aoptcpu.pas 46 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034
  1. {
  2. Copyright (c) 1998-2002 by Jonas Maebe, member of the Free Pascal
  3. Development Team
  4. This unit implements the ARM optimizer object
  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 aoptcpu;
  19. {$i fpcdefs.inc}
  20. Interface
  21. uses cgbase, cpubase, aasmtai, aopt, aoptcpub, aoptobj;
  22. Type
  23. { TCpuAsmOptimizer }
  24. TCpuAsmOptimizer = class(TAsmOptimizer)
  25. { uses the same constructor as TAopObj }
  26. function PeepHoleOptPass1Cpu(var p: tai): boolean; override;
  27. procedure PeepHoleOptPass2;override;
  28. Function RegInInstruction(Reg: TRegister; p1: tai): Boolean;override;
  29. procedure RemoveSuperfluousMove(const p: tai; movp: tai; const optimizer: string);
  30. function RegUsedAfterInstruction(reg: Tregister; p: tai;
  31. var AllUsedRegs: TAllUsedRegs): Boolean;
  32. End;
  33. TCpuPreRegallocScheduler = class(TAsmOptimizer)
  34. function PeepHoleOptPass1Cpu(var p: tai): boolean;override;
  35. end;
  36. TCpuThumb2AsmOptimizer = class(TCpuAsmOptimizer)
  37. { uses the same constructor as TAopObj }
  38. procedure PeepHoleOptPass2;override;
  39. End;
  40. Implementation
  41. uses
  42. cutils,
  43. verbose,
  44. cgutils,
  45. aasmbase,aasmdata,aasmcpu;
  46. function CanBeCond(p : tai) : boolean;
  47. begin
  48. result:=
  49. (p.typ=ait_instruction) and
  50. (taicpu(p).condition=C_None) and
  51. (taicpu(p).opcode<>A_PLD) and
  52. ((taicpu(p).opcode<>A_BLX) or
  53. (taicpu(p).oper[0]^.typ=top_reg));
  54. end;
  55. function RefsEqual(const r1, r2: treference): boolean;
  56. begin
  57. refsequal :=
  58. (r1.offset = r2.offset) and
  59. (r1.base = r2.base) and
  60. (r1.index = r2.index) and (r1.scalefactor = r2.scalefactor) and
  61. (r1.symbol=r2.symbol) and (r1.refaddr = r2.refaddr) and
  62. (r1.relsymbol = r2.relsymbol) and
  63. (r1.signindex = r2.signindex) and
  64. (r1.shiftimm = r2.shiftimm) and
  65. (r1.addressmode = r2.addressmode) and
  66. (r1.shiftmode = r2.shiftmode);
  67. end;
  68. function MatchInstruction(const instr: tai; const op: TAsmOp; const cond: TAsmConds; const postfix: TOpPostfixes): boolean;
  69. begin
  70. result :=
  71. (instr.typ = ait_instruction) and
  72. (taicpu(instr).opcode = op) and
  73. ((cond = []) or (taicpu(instr).condition in cond)) and
  74. ((postfix = []) or (taicpu(instr).oppostfix in postfix));
  75. end;
  76. function MatchOperand(const oper1: TOper; const oper2: TOper): boolean; inline;
  77. begin
  78. result := oper1.typ = oper2.typ;
  79. if result then
  80. case oper1.typ of
  81. top_const:
  82. Result:=oper1.val = oper2.val;
  83. top_reg:
  84. Result:=oper1.reg = oper2.reg;
  85. top_conditioncode:
  86. Result:=oper1.cc = oper2.cc;
  87. top_ref:
  88. Result:=RefsEqual(oper1.ref^, oper2.ref^);
  89. else Result:=false;
  90. end
  91. end;
  92. function MatchOperand(const oper: TOper; const reg: TRegister): boolean; inline;
  93. begin
  94. result := (oper.typ = top_reg) and (oper.reg = reg);
  95. end;
  96. procedure RemoveRedundantMove(const cmpp: tai; movp: tai; asml: TAsmList);
  97. begin
  98. if (taicpu(movp).condition = C_EQ) and
  99. (taicpu(cmpp).oper[0]^.reg = taicpu(movp).oper[0]^.reg) and
  100. (taicpu(cmpp).oper[1]^.val = taicpu(movp).oper[1]^.val) then
  101. begin
  102. asml.insertafter(tai_comment.Create(strpnew('Peephole CmpMovMov - Removed redundant moveq')), movp);
  103. asml.remove(movp);
  104. movp.free;
  105. end;
  106. end;
  107. function regLoadedWithNewValue(reg: tregister; hp: tai): boolean;
  108. var
  109. p: taicpu;
  110. begin
  111. p := taicpu(hp);
  112. regLoadedWithNewValue := false;
  113. if not ((assigned(hp)) and (hp.typ = ait_instruction)) then
  114. exit;
  115. case p.opcode of
  116. { These operands do not write into a register at all }
  117. A_CMP, A_CMN, A_TST, A_TEQ, A_B, A_BL, A_BX, A_BLX, A_SWI, A_MSR, A_PLD:
  118. exit;
  119. {Take care of post/preincremented store and loads, they will change their base register}
  120. A_STR, A_LDR:
  121. regLoadedWithNewValue :=
  122. (taicpu(p).oper[1]^.typ=top_ref) and
  123. (taicpu(p).oper[1]^.ref^.addressmode in [AM_PREINDEXED,AM_POSTINDEXED]) and
  124. (taicpu(p).oper[1]^.ref^.base = reg);
  125. { These four are writing into the first 2 register, UMLAL and SMLAL will also read from them }
  126. A_UMLAL, A_UMULL, A_SMLAL, A_SMULL:
  127. regLoadedWithNewValue :=
  128. (p.oper[1]^.typ = top_reg) and
  129. (p.oper[1]^.reg = reg);
  130. {Loads to oper2 from coprocessor}
  131. {
  132. MCR/MRC is currently not supported in FPC
  133. A_MRC:
  134. regLoadedWithNewValue :=
  135. (p.oper[2]^.typ = top_reg) and
  136. (p.oper[2]^.reg = reg);
  137. }
  138. {Loads to all register in the registerset}
  139. A_LDM:
  140. regLoadedWithNewValue := (getsupreg(reg) in p.oper[1]^.regset^);
  141. end;
  142. if regLoadedWithNewValue then
  143. exit;
  144. case p.oper[0]^.typ of
  145. {This is the case}
  146. top_reg:
  147. regLoadedWithNewValue := (p.oper[0]^.reg = reg);
  148. {LDM/STM might write a new value to their index register}
  149. top_ref:
  150. regLoadedWithNewValue :=
  151. (taicpu(p).oper[0]^.ref^.addressmode in [AM_PREINDEXED,AM_POSTINDEXED]) and
  152. (taicpu(p).oper[0]^.ref^.base = reg);
  153. end;
  154. end;
  155. function instructionLoadsFromReg(const reg: TRegister; const hp: tai): boolean;
  156. var
  157. p: taicpu;
  158. i: longint;
  159. begin
  160. instructionLoadsFromReg := false;
  161. if not (assigned(hp) and (hp.typ = ait_instruction)) then
  162. exit;
  163. p:=taicpu(hp);
  164. i:=1;
  165. {For these instructions we have to start on oper[0]}
  166. if (p.opcode in [A_STR, A_LDM, A_STM, A_PLD,
  167. A_CMP, A_CMN, A_TST, A_TEQ,
  168. A_B, A_BL, A_BX, A_BLX,
  169. A_SMLAL, A_UMLAL]) then i:=0;
  170. while(i<p.ops) do
  171. begin
  172. case p.oper[I]^.typ of
  173. top_reg:
  174. instructionLoadsFromReg := p.oper[I]^.reg = reg;
  175. top_regset:
  176. instructionLoadsFromReg := (getsupreg(reg) in p.oper[I]^.regset^);
  177. top_shifterop:
  178. instructionLoadsFromReg := p.oper[I]^.shifterop^.rs = reg;
  179. top_ref:
  180. instructionLoadsFromReg :=
  181. (p.oper[I]^.ref^.base = reg) or
  182. (p.oper[I]^.ref^.index = reg);
  183. end;
  184. if instructionLoadsFromReg then exit; {Bailout if we found something}
  185. Inc(I);
  186. end;
  187. end;
  188. function TCpuAsmOptimizer.RegUsedAfterInstruction(reg: Tregister; p: tai;
  189. var AllUsedRegs: TAllUsedRegs): Boolean;
  190. begin
  191. AllUsedRegs[getregtype(reg)].Update(tai(p.Next));
  192. RegUsedAfterInstruction :=
  193. AllUsedRegs[getregtype(reg)].IsUsed(reg) and
  194. not(regLoadedWithNewValue(reg,p)) and
  195. (
  196. not(GetNextInstruction(p,p)) or
  197. instructionLoadsFromReg(reg,p) or
  198. not(regLoadedWithNewValue(reg,p))
  199. );
  200. end;
  201. procedure TCpuAsmOptimizer.RemoveSuperfluousMove(const p: tai; movp: tai; const optimizer: string);
  202. var
  203. TmpUsedRegs: TAllUsedRegs;
  204. begin
  205. if MatchInstruction(movp, A_MOV, [taicpu(p).condition], [PF_None]) and
  206. (taicpu(movp).ops=2) and {We can't optimize if there is a shiftop}
  207. MatchOperand(taicpu(movp).oper[1]^, taicpu(p).oper[0]^.reg) and
  208. {There is a special requirement for MUL and MLA, oper[0] and oper[1] are not allowed to be the same}
  209. not (
  210. (taicpu(p).opcode in [A_MLA, A_MUL]) and
  211. (taicpu(p).oper[1]^.reg = taicpu(movp).oper[0]^.reg)
  212. ) then
  213. begin
  214. CopyUsedRegs(TmpUsedRegs);
  215. UpdateUsedRegs(TmpUsedRegs, tai(p.next));
  216. if not(RegUsedAfterInstruction(taicpu(p).oper[0]^.reg,movp,TmpUsedRegs)) then
  217. begin
  218. asml.insertbefore(tai_comment.Create(strpnew('Peephole '+optimizer+' removed superfluous mov')), movp);
  219. taicpu(p).loadreg(0,taicpu(movp).oper[0]^.reg);
  220. asml.remove(movp);
  221. movp.free;
  222. end;
  223. ReleaseUsedRegs(TmpUsedRegs);
  224. end;
  225. end;
  226. function TCpuAsmOptimizer.PeepHoleOptPass1Cpu(var p: tai): boolean;
  227. var
  228. hp1,hp2: tai;
  229. i: longint;
  230. TmpUsedRegs: TAllUsedRegs;
  231. tempop: tasmop;
  232. function IsPowerOf2(const value: DWord): boolean; inline;
  233. begin
  234. Result:=(value and (value - 1)) = 0;
  235. end;
  236. begin
  237. result := false;
  238. case p.typ of
  239. ait_instruction:
  240. begin
  241. (* optimization proved not to be safe, see tw4768.pp
  242. {
  243. change
  244. <op> reg,x,y
  245. cmp reg,#0
  246. into
  247. <op>s reg,x,y
  248. }
  249. { this optimization can applied only to the currently enabled operations because
  250. the other operations do not update all flags and FPC does not track flag usage }
  251. if (taicpu(p).opcode in [A_ADC,A_ADD,A_SUB {A_UDIV,A_SDIV,A_MUL,A_MVN,A_MOV,A_ORR,A_EOR,A_AND}]) and
  252. (taicpu(p).oppostfix = PF_None) and
  253. (taicpu(p).condition = C_None) and
  254. GetNextInstruction(p, hp1) and
  255. MatchInstruction(hp1, A_CMP, [C_None], [PF_None]) and
  256. (taicpu(hp1).oper[1]^.typ = top_const) and
  257. (taicpu(p).oper[0]^.reg = taicpu(hp1).oper[0]^.reg) and
  258. (taicpu(hp1).oper[1]^.val = 0) { and
  259. GetNextInstruction(hp1, hp2) and
  260. (tai(hp2).typ = ait_instruction) and
  261. // be careful here, following instructions could use other flags
  262. // however after a jump fpc never depends on the value of flags
  263. (taicpu(hp2).opcode = A_B) and
  264. (taicpu(hp2).condition in [C_EQ,C_NE,C_MI,C_PL])} then
  265. begin
  266. taicpu(p).oppostfix:=PF_S;
  267. asml.remove(hp1);
  268. hp1.free;
  269. end
  270. else
  271. *)
  272. case taicpu(p).opcode of
  273. A_STR:
  274. begin
  275. { change
  276. str reg1,ref
  277. ldr reg2,ref
  278. into
  279. str reg1,ref
  280. mov reg2,reg1
  281. }
  282. if (taicpu(p).oper[1]^.ref^.addressmode=AM_OFFSET) and
  283. (taicpu(p).oppostfix=PF_None) and
  284. GetNextInstruction(p,hp1) and
  285. MatchInstruction(hp1, A_LDR, [taicpu(p).condition, C_None], [PF_None]) and
  286. RefsEqual(taicpu(p).oper[1]^.ref^,taicpu(hp1).oper[1]^.ref^) and
  287. (taicpu(hp1).oper[1]^.ref^.addressmode=AM_OFFSET) then
  288. begin
  289. if taicpu(hp1).oper[0]^.reg=taicpu(p).oper[0]^.reg then
  290. begin
  291. asml.insertbefore(tai_comment.Create(strpnew('Peephole StrLdr2StrMov 1 done')), hp1);
  292. asml.remove(hp1);
  293. hp1.free;
  294. end
  295. else
  296. begin
  297. taicpu(hp1).opcode:=A_MOV;
  298. taicpu(hp1).oppostfix:=PF_None;
  299. taicpu(hp1).loadreg(1,taicpu(p).oper[0]^.reg);
  300. asml.insertbefore(tai_comment.Create(strpnew('Peephole StrLdr2StrMov 2 done')), hp1);
  301. end;
  302. result := true;
  303. end;
  304. end;
  305. A_LDR:
  306. begin
  307. { change
  308. ldr reg1,ref
  309. ldr reg2,ref
  310. into
  311. ldr reg1,ref
  312. mov reg2,reg1
  313. }
  314. if (taicpu(p).oper[1]^.ref^.addressmode=AM_OFFSET) and
  315. GetNextInstruction(p,hp1) and
  316. MatchInstruction(hp1, A_LDR, [taicpu(p).condition, C_None], [taicpu(p).oppostfix]) and
  317. RefsEqual(taicpu(p).oper[1]^.ref^,taicpu(hp1).oper[1]^.ref^) and
  318. (taicpu(p).oper[0]^.reg<>taicpu(hp1).oper[1]^.ref^.index) and
  319. (taicpu(p).oper[0]^.reg<>taicpu(hp1).oper[1]^.ref^.base) and
  320. (taicpu(hp1).oper[1]^.ref^.addressmode=AM_OFFSET) then
  321. begin
  322. if taicpu(hp1).oper[0]^.reg=taicpu(p).oper[0]^.reg then
  323. begin
  324. asml.insertbefore(tai_comment.Create(strpnew('Peephole LdrLdr2Ldr done')), hp1);
  325. asml.remove(hp1);
  326. hp1.free;
  327. end
  328. else
  329. begin
  330. asml.insertbefore(tai_comment.Create(strpnew('Peephole LdrLdr2LdrMov done')), hp1);
  331. taicpu(hp1).opcode:=A_MOV;
  332. taicpu(hp1).oppostfix:=PF_None;
  333. taicpu(hp1).loadreg(1,taicpu(p).oper[0]^.reg);
  334. end;
  335. result := true;
  336. end;
  337. { Remove superfluous mov after ldr
  338. changes
  339. ldr reg1, ref
  340. mov reg2, reg1
  341. to
  342. ldr reg2, ref
  343. conditions are:
  344. * reg1 must be released after mov
  345. * mov can not contain shifterops
  346. * ldr+mov have the same conditions
  347. * mov does not set flags
  348. }
  349. if GetNextInstruction(p, hp1) then
  350. RemoveSuperfluousMove(p, hp1, 'LdrMov2Ldr');
  351. end;
  352. A_MOV:
  353. begin
  354. { fold
  355. mov reg1,reg0, shift imm1
  356. mov reg1,reg1, shift imm2
  357. to
  358. mov reg1,reg0, shift imm1+imm2
  359. }
  360. if (taicpu(p).ops=3) and
  361. (taicpu(p).oper[2]^.typ = top_shifterop) and
  362. (taicpu(p).oper[2]^.shifterop^.rs = NR_NO) and
  363. getnextinstruction(p,hp1) and
  364. MatchInstruction(hp1, A_MOV, [taicpu(p).condition], [PF_None]) and
  365. (taicpu(hp1).ops=3) and
  366. MatchOperand(taicpu(hp1).oper[0]^, taicpu(p).oper[0]^.reg) and
  367. MatchOperand(taicpu(hp1).oper[1]^, taicpu(p).oper[0]^.reg) and
  368. (taicpu(hp1).oper[2]^.typ = top_shifterop) and
  369. (taicpu(hp1).oper[2]^.shifterop^.rs = NR_NO) then
  370. begin
  371. { fold
  372. mov reg1,reg0, lsl 16
  373. mov reg1,reg1, lsr 16
  374. strh reg1, ...
  375. dealloc reg1
  376. to
  377. strh reg1, ...
  378. dealloc reg1
  379. }
  380. if (taicpu(p).oper[2]^.shifterop^.shiftmode=SM_LSL) and
  381. (taicpu(p).oper[2]^.shifterop^.shiftimm=16) and
  382. (taicpu(hp1).oper[2]^.shifterop^.shiftmode in [SM_LSR,SM_ASR]) and
  383. (taicpu(hp1).oper[2]^.shifterop^.shiftimm=16) and
  384. getnextinstruction(hp1,hp2) and
  385. MatchInstruction(hp2, A_STR, [taicpu(p).condition], [PF_H]) and
  386. MatchOperand(taicpu(hp2).oper[0]^, taicpu(p).oper[0]^.reg) then
  387. begin
  388. CopyUsedRegs(TmpUsedRegs);
  389. UpdateUsedRegs(TmpUsedRegs, tai(p.next));
  390. UpdateUsedRegs(TmpUsedRegs, tai(hp1.next));
  391. if not(RegUsedAfterInstruction(taicpu(p).oper[0]^.reg,hp2,TmpUsedRegs)) then
  392. begin
  393. asml.insertbefore(tai_comment.Create(strpnew('Peephole optimizer removed superfluous 16 Bit zero extension')), hp1);
  394. taicpu(hp2).loadreg(0,taicpu(p).oper[1]^.reg);
  395. asml.remove(p);
  396. asml.remove(hp1);
  397. p.free;
  398. hp1.free;
  399. p:=hp2;
  400. end;
  401. ReleaseUsedRegs(TmpUsedRegs);
  402. end
  403. { fold
  404. mov reg1,reg0, shift imm1
  405. mov reg1,reg1, shift imm2
  406. to
  407. mov reg1,reg0, shift imm1+imm2
  408. }
  409. else if (taicpu(p).oper[2]^.shifterop^.shiftmode=taicpu(hp1).oper[2]^.shifterop^.shiftmode) then
  410. begin
  411. inc(taicpu(p).oper[2]^.shifterop^.shiftimm,taicpu(hp1).oper[2]^.shifterop^.shiftimm);
  412. { avoid overflows }
  413. if taicpu(p).oper[2]^.shifterop^.shiftimm>31 then
  414. case taicpu(p).oper[2]^.shifterop^.shiftmode of
  415. SM_ROR:
  416. taicpu(p).oper[2]^.shifterop^.shiftimm:=taicpu(p).oper[2]^.shifterop^.shiftimm and 31;
  417. SM_ASR:
  418. taicpu(p).oper[2]^.shifterop^.shiftimm:=31;
  419. SM_LSR,
  420. SM_LSL:
  421. begin
  422. hp1:=taicpu.op_reg_const(A_MOV,taicpu(p).oper[0]^.reg,0);
  423. InsertLLItem(p.previous, p.next, hp1);
  424. p.free;
  425. p:=hp1;
  426. end;
  427. else
  428. internalerror(2008072803);
  429. end;
  430. asml.insertbefore(tai_comment.Create(strpnew('Peephole ShiftShift2Shift done')), p);
  431. asml.remove(hp1);
  432. hp1.free;
  433. result := true;
  434. end;
  435. end;
  436. { Change the common
  437. mov r0, r0, lsr #24
  438. and r0, r0, #255
  439. and remove the superfluous and
  440. This could be extended to handle more cases.
  441. }
  442. if (taicpu(p).ops=3) and
  443. (taicpu(p).oper[2]^.typ = top_shifterop) and
  444. (taicpu(p).oper[2]^.shifterop^.rs = NR_NO) and
  445. (taicpu(p).oper[2]^.shifterop^.shiftmode = SM_LSR) and
  446. (taicpu(p).oper[2]^.shifterop^.shiftimm >= 24 ) and
  447. getnextinstruction(p,hp1) and
  448. MatchInstruction(hp1, A_AND, [taicpu(p).condition], [taicpu(p).oppostfix]) and
  449. (taicpu(hp1).ops=3) and
  450. MatchOperand(taicpu(p).oper[0]^, taicpu(hp1).oper[0]^) and
  451. MatchOperand(taicpu(p).oper[0]^, taicpu(hp1).oper[1]^) and
  452. (taicpu(hp1).oper[2]^.typ = top_const) and
  453. { Check if the AND actually would only mask out bits beeing already zero because of the shift
  454. For LSR #25 and an AndConst of 255 that whould go like this:
  455. 255 and ((2 shl (32-25))-1)
  456. which results in 127, which is one less a power-of-2, meaning all lower bits are set.
  457. LSR #25 and AndConst of 254:
  458. 254 and ((2 shl (32-25))-1) = 126 -> lowest bit is clear, so we can't remove it.
  459. }
  460. ispowerof2((taicpu(hp1).oper[2]^.val and ((2 shl (32-taicpu(p).oper[2]^.shifterop^.shiftimm))-1))+1) then
  461. begin
  462. asml.insertbefore(tai_comment.Create(strpnew('Peephole LsrAnd2Lsr done')), hp1);
  463. asml.remove(hp1);
  464. hp1.free;
  465. end;
  466. {
  467. This changes the very common
  468. mov r0, #0
  469. str r0, [...]
  470. mov r0, #0
  471. str r0, [...]
  472. and removes all superfluous mov instructions
  473. }
  474. if (taicpu(p).ops = 2) and
  475. (taicpu(p).oper[1]^.typ = top_const) and
  476. GetNextInstruction(p,hp1) then
  477. begin
  478. while MatchInstruction(hp1, A_STR, [taicpu(p).condition], []) and
  479. MatchOperand(taicpu(p).oper[0]^, taicpu(hp1).oper[0]^) and
  480. GetNextInstruction(hp1, hp2) and
  481. MatchInstruction(hp2, A_MOV, [taicpu(p).condition], [PF_None]) and
  482. (taicpu(hp2).ops = 2) and
  483. MatchOperand(taicpu(hp2).oper[0]^, taicpu(p).oper[0]^) and
  484. MatchOperand(taicpu(hp2).oper[1]^, taicpu(p).oper[1]^) do
  485. begin
  486. asml.insertbefore(tai_comment.Create(strpnew('Peephole MovStrMov done')), hp2);
  487. GetNextInstruction(hp2,hp1);
  488. asml.remove(hp2);
  489. hp2.free;
  490. if not assigned(hp1) then break;
  491. end;
  492. end;
  493. {
  494. change
  495. mov r1, r0
  496. add r1, r1, #1
  497. to
  498. add r1, r0, #1
  499. Todo: Make it work for mov+cmp too
  500. CAUTION! If this one is successful p might not be a mov instruction anymore!
  501. }
  502. if (taicpu(p).ops = 2) and
  503. (taicpu(p).oper[1]^.typ = top_reg) and
  504. (taicpu(p).oppostfix = PF_NONE) and
  505. GetNextInstruction(p, hp1) and
  506. (tai(hp1).typ = ait_instruction) and
  507. (taicpu(hp1).opcode in [A_ADD, A_ADC, A_RSB, A_RSC, A_SUB, A_SBC,
  508. A_AND, A_BIC, A_EOR, A_ORR, A_MOV, A_MVN]) and
  509. {MOV and MVN might only have 2 ops}
  510. (taicpu(hp1).ops = 3) and
  511. (taicpu(hp1).condition in [C_NONE, taicpu(hp1).condition]) and
  512. MatchOperand(taicpu(p).oper[0]^, taicpu(hp1).oper[0]^.reg) and
  513. (taicpu(hp1).oper[1]^.typ = top_reg) and
  514. (taicpu(hp1).oper[2]^.typ in [top_reg, top_const, top_shifterop]) then
  515. begin
  516. { When we get here we still don't know if the registers match}
  517. for I:=1 to 2 do
  518. {
  519. If the first loop was successful p will be replaced with hp1.
  520. The checks will still be ok, because all required information
  521. will also be in hp1 then.
  522. }
  523. if MatchOperand(taicpu(p).oper[0]^, taicpu(hp1).oper[I]^.reg) then
  524. begin
  525. asml.insertbefore(tai_comment.Create(strpnew('Peephole RedundantMovProcess done')), hp1);
  526. taicpu(hp1).oper[I]^.reg := taicpu(p).oper[1]^.reg;
  527. if p<>hp1 then
  528. begin
  529. asml.remove(p);
  530. p.free;
  531. p:=hp1;
  532. end;
  533. end;
  534. end;
  535. { This folds shifterops into following instructions
  536. mov r0, r1, lsl #8
  537. add r2, r3, r0
  538. to
  539. add r2, r3, r1, lsl #8
  540. CAUTION! If this one is successful p might not be a mov instruction anymore!
  541. }
  542. if (taicpu(p).opcode = A_MOV) and
  543. (taicpu(p).ops = 3) and
  544. (taicpu(p).oper[1]^.typ = top_reg) and
  545. (taicpu(p).oper[2]^.typ = top_shifterop) and
  546. (taicpu(p).oppostfix = PF_NONE) and
  547. GetNextInstruction(p, hp1) and
  548. (tai(hp1).typ = ait_instruction) and
  549. (taicpu(hp1).ops = 3) and {Currently we can't fold into another shifterop}
  550. (taicpu(hp1).oper[2]^.typ = top_reg) and
  551. (taicpu(hp1).oppostfix = PF_NONE) and
  552. (taicpu(hp1).condition = taicpu(p).condition) and
  553. (taicpu(hp1).opcode in [A_ADD, A_ADC, A_RSB, A_RSC, A_SUB, A_SBC,
  554. A_AND, A_BIC, A_EOR, A_ORR, A_TEQ, A_TST]) and
  555. (
  556. {Only ONE of the two src operands is allowed to match}
  557. MatchOperand(taicpu(p).oper[0]^, taicpu(hp1).oper[1]^) xor
  558. MatchOperand(taicpu(p).oper[0]^, taicpu(hp1).oper[2]^)
  559. ) then
  560. begin
  561. CopyUsedRegs(TmpUsedRegs);
  562. UpdateUsedRegs(TmpUsedRegs, tai(p.next));
  563. if not(RegUsedAfterInstruction(taicpu(p).oper[0]^.reg,hp1,TmpUsedRegs)) then
  564. for I:=1 to 2 do
  565. if MatchOperand(taicpu(p).oper[0]^, taicpu(hp1).oper[I]^.reg) then
  566. begin
  567. if I = 1 then
  568. begin
  569. {The SUB operators need to be changed when we swap parameters}
  570. case taicpu(hp1).opcode of
  571. A_SUB: tempop:=A_RSB;
  572. A_SBC: tempop:=A_RSC;
  573. A_RSB: tempop:=A_SUB;
  574. A_RSC: tempop:=A_SBC;
  575. else tempop:=taicpu(hp1).opcode;
  576. end;
  577. hp2:=taicpu.op_reg_reg_reg_shifterop(tempop,
  578. taicpu(hp1).oper[0]^.reg, taicpu(hp1).oper[2]^.reg,
  579. taicpu(p).oper[1]^.reg, taicpu(p).oper[2]^.shifterop^);
  580. end
  581. else
  582. hp2:=taicpu.op_reg_reg_reg_shifterop(taicpu(hp1).opcode,
  583. taicpu(hp1).oper[0]^.reg, taicpu(hp1).oper[1]^.reg,
  584. taicpu(p).oper[1]^.reg, taicpu(p).oper[2]^.shifterop^);
  585. asml.insertbefore(hp2, p);
  586. asml.remove(p);
  587. asml.remove(hp1);
  588. p.free;
  589. hp1.free;
  590. p:=hp2;
  591. GetNextInstruction(p,hp1);
  592. asml.insertbefore(tai_comment.Create(strpnew('Peephole FoldShiftProcess done')), p);
  593. break;
  594. end;
  595. ReleaseUsedRegs(TmpUsedRegs);
  596. end;
  597. {
  598. Often we see shifts and then a superfluous mov to another register
  599. In the future this might be handled in RedundantMovProcess when it uses RegisterTracking
  600. }
  601. if (taicpu(p).opcode = A_MOV) and
  602. GetNextInstruction(p, hp1) then
  603. RemoveSuperfluousMove(p, hp1, 'MovMov2Mov');
  604. end;
  605. A_ADD,
  606. A_ADC,
  607. A_RSB,
  608. A_RSC,
  609. A_SUB,
  610. A_SBC,
  611. A_AND,
  612. A_BIC,
  613. A_EOR,
  614. A_ORR,
  615. A_MLA,
  616. A_MUL:
  617. begin
  618. {
  619. change
  620. and reg2,reg1,const1
  621. and reg2,reg2,const2
  622. to
  623. and reg2,reg1,(const1 and const2)
  624. }
  625. if (taicpu(p).opcode = A_AND) and
  626. (taicpu(p).oper[1]^.typ = top_reg) and
  627. (taicpu(p).oper[2]^.typ = top_const) and
  628. GetNextInstruction(p, hp1) and
  629. MatchInstruction(hp1, A_AND, [taicpu(p).condition], [PF_None]) and
  630. MatchOperand(taicpu(hp1).oper[0]^, taicpu(p).oper[0]^.reg) and
  631. MatchOperand(taicpu(hp1).oper[1]^, taicpu(p).oper[0]^.reg) and
  632. (taicpu(hp1).oper[2]^.typ = top_const) then
  633. begin
  634. asml.insertbefore(tai_comment.Create(strpnew('Peephole AndAnd2And done')), p);
  635. taicpu(p).loadConst(2,taicpu(p).oper[2]^.val and taicpu(hp1).oper[2]^.val);
  636. taicpu(p).oppostfix:=taicpu(hp1).oppostfix;
  637. asml.remove(hp1);
  638. hp1.free;
  639. end;
  640. {
  641. change
  642. add reg1, ...
  643. mov reg2, reg1
  644. to
  645. add reg2, ...
  646. }
  647. if GetNextInstruction(p, hp1) then
  648. RemoveSuperfluousMove(p, hp1, 'DataMov2Data');
  649. end;
  650. A_CMP:
  651. begin
  652. {
  653. change
  654. cmp reg,const1
  655. moveq reg,const1
  656. movne reg,const2
  657. to
  658. cmp reg,const1
  659. movne reg,const2
  660. }
  661. if (taicpu(p).oper[1]^.typ = top_const) and
  662. GetNextInstruction(p, hp1) and
  663. MatchInstruction(hp1, A_MOV, [C_EQ, C_NE], [PF_NONE]) and
  664. (taicpu(hp1).oper[1]^.typ = top_const) and
  665. GetNextInstruction(hp1, hp2) and
  666. MatchInstruction(hp2, A_MOV, [C_EQ, C_NE], [PF_NONE]) and
  667. (taicpu(hp1).oper[1]^.typ = top_const) then
  668. begin
  669. RemoveRedundantMove(p, hp1, asml);
  670. RemoveRedundantMove(p, hp2, asml);
  671. end;
  672. end;
  673. end;
  674. end;
  675. end;
  676. end;
  677. { instructions modifying the CPSR can be only the last instruction }
  678. function MustBeLast(p : tai) : boolean;
  679. begin
  680. Result:=(p.typ=ait_instruction) and
  681. ((taicpu(p).opcode in [A_BL,A_BLX,A_CMP,A_CMN,A_SWI,A_TEQ,A_TST,A_CMF,A_CMFE {,A_MSR}]) or
  682. ((taicpu(p).ops>=1) and (taicpu(p).oper[0]^.typ=top_reg) and (taicpu(p).oper[0]^.reg=NR_PC)) or
  683. (taicpu(p).oppostfix=PF_S));
  684. end;
  685. procedure TCpuAsmOptimizer.PeepHoleOptPass2;
  686. var
  687. p,hp1,hp2: tai;
  688. l : longint;
  689. condition : tasmcond;
  690. hp3: tai;
  691. WasLast: boolean;
  692. { UsedRegs, TmpUsedRegs: TRegSet; }
  693. begin
  694. p := BlockStart;
  695. { UsedRegs := []; }
  696. while (p <> BlockEnd) Do
  697. begin
  698. { UpdateUsedRegs(UsedRegs, tai(p.next)); }
  699. case p.Typ Of
  700. Ait_Instruction:
  701. begin
  702. case taicpu(p).opcode Of
  703. A_B:
  704. if taicpu(p).condition<>C_None then
  705. begin
  706. { check for
  707. Bxx xxx
  708. <several instructions>
  709. xxx:
  710. }
  711. l:=0;
  712. WasLast:=False;
  713. GetNextInstruction(p, hp1);
  714. while assigned(hp1) and
  715. (l<=4) and
  716. CanBeCond(hp1) and
  717. { stop on labels }
  718. not(hp1.typ=ait_label) do
  719. begin
  720. inc(l);
  721. if MustBeLast(hp1) then
  722. begin
  723. WasLast:=True;
  724. GetNextInstruction(hp1,hp1);
  725. break;
  726. end
  727. else
  728. GetNextInstruction(hp1,hp1);
  729. end;
  730. if assigned(hp1) then
  731. begin
  732. if FindLabel(tasmlabel(taicpu(p).oper[0]^.ref^.symbol),hp1) then
  733. begin
  734. if (l<=4) and (l>0) then
  735. begin
  736. condition:=inverse_cond(taicpu(p).condition);
  737. hp2:=p;
  738. GetNextInstruction(p,hp1);
  739. p:=hp1;
  740. repeat
  741. if hp1.typ=ait_instruction then
  742. taicpu(hp1).condition:=condition;
  743. if MustBeLast(hp1) then
  744. begin
  745. GetNextInstruction(hp1,hp1);
  746. break;
  747. end
  748. else
  749. GetNextInstruction(hp1,hp1);
  750. until not(assigned(hp1)) or
  751. not(CanBeCond(hp1)) or
  752. (hp1.typ=ait_label);
  753. { wait with removing else GetNextInstruction could
  754. ignore the label if it was the only usage in the
  755. jump moved away }
  756. tasmlabel(taicpu(hp2).oper[0]^.ref^.symbol).decrefs;
  757. asml.remove(hp2);
  758. hp2.free;
  759. continue;
  760. end;
  761. end
  762. else
  763. { do not perform further optimizations if there is inctructon
  764. in block #1 which can not be optimized.
  765. }
  766. if not WasLast then
  767. begin
  768. { check further for
  769. Bcc xxx
  770. <several instructions 1>
  771. B yyy
  772. xxx:
  773. <several instructions 2>
  774. yyy:
  775. }
  776. { hp2 points to jmp yyy }
  777. hp2:=hp1;
  778. { skip hp1 to xxx }
  779. GetNextInstruction(hp1, hp1);
  780. if assigned(hp2) and
  781. assigned(hp1) and
  782. (l<=3) and
  783. (hp2.typ=ait_instruction) and
  784. (taicpu(hp2).is_jmp) and
  785. (taicpu(hp2).condition=C_None) and
  786. { real label and jump, no further references to the
  787. label are allowed }
  788. (tasmlabel(taicpu(p).oper[0]^.ref^.symbol).getrefs=2) and
  789. FindLabel(tasmlabel(taicpu(p).oper[0]^.ref^.symbol),hp1) then
  790. begin
  791. l:=0;
  792. { skip hp1 to <several moves 2> }
  793. GetNextInstruction(hp1, hp1);
  794. while assigned(hp1) and
  795. CanBeCond(hp1) do
  796. begin
  797. inc(l);
  798. GetNextInstruction(hp1, hp1);
  799. end;
  800. { hp1 points to yyy: }
  801. if assigned(hp1) and
  802. FindLabel(tasmlabel(taicpu(hp2).oper[0]^.ref^.symbol),hp1) then
  803. begin
  804. condition:=inverse_cond(taicpu(p).condition);
  805. GetNextInstruction(p,hp1);
  806. hp3:=p;
  807. p:=hp1;
  808. repeat
  809. if hp1.typ=ait_instruction then
  810. taicpu(hp1).condition:=condition;
  811. GetNextInstruction(hp1,hp1);
  812. until not(assigned(hp1)) or
  813. not(CanBeCond(hp1));
  814. { hp2 is still at jmp yyy }
  815. GetNextInstruction(hp2,hp1);
  816. { hp2 is now at xxx: }
  817. condition:=inverse_cond(condition);
  818. GetNextInstruction(hp1,hp1);
  819. { hp1 is now at <several movs 2> }
  820. repeat
  821. taicpu(hp1).condition:=condition;
  822. GetNextInstruction(hp1,hp1);
  823. until not(assigned(hp1)) or
  824. not(CanBeCond(hp1)) or
  825. (hp1.typ=ait_label);
  826. {
  827. asml.remove(hp1.next)
  828. hp1.next.free;
  829. asml.remove(hp1);
  830. hp1.free;
  831. }
  832. { remove Bcc }
  833. tasmlabel(taicpu(hp3).oper[0]^.ref^.symbol).decrefs;
  834. asml.remove(hp3);
  835. hp3.free;
  836. { remove jmp }
  837. tasmlabel(taicpu(hp2).oper[0]^.ref^.symbol).decrefs;
  838. asml.remove(hp2);
  839. hp2.free;
  840. continue;
  841. end;
  842. end;
  843. end;
  844. end;
  845. end;
  846. end;
  847. end;
  848. end;
  849. p := tai(p.next)
  850. end;
  851. end;
  852. function TCpuAsmOptimizer.RegInInstruction(Reg: TRegister; p1: tai): Boolean;
  853. begin
  854. If (p1.typ = ait_instruction) and (taicpu(p1).opcode=A_BL) then
  855. Result:=true
  856. else
  857. Result:=inherited RegInInstruction(Reg, p1);
  858. end;
  859. const
  860. { set of opcode which might or do write to memory }
  861. { TODO : extend armins.dat to contain r/w info }
  862. opcode_could_mem_write = [A_B,A_BL,A_BLX,A_BKPT,A_BX,A_STR,A_STRB,A_STRBT,
  863. A_STRH,A_STRT,A_STF,A_SFM,A_STM,A_FSTS,A_FSTD];
  864. function TCpuPreRegallocScheduler.PeepHoleOptPass1Cpu(var p: tai): boolean;
  865. { TODO : schedule also forward }
  866. { TODO : schedule distance > 1 }
  867. var
  868. hp1,hp2,hp3,hp4,hp5 : tai;
  869. list : TAsmList;
  870. begin
  871. result:=true;
  872. list:=TAsmList.Create;
  873. p := BlockStart;
  874. { UsedRegs := []; }
  875. while (p <> BlockEnd) Do
  876. begin
  877. if (p.typ=ait_instruction) and
  878. GetNextInstruction(p,hp1) and
  879. (hp1.typ=ait_instruction) and
  880. { for now we don't reschedule if the previous instruction changes potentially a memory location }
  881. ( (not(taicpu(p).opcode in opcode_could_mem_write) and
  882. not(RegModifiedByInstruction(NR_PC,p)) and
  883. (taicpu(hp1).opcode in [A_LDR,A_LDRB,A_LDRH,A_LDRSB,A_LDRSH])
  884. ) or
  885. ((taicpu(p).opcode in [A_STM,A_STRB,A_STRH,A_STR]) and
  886. (taicpu(hp1).opcode in [A_LDR,A_LDRB,A_LDRH,A_LDRSB,A_LDRSH]) and
  887. ((taicpu(hp1).oper[1]^.ref^.base=NR_PC) or
  888. (assigned(taicpu(hp1).oper[1]^.ref^.symboldata) and
  889. (taicpu(hp1).oper[1]^.ref^.offset=0)
  890. )
  891. ) or
  892. { try to prove that the memory accesses don't overlapp }
  893. ((taicpu(p).opcode in [A_STRB,A_STRH,A_STR]) and
  894. (taicpu(hp1).opcode in [A_LDR,A_LDRB,A_LDRH,A_LDRSB,A_LDRSH]) and
  895. (taicpu(p).oper[1]^.ref^.base=taicpu(hp1).oper[1]^.ref^.base) and
  896. (taicpu(p).oppostfix=PF_None) and
  897. (taicpu(hp1).oppostfix=PF_None) and
  898. (taicpu(p).oper[1]^.ref^.index=NR_NO) and
  899. (taicpu(hp1).oper[1]^.ref^.index=NR_NO) and
  900. { get operand sizes and check if the offset distance is large enough to ensure no overlapp }
  901. (abs(taicpu(p).oper[1]^.ref^.offset-taicpu(hp1).oper[1]^.ref^.offset)>=max(tcgsize2size[reg_cgsize(taicpu(p).oper[0]^.reg)],tcgsize2size[reg_cgsize(taicpu(hp1).oper[0]^.reg)]))
  902. )
  903. )
  904. ) and
  905. GetNextInstruction(hp1,hp2) and
  906. (hp2.typ=ait_instruction) and
  907. { loaded register used by next instruction? }
  908. (RegInInstruction(taicpu(hp1).oper[0]^.reg,hp2)) and
  909. { loaded register not used by previous instruction? }
  910. not(RegInInstruction(taicpu(hp1).oper[0]^.reg,p)) and
  911. { same condition? }
  912. (taicpu(p).condition=taicpu(hp1).condition) and
  913. { first instruction might not change the register used as base }
  914. ((taicpu(hp1).oper[1]^.ref^.base=NR_NO) or
  915. not(RegModifiedByInstruction(taicpu(hp1).oper[1]^.ref^.base,p))
  916. ) and
  917. { first instruction might not change the register used as index }
  918. ((taicpu(hp1).oper[1]^.ref^.index=NR_NO) or
  919. not(RegModifiedByInstruction(taicpu(hp1).oper[1]^.ref^.index,p))
  920. ) then
  921. begin
  922. hp3:=tai(p.Previous);
  923. hp5:=tai(p.next);
  924. asml.Remove(p);
  925. { if there is a reg. dealloc instruction associated with p, move it together with p }
  926. { before the instruction? }
  927. while assigned(hp3) and (hp3.typ<>ait_instruction) do
  928. begin
  929. if (hp3.typ=ait_regalloc) and (tai_regalloc(hp3).ratype in [ra_dealloc]) and
  930. RegInInstruction(tai_regalloc(hp3).reg,p) then
  931. begin
  932. hp4:=hp3;
  933. hp3:=tai(hp3.Previous);
  934. asml.Remove(hp4);
  935. list.Concat(hp4);
  936. end
  937. else
  938. hp3:=tai(hp3.Previous);
  939. end;
  940. list.Concat(p);
  941. { after the instruction? }
  942. while assigned(hp5) and (hp5.typ<>ait_instruction) do
  943. begin
  944. if (hp5.typ=ait_regalloc) and (tai_regalloc(hp5).ratype in [ra_dealloc]) and
  945. RegInInstruction(tai_regalloc(hp5).reg,p) then
  946. begin
  947. hp4:=hp5;
  948. hp5:=tai(hp5.next);
  949. asml.Remove(hp4);
  950. list.Concat(hp4);
  951. end
  952. else
  953. hp5:=tai(hp5.Next);
  954. end;
  955. asml.Remove(hp1);
  956. {$ifdef DEBUG_PREREGSCHEDULER}
  957. asml.InsertBefore(tai_comment.Create(strpnew('Rescheduled')),hp2);
  958. {$endif DEBUG_PREREGSCHEDULER}
  959. asml.InsertBefore(hp1,hp2);
  960. asml.InsertListBefore(hp2,list);
  961. end;
  962. p := tai(p.next)
  963. end;
  964. list.Free;
  965. end;
  966. procedure TCpuThumb2AsmOptimizer.PeepHoleOptPass2;
  967. begin
  968. { TODO: Add optimizer code }
  969. end;
  970. begin
  971. casmoptimizer:=TCpuAsmOptimizer;
  972. cpreregallocscheduler:=TCpuPreRegallocScheduler;
  973. End.