ncgset.pas 35 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856
  1. {
  2. Copyright (c) 1998-2002 by Florian Klaempfl and Carl Eric Codere
  3. Generate generic assembler for in set/case labels
  4. This program is free software; you can redistribute it and/or modify
  5. it under the terms of the GNU General Public License as published by
  6. the Free Software Foundation; either version 2 of the License, or
  7. (at your option) any later version.
  8. This program is distributed in the hope that it will be useful,
  9. but WITHOUT ANY WARRANTY; without even the implied warranty of
  10. MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
  11. GNU General Public License for more details.
  12. You should have received a copy of the GNU General Public License
  13. along with this program; if not, write to the Free Software
  14. Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  15. ****************************************************************************
  16. }
  17. unit ncgset;
  18. {$i fpcdefs.inc}
  19. interface
  20. uses
  21. globtype,globals,constexp,
  22. node,nset,cpubase,cgbase,cgutils,cgobj,aasmbase,aasmtai,aasmdata;
  23. type
  24. tcgsetelementnode = class(tsetelementnode)
  25. procedure pass_generate_code;override;
  26. end;
  27. Tsetpart=record
  28. range : boolean; {Part is a range.}
  29. start,stop : byte; {Start/stop when range; Stop=element when an element.}
  30. end;
  31. Tsetparts=array[1..8] of Tsetpart;
  32. tcginnode = class(tinnode)
  33. function pass_1: tnode;override;
  34. procedure pass_generate_code;override;
  35. protected
  36. function checkgenjumps(out setparts: Tsetparts; out numparts: byte; out use_small: boolean): boolean; virtual;
  37. function analizeset(const Aset:Tconstset;out setparts: Tsetparts; out numparts: byte;is_small:boolean):boolean;virtual;
  38. end;
  39. tcgcasenode = class(tcasenode)
  40. {
  41. Emits the case node statement. Contrary to the intel
  42. 80x86 version, this version does not emit jump tables,
  43. because of portability problems.
  44. }
  45. procedure pass_generate_code;override;
  46. protected
  47. with_sign : boolean;
  48. opsize : tcgsize;
  49. jmp_gt,jmp_lt,jmp_le : topcmp;
  50. { register with case expression }
  51. hregister,hregister2 : tregister;
  52. endlabel,elselabel : tasmlabel;
  53. { true, if we can omit the range check of the jump table }
  54. jumptable_no_range : boolean;
  55. { has the implementation jumptable support }
  56. min_label : tconstexprint;
  57. function blocklabel(id:longint):tasmlabel;
  58. procedure optimizevalues(var max_linear_list:aint;var max_dist:aword);virtual;
  59. function has_jumptable : boolean;virtual;
  60. procedure genjumptable(hp : pcaselabel;min_,max_ : aint); virtual;
  61. procedure genlinearlist(hp : pcaselabel); virtual;
  62. procedure genlinearcmplist(hp : pcaselabel); virtual;
  63. end;
  64. implementation
  65. uses
  66. systems,
  67. verbose,
  68. symconst,symdef,defutil,
  69. paramgr,
  70. procinfo,pass_2,tgobj,
  71. nbas,ncon,nflw,
  72. ncgutil;
  73. {*****************************************************************************
  74. TCGSETELEMENTNODE
  75. *****************************************************************************}
  76. procedure tcgsetelementnode.pass_generate_code;
  77. begin
  78. { load first value in 32bit register }
  79. secondpass(left);
  80. if left.location.loc in [LOC_REGISTER,LOC_CREGISTER] then
  81. location_force_reg(current_asmdata.CurrAsmList,left.location,OS_32,false);
  82. { also a second value ? }
  83. if assigned(right) then
  84. begin
  85. secondpass(right);
  86. if codegenerror then
  87. exit;
  88. if right.location.loc in [LOC_REGISTER,LOC_CREGISTER] then
  89. location_force_reg(current_asmdata.CurrAsmList,right.location,OS_32,false);
  90. end;
  91. { we doesn't modify the left side, we check only the type }
  92. location_copy(location,left.location);
  93. end;
  94. {*****************************************************************************
  95. *****************************************************************************}
  96. function tcginnode.analizeset(const Aset:Tconstset; out setparts:tsetparts; out numparts: byte; is_small:boolean):boolean;
  97. var
  98. compares,maxcompares:word;
  99. i:byte;
  100. begin
  101. analizeset:=false;
  102. numparts:=0;
  103. compares:=0;
  104. { Lots of comparisions take a lot of time, so do not allow
  105. too much comparisions. 8 comparisions are, however, still
  106. smalller than emitting the set }
  107. if cs_opt_size in current_settings.optimizerswitches then
  108. maxcompares:=8
  109. else
  110. maxcompares:=5;
  111. { when smallset is possible allow only 3 compares the smallset
  112. code is for littlesize also smaller when more compares are used }
  113. if is_small then
  114. maxcompares:=3;
  115. for i:=0 to 255 do
  116. if i in Aset then
  117. begin
  118. if (numparts=0) or (i<>setparts[numparts].stop+1) then
  119. begin
  120. {Set element is a separate element.}
  121. inc(compares);
  122. if compares>maxcompares then
  123. exit;
  124. inc(numparts);
  125. setparts[numparts].range:=false;
  126. setparts[numparts].stop:=i;
  127. end
  128. else
  129. {Set element is part of a range.}
  130. if not setparts[numparts].range then
  131. begin
  132. {Transform an element into a range.}
  133. setparts[numparts].range:=true;
  134. setparts[numparts].start:=setparts[numparts].stop;
  135. setparts[numparts].stop:=i;
  136. { there's only one compare per range anymore. Only a }
  137. { sub is added, but that's much faster than a }
  138. { cmp/jcc combo so neglect its effect }
  139. { inc(compares);
  140. if compares>maxcompares then
  141. exit; }
  142. end
  143. else
  144. begin
  145. {Extend a range.}
  146. setparts[numparts].stop:=i;
  147. end;
  148. end;
  149. analizeset:=true;
  150. end;
  151. function tcginnode.checkgenjumps(out setparts: Tsetparts; out numparts: byte;out use_small: boolean): boolean;
  152. begin
  153. { check if we can use smallset operation using btl which is limited
  154. to 32 bits, the left side may also not contain higher values !! }
  155. use_small:=(tsetdef(right.resultdef).settype=smallset) and not is_signed(left.resultdef) and
  156. ((left.resultdef.typ=orddef) and (torddef(left.resultdef).high<32) or
  157. (left.resultdef.typ=enumdef) and (tenumdef(left.resultdef).max<32));
  158. { Can we generate jumps? Possible for all types of sets }
  159. checkgenjumps:=(right.nodetype=setconstn) and
  160. analizeset(Tsetconstnode(right).value_set^,setparts,numparts,use_small);
  161. end;
  162. function tcginnode.pass_1: tnode;
  163. var
  164. setparts: Tsetparts;
  165. numparts: byte;
  166. use_small: boolean;
  167. begin
  168. result := inherited pass_1;
  169. if not(assigned(result)) and
  170. checkgenjumps(setparts,numparts,use_small) then
  171. expectloc := LOC_JUMP;
  172. end;
  173. procedure tcginnode.pass_generate_code;
  174. var
  175. adjustment,
  176. setbase : aint;
  177. l, l2 : tasmlabel;
  178. otl, ofl : tasmlabel;
  179. hr,
  180. pleftreg : tregister;
  181. setparts : Tsetparts;
  182. opsize : tcgsize;
  183. uopsize : tcgsize;
  184. orgopsize : tcgsize;
  185. genjumps,
  186. use_small,
  187. isjump : boolean;
  188. i,numparts : byte;
  189. needslabel : Boolean;
  190. begin
  191. { We check first if we can generate jumps, this can be done
  192. because the resultdef is already set in firstpass }
  193. genjumps := checkgenjumps(setparts,numparts,use_small);
  194. orgopsize := def_cgsize(left.resultdef);
  195. uopsize := OS_32;
  196. if is_signed(left.resultdef) then
  197. opsize := tcgsize(ord(uopsize)+(ord(OS_S8)-ord(OS_8)))
  198. else
  199. opsize := uopsize;
  200. needslabel := false;
  201. isjump:=false;
  202. if (left.expectloc=LOC_JUMP) then
  203. begin
  204. otl:=current_procinfo.CurrTrueLabel;
  205. current_asmdata.getjumplabel(current_procinfo.CurrTrueLabel);
  206. ofl:=current_procinfo.CurrFalseLabel;
  207. current_asmdata.getjumplabel(current_procinfo.CurrFalseLabel);
  208. isjump:=true;
  209. end
  210. else if not genjumps then
  211. { calculate both operators }
  212. { the complex one first }
  213. { only if left will not be a LOC_JUMP, to keep complexity in the }
  214. { code generator down. This almost never happens anyway, only in }
  215. { case like "if ((a in someset) in someboolset) then" etc }
  216. { also not in case of genjumps, because then we don't secondpass }
  217. { right at all (so we have to make sure that "right" really is }
  218. { "right" and not "swapped left" in that case) }
  219. firstcomplex(self);
  220. secondpass(left);
  221. if isjump then
  222. begin
  223. location_force_reg(current_asmdata.CurrAsmList,left.location,opsize,true);
  224. current_procinfo.CurrTrueLabel:=otl;
  225. current_procinfo.CurrFalseLabel:=ofl;
  226. end
  227. else if (left.location.loc=LOC_JUMP) then
  228. internalerror(2007070101);
  229. { Only process the right if we are not generating jumps }
  230. if not genjumps then
  231. secondpass(right);
  232. if codegenerror then
  233. exit;
  234. { ofcourse not commutative }
  235. if nf_swapped in flags then
  236. swapleftright;
  237. setbase:=tsetdef(right.resultdef).setbase;
  238. if genjumps then
  239. begin
  240. { location is always LOC_JUMP }
  241. location_reset(location,LOC_JUMP,OS_NO);
  242. { If register is used, use only lower 8 bits }
  243. location_force_reg(current_asmdata.CurrAsmList,left.location,opsize,false);
  244. pleftreg := left.location.register;
  245. { how much have we already substracted from the x in the }
  246. { "x in [y..z]" expression }
  247. adjustment := 0;
  248. hr:=NR_NO;
  249. for i:=1 to numparts do
  250. if setparts[i].range then
  251. { use fact that a <= x <= b <=> aword(x-a) <= aword(b-a) }
  252. begin
  253. { is the range different from all legal values? }
  254. if (setparts[i].stop-setparts[i].start <> 255) or not (orgopsize = OS_8) then
  255. begin
  256. { yes, is the lower bound <> 0? }
  257. if (setparts[i].start <> 0) then
  258. { we're going to substract from the left register, }
  259. { so in case of a LOC_CREGISTER first move the value }
  260. { to edi (not done before because now we can do the }
  261. { move and substract in one instruction with LEA) }
  262. if (left.location.loc = LOC_CREGISTER) and
  263. (hr<>pleftreg) then
  264. begin
  265. { don't change this back to a_op_const_reg/a_load_reg_reg, since pleftreg must not be modified }
  266. hr:=cg.getintregister(current_asmdata.CurrAsmList,opsize);
  267. cg.a_op_const_reg_reg(current_asmdata.CurrAsmList,OP_SUB,opsize,setparts[i].start,pleftreg,hr);
  268. pleftreg:=hr;
  269. end
  270. else
  271. begin
  272. { otherwise, the value is already in a register }
  273. { that can be modified }
  274. cg.a_op_const_reg(current_asmdata.CurrAsmList,OP_SUB,opsize,
  275. setparts[i].start-adjustment,pleftreg)
  276. end;
  277. { new total value substracted from x: }
  278. { adjustment + (setparts[i].start - adjustment) }
  279. adjustment := setparts[i].start;
  280. { check if result < b-a+1 (not "result <= b-a", since }
  281. { we need a carry in case the element is in the range }
  282. { (this will never overflow since we check at the }
  283. { beginning whether stop-start <> 255) }
  284. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, opsize, OC_B,
  285. setparts[i].stop-setparts[i].start+1,pleftreg,current_procinfo.CurrTrueLabel);
  286. end
  287. else
  288. { if setparts[i].start = 0 and setparts[i].stop = 255, }
  289. { it's always true since "in" is only allowed for bytes }
  290. begin
  291. cg.a_jmp_always(current_asmdata.CurrAsmList,current_procinfo.CurrTrueLabel);
  292. end;
  293. end
  294. else
  295. begin
  296. { Emit code to check if left is an element }
  297. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, opsize, OC_EQ,
  298. setparts[i].stop-adjustment,pleftreg,current_procinfo.CurrTrueLabel);
  299. end;
  300. { To compensate for not doing a second pass }
  301. right.location.reference.symbol:=nil;
  302. cg.a_jmp_always(current_asmdata.CurrAsmList,current_procinfo.CurrFalseLabel);
  303. end
  304. else
  305. {*****************************************************************}
  306. { NO JUMP TABLE GENERATION }
  307. {*****************************************************************}
  308. begin
  309. { location is always LOC_REGISTER }
  310. location_reset(location, LOC_REGISTER, uopsize{def_cgsize(resultdef)});
  311. { allocate a register for the result }
  312. location.register := cg.getintregister(current_asmdata.CurrAsmList, uopsize);
  313. { We will now generated code to check the set itself, no jmps,
  314. handle smallsets separate, because it allows faster checks }
  315. if use_small then
  316. begin
  317. {**************************** SMALL SET **********************}
  318. if left.location.loc=LOC_CONSTANT then
  319. begin
  320. cg.a_bit_test_const_loc_reg(current_asmdata.CurrAsmList,location.size,
  321. left.location.value-setbase,right.location,
  322. location.register);
  323. end
  324. else
  325. begin
  326. location_force_reg(current_asmdata.CurrAsmList,left.location,opsize,true);
  327. register_maybe_adjust_setbase(current_asmdata.CurrAsmList,left.location,setbase);
  328. cg.a_bit_test_reg_loc_reg(current_asmdata.CurrAsmList,left.location.size,
  329. location.size,left.location.register,right.location,location.register);
  330. end;
  331. end
  332. else
  333. {************************** NOT SMALL SET ********************}
  334. begin
  335. if right.location.loc=LOC_CONSTANT then
  336. begin
  337. { can it actually occur currently? CEC }
  338. { yes: "if bytevar in [1,3,5,7,9,11,13,15]" (JM) }
  339. { note: this code assumes that left in [0..255], which is a valid }
  340. { assumption (other cases will be caught by range checking) (JM) }
  341. { load left in register }
  342. location_force_reg(current_asmdata.CurrAsmList,left.location,location.size,true);
  343. register_maybe_adjust_setbase(current_asmdata.CurrAsmList,left.location,setbase);
  344. location_force_reg(current_asmdata.CurrAsmList,right.location,opsize,true);
  345. { emit bit test operation }
  346. cg.a_bit_test_reg_reg_reg(current_asmdata.CurrAsmList,
  347. left.location.size,right.location.size,location.size,
  348. left.location.register,right.location.register,location.register);
  349. { now zero the result if left > nr_of_bits_in_right_register }
  350. hr := cg.getintregister(current_asmdata.CurrAsmList,location.size);
  351. { if left > tcgsize2size[opsize]*8 then hr := 0 else hr := $ffffffff }
  352. { (left.location.size = location.size at this point) }
  353. cg.a_op_const_reg_reg(current_asmdata.CurrAsmList, OP_SUB, location.size, tcgsize2size[opsize]*8, left.location.register, hr);
  354. cg.a_op_const_reg(current_asmdata.CurrAsmList, OP_SAR, location.size, (tcgsize2size[opsize]*8)-1, hr);
  355. { if left > tcgsize2size[opsize]*8-1, then result := 0 else result := result of bit test }
  356. cg.a_op_reg_reg(current_asmdata.CurrAsmList, OP_AND, location.size, hr, location.register);
  357. end { of right.location.loc=LOC_CONSTANT }
  358. { do search in a normal set which could have >32 elements
  359. but also used if the left side contains higher values > 32 }
  360. else if (left.location.loc=LOC_CONSTANT) then
  361. begin
  362. if (left.location.value < setbase) or (((left.location.value-setbase) shr 3) >= right.resultdef.size) then
  363. {should be caught earlier }
  364. internalerror(2007020402);
  365. cg.a_bit_test_const_loc_reg(current_asmdata.CurrAsmList,location.size,left.location.value-setbase,
  366. right.location,location.register);
  367. end
  368. else
  369. begin
  370. location_force_reg(current_asmdata.CurrAsmList, left.location, opsize, true);
  371. pleftreg := left.location.register;
  372. if (opsize >= OS_S8) or { = if signed }
  373. ((left.resultdef.typ=orddef) and (torddef(left.resultdef).high > tsetdef(right.resultdef).setmax)) or
  374. ((left.resultdef.typ=enumdef) and (tenumdef(left.resultdef).max > tsetdef(right.resultdef).setmax)) then
  375. begin
  376. current_asmdata.getjumplabel(l);
  377. current_asmdata.getjumplabel(l2);
  378. needslabel := True;
  379. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, left.location.size, OC_BE, tsetdef(right.resultdef).setmax, pleftreg, l);
  380. cg.a_load_const_reg(current_asmdata.CurrAsmList, location.size, 0, location.register);
  381. cg.a_jmp_always(current_asmdata.CurrAsmList, l2);
  382. cg.a_label(current_asmdata.CurrAsmList, l);
  383. end;
  384. register_maybe_adjust_setbase(current_asmdata.CurrAsmList,left.location,setbase);
  385. pleftreg:=left.location.register;
  386. cg.a_bit_test_reg_loc_reg(current_asmdata.CurrAsmList,left.location.size,location.size,
  387. pleftreg,right.location,location.register);
  388. if needslabel then
  389. cg.a_label(current_asmdata.CurrAsmList, l2);
  390. end;
  391. end;
  392. end;
  393. location_freetemp(current_asmdata.CurrAsmList, right.location);
  394. location.size := def_cgsize(resultdef);
  395. location.register := cg.makeregsize(current_asmdata.CurrAsmList, location.register, location.size);
  396. end;
  397. {*****************************************************************************
  398. TCGCASENODE
  399. *****************************************************************************}
  400. function tcgcasenode.blocklabel(id:longint):tasmlabel;
  401. begin
  402. if not assigned(blocks[id]) then
  403. internalerror(200411301);
  404. result:=pcaseblock(blocks[id])^.blocklabel;
  405. end;
  406. procedure tcgcasenode.optimizevalues(var max_linear_list:aint;var max_dist:aword);
  407. begin
  408. { no changes by default }
  409. end;
  410. function tcgcasenode.has_jumptable : boolean;
  411. begin
  412. { No jumptable support in the default implementation }
  413. has_jumptable:=false;
  414. end;
  415. procedure tcgcasenode.genjumptable(hp : pcaselabel;min_,max_ : aint);
  416. begin
  417. internalerror(200209161);
  418. end;
  419. procedure tcgcasenode.genlinearlist(hp : pcaselabel);
  420. var
  421. first : boolean;
  422. last : TConstExprInt;
  423. scratch_reg: tregister;
  424. procedure genitem(t : pcaselabel);
  425. procedure gensub(value:aint);
  426. begin
  427. { here, since the sub and cmp are separate we need
  428. to move the result before subtract to help
  429. the register allocator
  430. }
  431. cg.a_load_reg_reg(current_asmdata.CurrAsmList, opsize, opsize, hregister, scratch_reg);
  432. cg.a_op_const_reg(current_asmdata.CurrAsmList, OP_SUB, opsize, value, hregister);
  433. end;
  434. begin
  435. if assigned(t^.less) then
  436. genitem(t^.less);
  437. { do we need to test the first value? }
  438. if first and (t^._low>get_min_value(left.resultdef)) then
  439. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList,opsize,jmp_lt,aint(t^._low.svalue),hregister,elselabel);
  440. if t^._low=t^._high then
  441. begin
  442. if t^._low-last=0 then
  443. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList,opsize,OC_EQ,0,hregister,blocklabel(t^.blockid))
  444. else
  445. begin
  446. gensub(aint(t^._low.svalue-last.svalue));
  447. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList,opsize,
  448. OC_EQ,aint(t^._low.svalue-last.svalue),scratch_reg,blocklabel(t^.blockid));
  449. end;
  450. last:=t^._low;
  451. end
  452. else
  453. begin
  454. { it begins with the smallest label, if the value }
  455. { is even smaller then jump immediately to the }
  456. { ELSE-label }
  457. if first then
  458. begin
  459. { have we to ajust the first value ? }
  460. if (t^._low>get_min_value(left.resultdef)) then
  461. gensub(aint(t^._low.svalue));
  462. end
  463. else
  464. begin
  465. { if there is no unused label between the last and the }
  466. { present label then the lower limit can be checked }
  467. { immediately. else check the range in between: }
  468. gensub(aint(t^._low.svalue-last.svalue));
  469. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, opsize,jmp_lt,aint(t^._low.svalue-last.svalue),scratch_reg,elselabel);
  470. end;
  471. gensub(aint(t^._high.svalue-t^._low.svalue));
  472. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList,opsize,jmp_le,aint(t^._high.svalue-t^._low.svalue),scratch_reg,blocklabel(t^.blockid));
  473. last:=t^._high;
  474. end;
  475. first:=false;
  476. if assigned(t^.greater) then
  477. genitem(t^.greater);
  478. end;
  479. begin
  480. { do we need to generate cmps? }
  481. if (with_sign and (min_label<0)) then
  482. genlinearcmplist(hp)
  483. else
  484. begin
  485. last:=0;
  486. first:=true;
  487. scratch_reg:=cg.getintregister(current_asmdata.CurrAsmList,opsize);
  488. genitem(hp);
  489. cg.a_jmp_always(current_asmdata.CurrAsmList,elselabel);
  490. end;
  491. end;
  492. procedure tcgcasenode.genlinearcmplist(hp : pcaselabel);
  493. var
  494. last : TConstExprInt;
  495. lastwasrange: boolean;
  496. procedure genitem(t : pcaselabel);
  497. {$ifndef cpu64bit}
  498. var
  499. l1 : tasmlabel;
  500. {$endif cpu64bit}
  501. begin
  502. if assigned(t^.less) then
  503. genitem(t^.less);
  504. if t^._low=t^._high then
  505. begin
  506. {$ifndef cpu64bit}
  507. if opsize in [OS_S64,OS_64] then
  508. begin
  509. current_asmdata.getjumplabel(l1);
  510. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, OS_32, OC_NE, aint(hi(int64(t^._low.svalue))),hregister2,l1);
  511. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, OS_32, OC_EQ, aint(lo(int64(t^._low.svalue))),hregister, blocklabel(t^.blockid));
  512. cg.a_label(current_asmdata.CurrAsmList,l1);
  513. end
  514. else
  515. {$endif cpu64bit}
  516. begin
  517. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, opsize, OC_EQ, aint(t^._low.svalue),hregister, blocklabel(t^.blockid));
  518. end;
  519. { Reset last here, because we've only checked for one value and need to compare
  520. for the next range both the lower and upper bound }
  521. lastwasrange := false;
  522. end
  523. else
  524. begin
  525. { it begins with the smallest label, if the value }
  526. { is even smaller then jump immediately to the }
  527. { ELSE-label }
  528. if not lastwasrange or (t^._low-last>1) then
  529. begin
  530. {$ifndef cpu64bit}
  531. if opsize in [OS_64,OS_S64] then
  532. begin
  533. current_asmdata.getjumplabel(l1);
  534. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, OS_32, jmp_lt, aint(hi(int64(t^._low.svalue))),
  535. hregister2, elselabel);
  536. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, OS_32, jmp_gt, aint(hi(int64(t^._low.svalue))),
  537. hregister2, l1);
  538. { the comparisation of the low dword must be always unsigned! }
  539. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, OS_32, OC_B, aint(lo(int64(t^._low.svalue))), hregister, elselabel);
  540. cg.a_label(current_asmdata.CurrAsmList,l1);
  541. end
  542. else
  543. {$endif cpu64bit}
  544. begin
  545. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, opsize, jmp_lt, aint(t^._low.svalue), hregister,
  546. elselabel);
  547. end;
  548. end;
  549. {$ifndef cpu64bit}
  550. if opsize in [OS_S64,OS_64] then
  551. begin
  552. current_asmdata.getjumplabel(l1);
  553. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, OS_32, jmp_lt, aint(hi(int64(t^._high.svalue))), hregister2,
  554. blocklabel(t^.blockid));
  555. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, OS_32, jmp_gt, aint(hi(int64(t^._high.svalue))), hregister2,
  556. l1);
  557. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, OS_32, OC_BE, aint(lo(int64(t^._high.svalue))), hregister, blocklabel(t^.blockid));
  558. cg.a_label(current_asmdata.CurrAsmList,l1);
  559. end
  560. else
  561. {$endif cpu64bit}
  562. begin
  563. cg.a_cmp_const_reg_label(current_asmdata.CurrAsmList, opsize, jmp_le, aint(t^._high.svalue), hregister, blocklabel(t^.blockid));
  564. end;
  565. last:=t^._high;
  566. lastwasrange := true;
  567. end;
  568. if assigned(t^.greater) then
  569. genitem(t^.greater);
  570. end;
  571. begin
  572. last:=0;
  573. lastwasrange:=false;
  574. genitem(hp);
  575. cg.a_jmp_always(current_asmdata.CurrAsmList,elselabel);
  576. end;
  577. procedure tcgcasenode.pass_generate_code;
  578. var
  579. oldflowcontrol: tflowcontrol;
  580. i : longint;
  581. lv,hv,
  582. max_label: tconstexprint;
  583. labelcnt : aint;
  584. max_linear_list : aint;
  585. otl, ofl: tasmlabel;
  586. isjump : boolean;
  587. max_dist,
  588. dist : aword;
  589. begin
  590. location_reset(location,LOC_VOID,OS_NO);
  591. oldflowcontrol := flowcontrol;
  592. include(flowcontrol,fc_inflowcontrol);
  593. { Allocate labels }
  594. current_asmdata.getjumplabel(endlabel);
  595. current_asmdata.getjumplabel(elselabel);
  596. for i:=0 to blocks.count-1 do
  597. current_asmdata.getjumplabel(pcaseblock(blocks[i])^.blocklabel);
  598. with_sign:=is_signed(left.resultdef);
  599. if with_sign then
  600. begin
  601. jmp_gt:=OC_GT;
  602. jmp_lt:=OC_LT;
  603. jmp_le:=OC_LTE;
  604. end
  605. else
  606. begin
  607. jmp_gt:=OC_A;
  608. jmp_lt:=OC_B;
  609. jmp_le:=OC_BE;
  610. end;
  611. { save current current_procinfo.CurrTrueLabel and current_procinfo.CurrFalseLabel }
  612. isjump:=false;
  613. if left.expectloc=LOC_JUMP then
  614. begin
  615. otl:=current_procinfo.CurrTrueLabel;
  616. current_asmdata.getjumplabel(current_procinfo.CurrTrueLabel);
  617. ofl:=current_procinfo.CurrFalseLabel;
  618. current_asmdata.getjumplabel(current_procinfo.CurrFalseLabel);
  619. isjump:=true;
  620. end;
  621. secondpass(left);
  622. { determines the size of the operand }
  623. opsize:=def_cgsize(left.resultdef);
  624. { copy the case expression to a register }
  625. location_force_reg(current_asmdata.CurrAsmList,left.location,opsize,false);
  626. {$ifndef cpu64bit}
  627. if opsize in [OS_S64,OS_64] then
  628. begin
  629. hregister:=left.location.register64.reglo;
  630. hregister2:=left.location.register64.reghi;
  631. end
  632. else
  633. {$endif cpu64bit}
  634. hregister:=left.location.register;
  635. if isjump then
  636. begin
  637. current_procinfo.CurrTrueLabel:=otl;
  638. current_procinfo.CurrFalseLabel:=ofl;
  639. end
  640. else
  641. if (left.location.loc=LOC_JUMP) then
  642. internalerror(2006050501);
  643. { we need the min_label always to choose between }
  644. { cmps and subs/decs }
  645. min_label:=case_get_min(labels);
  646. { Generate the jumps }
  647. {$ifdef OLDREGVARS}
  648. load_all_regvars(current_asmdata.CurrAsmList);
  649. {$endif OLDREGVARS}
  650. {$ifndef cpu64bit}
  651. if opsize in [OS_64,OS_S64] then
  652. genlinearcmplist(labels)
  653. else
  654. {$endif cpu64bit}
  655. begin
  656. if cs_opt_level1 in current_settings.optimizerswitches then
  657. begin
  658. { procedures are empirically passed on }
  659. { consumption can also be calculated }
  660. { but does it pay on the different }
  661. { processors? }
  662. { moreover can the size only be appro- }
  663. { ximated as it is not known if rel8, }
  664. { rel16 or rel32 jumps are used }
  665. max_label:=case_get_max(labels);
  666. labelcnt:=case_count_labels(labels);
  667. { can we omit the range check of the jump table ? }
  668. getrange(left.resultdef,lv,hv);
  669. jumptable_no_range:=(lv=min_label) and (hv=max_label);
  670. { hack a little bit, because the range can be greater }
  671. { than the positive range of a aint }
  672. if (min_label<0) and (max_label>0) then
  673. begin
  674. if min_label=TConstExprInt(int64(low(aint))) then
  675. dist:=aword(max_label.uvalue)+aword(low(aint))
  676. else
  677. dist:=aword(max_label.uvalue)+aword(-min_label.svalue)
  678. end
  679. else
  680. dist:=max_label.uvalue-min_label.uvalue;
  681. { optimize for size ? }
  682. if cs_opt_size in current_settings.optimizerswitches then
  683. begin
  684. if has_jumptable and
  685. (min_label>=int64(low(aint))) and
  686. (max_label<=high(aint)) and
  687. not((labelcnt<=2) or
  688. ((max_label-min_label)<0) or
  689. ((max_label-min_label)>3*labelcnt)) then
  690. begin
  691. { if the labels less or more a continuum then }
  692. genjumptable(labels,min_label.svalue,max_label.svalue);
  693. end
  694. else
  695. begin
  696. { a linear list is always smaller than a jump tree }
  697. genlinearlist(labels);
  698. end;
  699. end
  700. else
  701. begin
  702. max_dist:=4*labelcnt;
  703. if jumptable_no_range then
  704. max_linear_list:=4
  705. else
  706. max_linear_list:=2;
  707. { allow processor specific values }
  708. optimizevalues(max_linear_list,max_dist);
  709. if (labelcnt<=max_linear_list) then
  710. genlinearlist(labels)
  711. else
  712. begin
  713. if (has_jumptable) and
  714. (dist<max_dist) and
  715. (min_label>=int64(low(aint))) and
  716. (max_label<=high(aint)) then
  717. genjumptable(labels,min_label.svalue,max_label.svalue)
  718. else
  719. genlinearlist(labels);
  720. end;
  721. end;
  722. end
  723. else
  724. { it's always not bad }
  725. genlinearlist(labels);
  726. end;
  727. { generate the instruction blocks }
  728. for i:=0 to blocks.count-1 do
  729. begin
  730. current_asmdata.CurrAsmList.concat(Tai_align_abstract.create(current_settings.alignment.jumpalign));
  731. cg.a_label(current_asmdata.CurrAsmList,pcaseblock(blocks[i])^.blocklabel);
  732. secondpass(pcaseblock(blocks[i])^.statement);
  733. { don't come back to case line }
  734. current_filepos:=current_asmdata.CurrAsmList.getlasttaifilepos^;
  735. {$ifdef OLDREGVARS}
  736. load_all_regvars(current_asmdata.CurrAsmList);
  737. {$endif OLDREGVARS}
  738. cg.a_jmp_always(current_asmdata.CurrAsmList,endlabel);
  739. end;
  740. current_asmdata.CurrAsmList.concat(Tai_align_abstract.create(current_settings.alignment.jumpalign));
  741. { ...and the else block }
  742. cg.a_label(current_asmdata.CurrAsmList,elselabel);
  743. if assigned(elseblock) then
  744. begin
  745. secondpass(elseblock);
  746. {$ifdef OLDREGVARS}
  747. load_all_regvars(current_asmdata.CurrAsmList);
  748. {$endif OLDREGVARS}
  749. end;
  750. current_asmdata.CurrAsmList.concat(Tai_align_abstract.create(current_settings.alignment.jumpalign));
  751. cg.a_label(current_asmdata.CurrAsmList,endlabel);
  752. { Reset labels }
  753. for i:=0 to blocks.count-1 do
  754. pcaseblock(blocks[i])^.blocklabel:=nil;
  755. flowcontrol := oldflowcontrol + (flowcontrol - [fc_inflowcontrol]);
  756. end;
  757. begin
  758. csetelementnode:=tcgsetelementnode;
  759. cinnode:=tcginnode;
  760. ccasenode:=tcgcasenode;
  761. end.