx:=true;
  113. end;
  114. if hreg1<>R_EAX then
  115. begin
  116. if not(R_EAX in unused) then
  117. begin
  118. exprasmlist^.concat(new(pai386,op_reg(A_PUSH,S_L,R_EAX)));
  119. popeax:=true;
  120. end;
  121. emit_reg_reg(A_MOV,S_L,hreg1,R_EAX);
  122. end;
  123. end;
  124. { sign extension depends on the left type }
  125. if porddef(p^.left^.resulttype)^.typ=u32bit then
  126. exprasmlist^.concat(new(pai386,op_reg_reg(A_XOR,S_L,R_EDX,R_EDX)))
  127. else
  128. exprasmlist^.concat(new(pai386,op_none(A_CLTD,S_NO)));
  129. { division depends on the right type }
  130. if porddef(p^.right^.resulttype)^.typ=u32bit then
  131. exprasmlist^.concat(new(pai386,op_reg(A_DIV,S_L,R_EDI)))
  132. else
  133. exprasmlist^.concat(new(pai386,op_reg(A_IDIV,S_L,R_EDI)));
  134. if p^.treetype=divn then
  135. begin
  136. { if result register is busy then copy }
  137. if popeax then
  138. begin
  139. if hreg1=R_EAX then
  140. internalerror(112);
  141. emit_reg_reg(A_MOV,S_L,R_EAX,hreg1)
  142. end
  143. else
  144. if hreg1<>R_EAX then
  145. emit_reg_reg(A_MOV,S_L,R_EAX,hreg1);
  146. end
  147. else
  148. emit_reg_reg(A_MOV,S_L,R_EDX,hreg1);
  149. if popeax then
  150. exprasmlist^.concat(new(pai386,op_reg(A_POP,S_L,R_EAX)));
  151. if popedx then
  152. exprasmlist^.concat(new(pai386,op_reg(A_POP,S_L,R_EDX)));
  153. end;
  154. { this registers are always used when div/mod are present }
  155. usedinproc:=usedinproc or ($80 shr byte(R_EAX));
  156. usedinproc:=usedinproc or ($80 shr byte(R_EDX));
  157. p^.location.loc:=LOC_REGISTER;
  158. p^.location.register:=hreg1;
  159. end;
  160. {*****************************************************************************
  161. SecondShlShr
  162. *****************************************************************************}
  163. procedure secondshlshr(var p : ptree);
  164. var
  165. hregister1,hregister2,hregister3 : tregister;
  166. pushed,popecx : boolean;
  167. op : tasmop;
  168. begin
  169. popecx:=false;
  170. secondpass(p^.left);
  171. pushed:=maybe_push(p^.right^.registers32,p);
  172. secondpass(p^.right);
  173. if pushed then restore(p);
  174. { load left operators in a register }
  175. if p^.left^.location.loc<>LOC_REGISTER then
  176. begin
  177. if p^.left^.location.loc=LOC_CREGISTER then
  178. begin
  179. hregister1:=getregister32;
  180. emit_reg_reg(A_MOV,S_L,p^.left^.location.register,
  181. hregister1);
  182. end
  183. else
  184. begin
  185. del_reference(p^.left^.location.reference);
  186. hregister1:=getregister32;
  187. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.left^.location.reference),
  188. hregister1)));
  189. end;
  190. end
  191. else hregister1:=p^.left^.location.register;
  192. { determine operator }
  193. if p^.treetype=shln then
  194. op:=A_SHL
  195. else
  196. op:=A_SHR;
  197. { shifting by a constant directly decode: }
  198. if (p^.right^.treetype=ordconstn) then
  199. begin
  200. exprasmlist^.concat(new(pai386,op_const_reg(op,S_L,p^.right^.location.reference.offset and 31,
  201. hregister1)));
  202. p^.location.loc:=LOC_REGISTER;
  203. p^.location.register:=hregister1;
  204. end
  205. else
  206. begin
  207. { load right operators in a register }
  208. if p^.right^.location.loc<>LOC_REGISTER then
  209. begin
  210. if p^.right^.location.loc=LOC_CREGISTER then
  211. begin
  212. hregister2:=getregister32;
  213. emit_reg_reg(A_MOV,S_L,p^.right^.location.register,
  214. hregister2);
  215. end
  216. else
  217. begin
  218. del_reference(p^.right^.location.reference);
  219. hregister2:=getregister32;
  220. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,newreference(p^.right^.location.reference),
  221. hregister2)));
  222. end;
  223. end
  224. else hregister2:=p^.right^.location.register;
  225. { left operator is already in a register }
  226. { hence are both in a register }
  227. { is it in the case ECX ? }
  228. if (hregister1=R_ECX) then
  229. begin
  230. { then only swap }
  231. emit_reg_reg(A_XCHG,S_L,hregister1,
  232. hregister2);
  233. hregister3:=hregister1;
  234. hregister1:=hregister2;
  235. hregister2:=hregister3;
  236. end
  237. { if second operator not in ECX ? }
  238. else if (hregister2<>R_ECX) then
  239. begin
  240. { ECX not occupied then swap with right register }
  241. if R_ECX in unused then
  242. begin
  243. emit_reg_reg(A_MOV,S_L,hregister2,R_ECX);
  244. ungetregister32(hregister2);
  245. end
  246. else
  247. begin
  248. { else save ECX and then copy it }
  249. popecx:=true;
  250. exprasmlist^.concat(new(pai386,op_reg(A_PUSH,S_L,R_ECX)));
  251. emit_reg_reg(A_MOV,S_L,hregister2,R_ECX);
  252. ungetregister32(hregister2);
  253. end;
  254. end;
  255. { right operand is in ECX }
  256. emit_reg_reg(op,S_L,R_CL,hregister1);
  257. { maybe ECX back }
  258. if popecx then
  259. exprasmlist^.concat(new(pai386,op_reg(A_POP,S_L,R_ECX)));
  260. p^.location.register:=hregister1;
  261. end;
  262. { this register is always used when shl/shr are present }
  263. usedinproc:=usedinproc or ($80 shr byte(R_ECX));
  264. end;
  265. {*****************************************************************************
  266. SecondUmMinus
  267. *****************************************************************************}
  268. procedure secondumminus(var p : ptree);
  269. {$ifdef SUPPORT_MMX}
  270. procedure do_mmx_neg;
  271. var
  272. op : tasmop;
  273. begin
  274. p^.location.loc:=LOC_MMXREGISTER;
  275. if cs_mmx_saturation in aktlocalswitches then
  276. case mmx_type(p^.resulttype) of
  277. mmxs8bit:
  278. op:=A_PSUBSB;
  279. mmxu8bit:
  280. op:=A_PSUBUSB;
  281. mmxs16bit,mmxfixed16:
  282. op:=A_PSUBSW;
  283. mmxu16bit:
  284. op:=A_PSUBUSW;
  285. end
  286. else
  287. case mmx_type(p^.resulttype) of
  288. mmxs8bit,mmxu8bit:
  289. op:=A_PSUBB;
  290. mmxs16bit,mmxu16bit,mmxfixed16:
  291. op:=A_PSUBW;
  292. mmxs32bit,mmxu32bit:
  293. op:=A_PSUBD;
  294. end;
  295. emit_reg_reg(op,S_NO,p^.location.register,R_MM7);
  296. emit_reg_reg(A_MOVQ,S_NO,R_MM7,p^.location.register);
  297. end;
  298. {$endif}
  299. begin
  300. secondpass(p^.left);
  301. p^.location.loc:=LOC_REGISTER;
  302. case p^.left^.location.loc of
  303. LOC_REGISTER:
  304. begin
  305. p^.location.register:=p^.left^.location.register;
  306. exprasmlist^.concat(new(pai386,op_reg(A_NEG,S_L,p^.location.register)));
  307. end;
  308. LOC_CREGISTER:
  309. begin
  310. p^.location.register:=getregister32;
  311. emit_reg_reg(A_MOV,S_L,p^.location.register,
  312. p^.location.register);
  313. exprasmlist^.concat(new(pai386,op_reg(A_NEG,S_L,p^.location.register)));
  314. end;
  315. {$ifdef SUPPORT_MMX}
  316. LOC_MMXREGISTER:
  317. begin
  318. p^.location:=p^.left^.location;
  319. emit_reg_reg(A_PXOR,S_NO,R_MM7,R_MM7);
  320. do_mmx_neg;
  321. end;
  322. LOC_CMMXREGISTER:
  323. begin
  324. p^.location.register:=getregistermmx;
  325. emit_reg_reg(A_PXOR,S_NO,R_MM7,R_MM7);
  326. emit_reg_reg(A_MOVQ,S_NO,p^.left^.location.register,
  327. p^.location.register);
  328. do_mmx_neg;
  329. end;
  330. {$endif SUPPORT_MMX}
  331. LOC_REFERENCE,LOC_MEM:
  332. begin
  333. del_reference(p^.left^.location.reference);
  334. if (p^.left^.resulttype^.deftype=floatdef) and
  335. (pfloatdef(p^.left^.resulttype)^.typ<>f32bit) then
  336. begin
  337. p^.location.loc:=LOC_FPU;
  338. floatload(pfloatdef(p^.left^.resulttype)^.typ,
  339. p^.left^.location.reference);
  340. exprasmlist^.concat(new(pai386,op_none(A_FCHS,S_NO)));
  341. end
  342. {$ifdef SUPPORT_MMX}
  343. else if (cs_mmx in aktlocalswitches) and is_mmx_able_array(p^.left^.resulttype) then
  344. begin
  345. p^.location.register:=getregistermmx;
  346. emit_reg_reg(A_PXOR,S_NO,R_MM7,R_MM7);
  347. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOVQ,S_NO,
  348. newreference(p^.left^.location.reference),
  349. p^.location.register)));
  350. do_mmx_neg;
  351. end
  352. {$endif SUPPORT_MMX}
  353. else
  354. begin
  355. p^.location.register:=getregister32;
  356. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,
  357. newreference(p^.left^.location.reference),
  358. p^.location.register)));
  359. exprasmlist^.concat(new(pai386,op_reg(A_NEG,S_L,p^.location.register)));
  360. end;
  361. end;
  362. LOC_FPU:
  363. begin
  364. p^.location.loc:=LOC_FPU;
  365. exprasmlist^.concat(new(pai386,op_none(A_FCHS,S_NO)));
  366. end;
  367. end;
  368. { Here was a problem... }
  369. { Operand to be negated always }
  370. { seems to be converted to signed }
  371. { 32-bit before doing neg!! }
  372. { So this is useless... }
  373. { emitoverflowcheck(p);}
  374. end;
  375. {*****************************************************************************
  376. SecondNot
  377. *****************************************************************************}
  378. procedure secondnot(var p : ptree);
  379. const
  380. flagsinvers : array[F_E..F_BE] of tresflags =
  381. (F_NE,F_E,F_LE,F_GE,F_L,F_G,F_NC,F_C,
  382. F_A,F_AE,F_B,F_BE);
  383. var
  384. hl : plabel;
  385. opsize : topsize;
  386. begin
  387. if (p^.resulttype^.deftype=orddef) and
  388. (porddef(p^.resulttype)^.typ in [bool8bit,bool16bit,bool32bit]) then
  389. begin
  390. case porddef(p^.resulttype)^.typ of
  391. bool8bit : opsize:=S_B;
  392. bool16bit : opsize:=S_W;
  393. bool32bit : opsize:=S_L;
  394. end;
  395. case p^.location.loc of
  396. LOC_JUMP : begin
  397. hl:=truelabel;
  398. truelabel:=falselabel;
  399. falselabel:=hl;
  400. secondpass(p^.left);
  401. maketojumpbool(p^.left);
  402. hl:=truelabel;
  403. truelabel:=falselabel;
  404. falselabel:=hl;
  405. end;
  406. LOC_FLAGS : begin
  407. secondpass(p^.left);
  408. p^.location.resflags:=flagsinvers[p^.left^.location.resflags];
  409. end;
  410. LOC_REGISTER : begin
  411. secondpass(p^.left);
  412. p^.location.register:=p^.left^.location.register;
  413. exprasmlist^.concat(new(pai386,op_const_reg(A_XOR,opsize,1,p^.location.register)));
  414. end;
  415. LOC_CREGISTER : begin
  416. secondpass(p^.left);
  417. p^.location.loc:=LOC_REGISTER;
  418. case porddef(p^.resulttype)^.typ of
  419. bool8bit : p^.location.register:=reg32toreg8(getregister32);
  420. bool16bit : p^.location.register:=reg32toreg16(getregister32);
  421. bool32bit : p^.location.register:=getregister32;
  422. end;
  423. emit_reg_reg(A_MOV,opsize,p^.left^.location.register,p^.location.register);
  424. exprasmlist^.concat(new(pai386,op_const_reg(A_XOR,opsize,1,p^.location.register)));
  425. end;
  426. LOC_REFERENCE,
  427. LOC_MEM : begin
  428. secondpass(p^.left);
  429. del_reference(p^.left^.location.reference);
  430. p^.location.loc:=LOC_REGISTER;
  431. case porddef(p^.resulttype)^.typ of
  432. bool8bit : p^.location.register:=reg32toreg8(getregister32);
  433. bool16bit : p^.location.register:=reg32toreg16(getregister32);
  434. bool32bit : p^.location.register:=getregister32;
  435. end;
  436. if p^.left^.location.loc=LOC_CREGISTER then
  437. emit_reg_reg(A_MOV,opsize,p^.left^.location.register,p^.location.register)
  438. else
  439. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,opsize,
  440. newreference(p^.left^.location.reference),p^.location.register)));
  441. exprasmlist^.concat(new(pai386,op_const_reg(A_XOR,opsize,1,p^.location.register)));
  442. end;
  443. end;
  444. end
  445. {$ifdef SUPPORT_MMX}
  446. else if (cs_mmx in aktlocalswitches) and is_mmx_able_array(p^.left^.resulttype) then
  447. begin
  448. secondpass(p^.left);
  449. p^.location.loc:=LOC_MMXREGISTER;
  450. { prepare EDI }
  451. exprasmlist^.concat(new(pai386,op_const_reg(A_MOV,S_L,$ffffffff,R_EDI)));
  452. { load operand }
  453. case p^.left^.location.loc of
  454. LOC_MMXREGISTER:
  455. p^.location:=p^.left^.location;
  456. LOC_CMMXREGISTER:
  457. begin
  458. p^.location.register:=getregistermmx;
  459. emit_reg_reg(A_MOVQ,S_NO,p^.left^.location.register,
  460. p^.location.register);
  461. end;
  462. LOC_REFERENCE,LOC_MEM:
  463. begin
  464. del_reference(p^.left^.location.reference);
  465. p^.location.register:=getregistermmx;
  466. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOVQ,S_NO,
  467. newreference(p^.left^.location.reference),
  468. p^.location.register)));
  469. end;
  470. end;
  471. { load mask }
  472. emit_reg_reg(A_MOV,S_D,R_EDI,R_MM7);
  473. { lower 32 bit }
  474. emit_reg_reg(A_PXOR,S_D,R_MM7,p^.location.register);
  475. { shift mask }
  476. exprasmlist^.concat(new(pai386,op_const_reg(A_PSLLQ,S_NO,
  477. 32,R_MM7)));
  478. { higher 32 bit }
  479. emit_reg_reg(A_PXOR,S_D,R_MM7,p^.location.register);
  480. end
  481. {$endif SUPPORT_MMX}
  482. else
  483. begin
  484. secondpass(p^.left);
  485. p^.location.loc:=LOC_REGISTER;
  486. case p^.left^.location.loc of
  487. LOC_REGISTER : begin
  488. p^.location.register:=p^.left^.location.register;
  489. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.register)));
  490. end;
  491. LOC_CREGISTER : begin
  492. p^.location.register:=getregister32;
  493. emit_reg_reg(A_MOV,S_L,p^.left^.location.register,
  494. p^.location.register);
  495. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.register)));
  496. end;
  497. LOC_REFERENCE,LOC_MEM :
  498. begin
  499. del_reference(p^.left^.location.reference);
  500. p^.location.register:=getregister32;
  501. exprasmlist^.concat(new(pai386,op_ref_reg(A_MOV,S_L,
  502. newreference(p^.left^.location.reference),
  503. p^.location.register)));
  504. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.register)));
  505. end;
  506. end;
  507. {if p^.left^.location.loc=loc_register then
  508. p^.location.register:=p^.left^.location.register
  509. else
  510. begin
  511. del_locref(p^.left^.location);
  512. p^.location.register:=getregister32;
  513. exprasmlist^.concat(new(pai386,op_loc_reg(A_MOV,S_L,
  514. p^.left^.location,
  515. p^.location.register)));
  516. end;
  517. exprasmlist^.concat(new(pai386,op_reg(A_NOT,S_L,p^.location.register)));}
  518. end;
  519. end;
  520. end.
  521. {
  522. $Log$
  523. Revision 1.7 1998-09-17 09:42:17 peter
  524. + pass_2 for cg386
  525. * Message() -> CGMessage() for pass_1/pass_2
  526. Revision 1.6 1998/09/09 14:37:37 florian
  527. * mod/div for cardinal type fixed
  528. Revision 1.5 1998/08/23 16:07:20 florian
  529. * internalerror with mod/div fixed
  530. Revision 1.4 1998/08/18 09:24:38 pierre
  531. * small warning position bug fixed
  532. * support_mmx switches splitting was missing
  533. * rhide error and warning output corrected
  534. Revision 1.3 1998/06/05 17:44:12 peter
  535. * splitted cgi386
  536. Revision 1.2 1998/06/02 17:02:59 pierre
  537. * with node corrected for objects
  538. * small bugs for SUPPORT_MMX fixed
  539. Revision 1.1 1998/06/01 16:50:18 peter
  540. + boolean -> ord conversion
  541. * fixed ord -> boolean conversion
  542. }