1 -----------------------------------------------------------------------------
3 -- Machine-dependent assembly language
5 -- (c) The University of Glasgow 1993-2004
7 -----------------------------------------------------------------------------
9 #include "HsVersions.h"
10 #include "nativeGen/NCG.h"
28 import Constants (rESERVED_C_STACK_BYTES)
34 -- Size of a PPC memory address, in bytes.
39 #elif x86_64_TARGET_ARCH
42 archWordSize = panic "X86.Instr.archWordSize: not defined"
45 -- | Instruction instance for x86 instruction set.
46 instance Instruction Instr where
47 regUsageOfInstr = x86_regUsageOfInstr
48 patchRegsOfInstr = x86_patchRegsOfInstr
49 isJumpishInstr = x86_isJumpishInstr
50 jumpDestsOfInstr = x86_jumpDestsOfInstr
51 patchJumpInstr = x86_patchJumpInstr
52 mkSpillInstr = x86_mkSpillInstr
53 mkLoadInstr = x86_mkLoadInstr
54 takeDeltaInstr = x86_takeDeltaInstr
55 isMetaInstr = x86_isMetaInstr
56 mkRegRegMoveInstr = x86_mkRegRegMoveInstr
57 takeRegRegMoveInstr = x86_takeRegRegMoveInstr
58 mkJumpInstr = x86_mkJumpInstr
61 -- -----------------------------------------------------------------------------
62 -- Intel x86 instructions
65 Intel, in their infinite wisdom, selected a stack model for floating
66 point registers on x86. That might have made sense back in 1979 --
67 nowadays we can see it for the nonsense it really is. A stack model
68 fits poorly with the existing nativeGen infrastructure, which assumes
69 flat integer and FP register sets. Prior to this commit, nativeGen
70 could not generate correct x86 FP code -- to do so would have meant
71 somehow working the register-stack paradigm into the register
72 allocator and spiller, which sounds very difficult.
74 We have decided to cheat, and go for a simple fix which requires no
75 infrastructure modifications, at the expense of generating ropey but
76 correct FP code. All notions of the x86 FP stack and its insns have
77 been removed. Instead, we pretend (to the instruction selector and
78 register allocator) that x86 has six floating point registers, %fake0
79 .. %fake5, which can be used in the usual flat manner. We further
80 claim that x86 has floating point instructions very similar to SPARC
81 and Alpha, that is, a simple 3-operand register-register arrangement.
82 Code generation and register allocation proceed on this basis.
84 When we come to print out the final assembly, our convenient fiction
85 is converted to dismal reality. Each fake instruction is
86 independently converted to a series of real x86 instructions.
87 %fake0 .. %fake5 are mapped to %st(0) .. %st(5). To do reg-reg
88 arithmetic operations, the two operands are pushed onto the top of the
89 FP stack, the operation done, and the result copied back into the
90 relevant register. There are only six %fake registers because 2 are
91 needed for the translation, and x86 has 8 in total.
93 The translation is inefficient but is simple and it works. A cleverer
94 translation would handle a sequence of insns, simulating the FP stack
95 contents, would not impose a fixed mapping from %fake to %st regs, and
96 hopefully could avoid most of the redundant reg-reg moves of the
99 We might as well make use of whatever unique FP facilities Intel have
100 chosen to bless us with (let's not be churlish, after all).
101 Hence GLDZ and GLD1. Bwahahahahahahaha!
105 MORE FLOATING POINT MUSINGS...
107 Intel's internal floating point registers are by default 80 bit
108 extended precision. This means that all operations done on values in
109 registers are done at 80 bits, and unless the intermediate values are
110 truncated to the appropriate size (32 or 64 bits) by storing in
111 memory, calculations in registers will give different results from
112 calculations which pass intermediate values in memory (eg. via
115 One solution is to set the FPU into 64 bit precision mode. Some OSs
116 do this (eg. FreeBSD) and some don't (eg. Linux). The problem here is
117 that this will only affect 64-bit precision arithmetic; 32-bit
118 calculations will still be done at 64-bit precision in registers. So
119 it doesn't solve the whole problem.
121 There's also the issue of what the C library is expecting in terms of
122 precision. It seems to be the case that glibc on Linux expects the
123 FPU to be set to 80 bit precision, so setting it to 64 bit could have
124 unexpected effects. Changing the default could have undesirable
125 effects on other 3rd-party library code too, so the right thing would
126 be to save/restore the FPU control word across Haskell code if we were
129 gcc's -ffloat-store gives consistent results by always storing the
130 results of floating-point calculations in memory, which works for both
131 32 and 64-bit precision. However, it only affects the values of
132 user-declared floating point variables in C, not intermediate results.
133 GHC in -fvia-C mode uses -ffloat-store (see the -fexcess-precision
136 Another problem is how to spill floating point registers in the
137 register allocator. Should we spill the whole 80 bits, or just 64?
138 On an OS which is set to 64 bit precision, spilling 64 is fine. On
139 Linux, spilling 64 bits will round the results of some operations.
140 This is what gcc does. Spilling at 80 bits requires taking up a full
141 128 bit slot (so we get alignment). We spill at 80-bits and ignore
142 the alignment problems.
144 In the future, we'll use the SSE registers for floating point. This
145 requires a CPU that supports SSE2 (ordinary SSE only supports 32 bit
146 precision float ops), which means P4 or Xeon and above. Using SSE
147 will solve all these problems, because the SSE registers use fixed 32
148 bit or 64 bit precision.
158 -- some static data spat out during code
159 -- generation. Will be extracted before
161 | LDATA Section [CmmStatic]
163 -- start a new basic block. Useful during
164 -- codegen, removed later. Preceding
165 -- instruction should be a jump, as per the
166 -- invariants for a BasicBlock (see Cmm).
169 -- specify current stack offset for
170 -- benefit of subsequent passes
174 | MOV Size Operand Operand
175 | MOVZxL Size Operand Operand -- size is the size of operand 1
176 | MOVSxL Size Operand Operand -- size is the size of operand 1
177 -- x86_64 note: plain mov into a 32-bit register always zero-extends
178 -- into the 64-bit reg, in contrast to the 8 and 16-bit movs which
179 -- don't affect the high bits of the register.
181 -- Load effective address (also a very useful three-operand add instruction :-)
182 | LEA Size Operand Operand
185 | ADD Size Operand Operand
186 | ADC Size Operand Operand
187 | SUB Size Operand Operand
189 | MUL Size Operand Operand
190 | IMUL Size Operand Operand -- signed int mul
191 | IMUL2 Size Operand -- %edx:%eax = operand * %eax
193 | DIV Size Operand -- eax := eax:edx/op, edx := eax:edx%op
194 | IDIV Size Operand -- ditto, but signed
196 -- Simple bit-twiddling.
197 | AND Size Operand Operand
198 | OR Size Operand Operand
199 | XOR Size Operand Operand
201 | NEGI Size Operand -- NEG instruction (name clash with Cond)
203 -- Shifts (amount may be immediate or %cl only)
204 | SHL Size Operand{-amount-} Operand
205 | SAR Size Operand{-amount-} Operand
206 | SHR Size Operand{-amount-} Operand
208 | BT Size Imm Operand
211 -- x86 Float Arithmetic.
212 -- Note that we cheat by treating G{ABS,MOV,NEG} of doubles
213 -- as single instructions right up until we spit them out.
214 -- all the 3-operand fake fp insns are src1 src2 dst
215 -- and furthermore are constrained to be fp regs only.
216 -- IMPORTANT: keep is_G_insn up to date with any changes here
217 | GMOV Reg Reg -- src(fpreg), dst(fpreg)
218 | GLD Size AddrMode Reg -- src, dst(fpreg)
219 | GST Size Reg AddrMode -- src(fpreg), dst
221 | GLDZ Reg -- dst(fpreg)
222 | GLD1 Reg -- dst(fpreg)
224 | GFTOI Reg Reg -- src(fpreg), dst(intreg)
225 | GDTOI Reg Reg -- src(fpreg), dst(intreg)
227 | GITOF Reg Reg -- src(intreg), dst(fpreg)
228 | GITOD Reg Reg -- src(intreg), dst(fpreg)
230 | GADD Size Reg Reg Reg -- src1, src2, dst
231 | GDIV Size Reg Reg Reg -- src1, src2, dst
232 | GSUB Size Reg Reg Reg -- src1, src2, dst
233 | GMUL Size Reg Reg Reg -- src1, src2, dst
235 -- FP compare. Cond must be `elem` [EQQ, NE, LE, LTT, GE, GTT]
236 -- Compare src1 with src2; set the Zero flag iff the numbers are
237 -- comparable and the comparison is True. Subsequent code must
238 -- test the %eflags zero flag regardless of the supplied Cond.
239 | GCMP Cond Reg Reg -- src1, src2
241 | GABS Size Reg Reg -- src, dst
242 | GNEG Size Reg Reg -- src, dst
243 | GSQRT Size Reg Reg -- src, dst
244 | GSIN Size CLabel CLabel Reg Reg -- src, dst
245 | GCOS Size CLabel CLabel Reg Reg -- src, dst
246 | GTAN Size CLabel CLabel Reg Reg -- src, dst
248 | GFREE -- do ffree on all x86 regs; an ugly hack
251 -- SSE2 floating point: we use a restricted set of the available SSE2
252 -- instructions for floating-point.
253 -- use MOV for moving (either movss or movsd (movlpd better?))
254 | CVTSS2SD Reg Reg -- F32 to F64
255 | CVTSD2SS Reg Reg -- F64 to F32
256 | CVTTSS2SIQ Size Operand Reg -- F32 to I32/I64 (with truncation)
257 | CVTTSD2SIQ Size Operand Reg -- F64 to I32/I64 (with truncation)
258 | CVTSI2SS Size Operand Reg -- I32/I64 to F32
259 | CVTSI2SD Size Operand Reg -- I32/I64 to F64
261 -- use ADD & SUB for arithmetic. In both cases, operands
264 -- SSE2 floating-point division:
265 | FDIV Size Operand Operand -- divisor, dividend(dst)
267 -- use CMP for comparisons. ucomiss and ucomisd instructions
268 -- compare single/double prec floating point respectively.
270 | SQRT Size Operand Reg -- src, dst
274 | TEST Size Operand Operand
275 | CMP Size Operand Operand
281 -- both unused (SDM):
287 | JXX Cond BlockId -- includes unconditional branches
288 | JXX_GBL Cond Imm -- non-local version of JXX
289 | JMP_TBL Operand [BlockId] -- table jump
290 | CALL (Either Imm Reg) [Reg]
293 | CLTD Size -- sign extend %eax into %edx:%eax
295 | FETCHGOT Reg -- pseudo-insn for ELF position-independent code
299 -- addl __GLOBAL_OFFSET_TABLE__+.-1b, %reg
300 | FETCHPC Reg -- pseudo-insn for Darwin position-independent code
307 = OpReg Reg -- register
308 | OpImm Imm -- immediate value
309 | OpAddr AddrMode -- memory reference
313 x86_regUsageOfInstr :: Instr -> RegUsage
314 x86_regUsageOfInstr instr
316 MOV _ src dst -> usageRW src dst
317 MOVZxL _ src dst -> usageRW src dst
318 MOVSxL _ src dst -> usageRW src dst
319 LEA _ src dst -> usageRW src dst
320 ADD _ src dst -> usageRM src dst
321 ADC _ src dst -> usageRM src dst
322 SUB _ src dst -> usageRM src dst
323 IMUL _ src dst -> usageRM src dst
324 IMUL2 _ src -> mkRU (eax:use_R src) [eax,edx]
325 MUL _ src dst -> usageRM src dst
326 DIV _ op -> mkRU (eax:edx:use_R op) [eax,edx]
327 IDIV _ op -> mkRU (eax:edx:use_R op) [eax,edx]
328 AND _ src dst -> usageRM src dst
329 OR _ src dst -> usageRM src dst
331 XOR _ (OpReg src) (OpReg dst)
332 | src == dst -> mkRU [] [dst]
334 XOR _ src dst -> usageRM src dst
335 NOT _ op -> usageM op
336 NEGI _ op -> usageM op
337 SHL _ imm dst -> usageRM imm dst
338 SAR _ imm dst -> usageRM imm dst
339 SHR _ imm dst -> usageRM imm dst
340 BT _ _ src -> mkRUR (use_R src)
342 PUSH _ op -> mkRUR (use_R op)
343 POP _ op -> mkRU [] (def_W op)
344 TEST _ src dst -> mkRUR (use_R src ++ use_R dst)
345 CMP _ src dst -> mkRUR (use_R src ++ use_R dst)
346 SETCC _ op -> mkRU [] (def_W op)
347 JXX _ _ -> mkRU [] []
348 JXX_GBL _ _ -> mkRU [] []
349 JMP op -> mkRUR (use_R op)
350 JMP_TBL op _ -> mkRUR (use_R op)
351 CALL (Left _) params -> mkRU params callClobberedRegs
352 CALL (Right reg) params -> mkRU (reg:params) callClobberedRegs
353 CLTD _ -> mkRU [eax] [edx]
356 GMOV src dst -> mkRU [src] [dst]
357 GLD _ src dst -> mkRU (use_EA src) [dst]
358 GST _ src dst -> mkRUR (src : use_EA dst)
360 GLDZ dst -> mkRU [] [dst]
361 GLD1 dst -> mkRU [] [dst]
363 GFTOI src dst -> mkRU [src] [dst]
364 GDTOI src dst -> mkRU [src] [dst]
366 GITOF src dst -> mkRU [src] [dst]
367 GITOD src dst -> mkRU [src] [dst]
369 GADD _ s1 s2 dst -> mkRU [s1,s2] [dst]
370 GSUB _ s1 s2 dst -> mkRU [s1,s2] [dst]
371 GMUL _ s1 s2 dst -> mkRU [s1,s2] [dst]
372 GDIV _ s1 s2 dst -> mkRU [s1,s2] [dst]
374 GCMP _ src1 src2 -> mkRUR [src1,src2]
375 GABS _ src dst -> mkRU [src] [dst]
376 GNEG _ src dst -> mkRU [src] [dst]
377 GSQRT _ src dst -> mkRU [src] [dst]
378 GSIN _ _ _ src dst -> mkRU [src] [dst]
379 GCOS _ _ _ src dst -> mkRU [src] [dst]
380 GTAN _ _ _ src dst -> mkRU [src] [dst]
382 CVTSS2SD src dst -> mkRU [src] [dst]
383 CVTSD2SS src dst -> mkRU [src] [dst]
384 CVTTSS2SIQ _ src dst -> mkRU (use_R src) [dst]
385 CVTTSD2SIQ _ src dst -> mkRU (use_R src) [dst]
386 CVTSI2SS _ src dst -> mkRU (use_R src) [dst]
387 CVTSI2SD _ src dst -> mkRU (use_R src) [dst]
388 FDIV _ src dst -> usageRM src dst
390 FETCHGOT reg -> mkRU [] [reg]
391 FETCHPC reg -> mkRU [] [reg]
396 _other -> panic "regUsage: unrecognised instr"
399 -- 2 operand form; first operand Read; second Written
400 usageRW :: Operand -> Operand -> RegUsage
401 usageRW op (OpReg reg) = mkRU (use_R op) [reg]
402 usageRW op (OpAddr ea) = mkRUR (use_R op ++ use_EA ea)
403 usageRW _ _ = panic "X86.RegInfo.usageRW: no match"
405 -- 2 operand form; first operand Read; second Modified
406 usageRM :: Operand -> Operand -> RegUsage
407 usageRM op (OpReg reg) = mkRU (use_R op ++ [reg]) [reg]
408 usageRM op (OpAddr ea) = mkRUR (use_R op ++ use_EA ea)
409 usageRM _ _ = panic "X86.RegInfo.usageRM: no match"
411 -- 1 operand form; operand Modified
412 usageM :: Operand -> RegUsage
413 usageM (OpReg reg) = mkRU [reg] [reg]
414 usageM (OpAddr ea) = mkRUR (use_EA ea)
415 usageM _ = panic "X86.RegInfo.usageM: no match"
417 -- Registers defd when an operand is written.
418 def_W (OpReg reg) = [reg]
419 def_W (OpAddr _ ) = []
420 def_W _ = panic "X86.RegInfo.def_W: no match"
422 -- Registers used when an operand is read.
423 use_R (OpReg reg) = [reg]
425 use_R (OpAddr ea) = use_EA ea
427 -- Registers used to compute an effective address.
428 use_EA (ImmAddr _ _) = []
429 use_EA (AddrBaseIndex base index _) =
430 use_base base $! use_index index
431 where use_base (EABaseReg r) x = r : x
433 use_index EAIndexNone = []
434 use_index (EAIndex i _) = [i]
436 mkRUR src = src' `seq` RU src' []
437 where src' = filter interesting src
439 mkRU src dst = src' `seq` dst' `seq` RU src' dst'
440 where src' = filter interesting src
441 dst' = filter interesting dst
443 interesting :: Reg -> Bool
444 interesting (RegVirtual _) = True
445 interesting (RegReal (RealRegSingle i)) = isFastTrue (freeReg i)
446 interesting (RegReal (RealRegPair{})) = panic "X86.interesting: no reg pairs on this arch"
450 x86_patchRegsOfInstr :: Instr -> (Reg -> Reg) -> Instr
451 x86_patchRegsOfInstr instr env
453 MOV sz src dst -> patch2 (MOV sz) src dst
454 MOVZxL sz src dst -> patch2 (MOVZxL sz) src dst
455 MOVSxL sz src dst -> patch2 (MOVSxL sz) src dst
456 LEA sz src dst -> patch2 (LEA sz) src dst
457 ADD sz src dst -> patch2 (ADD sz) src dst
458 ADC sz src dst -> patch2 (ADC sz) src dst
459 SUB sz src dst -> patch2 (SUB sz) src dst
460 IMUL sz src dst -> patch2 (IMUL sz) src dst
461 IMUL2 sz src -> patch1 (IMUL2 sz) src
462 MUL sz src dst -> patch2 (MUL sz) src dst
463 IDIV sz op -> patch1 (IDIV sz) op
464 DIV sz op -> patch1 (DIV sz) op
465 AND sz src dst -> patch2 (AND sz) src dst
466 OR sz src dst -> patch2 (OR sz) src dst
467 XOR sz src dst -> patch2 (XOR sz) src dst
468 NOT sz op -> patch1 (NOT sz) op
469 NEGI sz op -> patch1 (NEGI sz) op
470 SHL sz imm dst -> patch1 (SHL sz imm) dst
471 SAR sz imm dst -> patch1 (SAR sz imm) dst
472 SHR sz imm dst -> patch1 (SHR sz imm) dst
473 BT sz imm src -> patch1 (BT sz imm) src
474 TEST sz src dst -> patch2 (TEST sz) src dst
475 CMP sz src dst -> patch2 (CMP sz) src dst
476 PUSH sz op -> patch1 (PUSH sz) op
477 POP sz op -> patch1 (POP sz) op
478 SETCC cond op -> patch1 (SETCC cond) op
479 JMP op -> patch1 JMP op
480 JMP_TBL op ids -> patch1 JMP_TBL op $ ids
482 GMOV src dst -> GMOV (env src) (env dst)
483 GLD sz src dst -> GLD sz (lookupAddr src) (env dst)
484 GST sz src dst -> GST sz (env src) (lookupAddr dst)
486 GLDZ dst -> GLDZ (env dst)
487 GLD1 dst -> GLD1 (env dst)
489 GFTOI src dst -> GFTOI (env src) (env dst)
490 GDTOI src dst -> GDTOI (env src) (env dst)
492 GITOF src dst -> GITOF (env src) (env dst)
493 GITOD src dst -> GITOD (env src) (env dst)
495 GADD sz s1 s2 dst -> GADD sz (env s1) (env s2) (env dst)
496 GSUB sz s1 s2 dst -> GSUB sz (env s1) (env s2) (env dst)
497 GMUL sz s1 s2 dst -> GMUL sz (env s1) (env s2) (env dst)
498 GDIV sz s1 s2 dst -> GDIV sz (env s1) (env s2) (env dst)
500 GCMP sz src1 src2 -> GCMP sz (env src1) (env src2)
501 GABS sz src dst -> GABS sz (env src) (env dst)
502 GNEG sz src dst -> GNEG sz (env src) (env dst)
503 GSQRT sz src dst -> GSQRT sz (env src) (env dst)
504 GSIN sz l1 l2 src dst -> GSIN sz l1 l2 (env src) (env dst)
505 GCOS sz l1 l2 src dst -> GCOS sz l1 l2 (env src) (env dst)
506 GTAN sz l1 l2 src dst -> GTAN sz l1 l2 (env src) (env dst)
508 CVTSS2SD src dst -> CVTSS2SD (env src) (env dst)
509 CVTSD2SS src dst -> CVTSD2SS (env src) (env dst)
510 CVTTSS2SIQ sz src dst -> CVTTSS2SIQ sz (patchOp src) (env dst)
511 CVTTSD2SIQ sz src dst -> CVTTSD2SIQ sz (patchOp src) (env dst)
512 CVTSI2SS sz src dst -> CVTSI2SS sz (patchOp src) (env dst)
513 CVTSI2SD sz src dst -> CVTSI2SD sz (patchOp src) (env dst)
514 FDIV sz src dst -> FDIV sz (patchOp src) (patchOp dst)
516 CALL (Left _) _ -> instr
517 CALL (Right reg) p -> CALL (Right (env reg)) p
519 FETCHGOT reg -> FETCHGOT (env reg)
520 FETCHPC reg -> FETCHPC (env reg)
530 _other -> panic "patchRegs: unrecognised instr"
533 patch1 insn op = insn $! patchOp op
534 patch2 insn src dst = (insn $! patchOp src) $! patchOp dst
536 patchOp (OpReg reg) = OpReg $! env reg
537 patchOp (OpImm imm) = OpImm imm
538 patchOp (OpAddr ea) = OpAddr $! lookupAddr ea
540 lookupAddr (ImmAddr imm off) = ImmAddr imm off
541 lookupAddr (AddrBaseIndex base index disp)
542 = ((AddrBaseIndex $! lookupBase base) $! lookupIndex index) disp
544 lookupBase EABaseNone = EABaseNone
545 lookupBase EABaseRip = EABaseRip
546 lookupBase (EABaseReg r) = EABaseReg (env r)
548 lookupIndex EAIndexNone = EAIndexNone
549 lookupIndex (EAIndex r i) = EAIndex (env r) i
552 --------------------------------------------------------------------------------
556 x86_isJumpishInstr instr
570 x86_jumpDestsOfInstr insn
578 :: Instr -> (BlockId -> BlockId) -> Instr
580 x86_patchJumpInstr insn patchF
582 JXX cc id -> JXX cc (patchF id)
583 JMP_TBL _ _ -> error "Cannot patch JMP_TBL"
589 -- -----------------------------------------------------------------------------
590 -- | Make a spill instruction.
592 :: Reg -- register to spill
593 -> Int -- current stack delta
594 -> Int -- spill slot to use
597 x86_mkSpillInstr reg delta slot
598 = let off = spillSlotToOffset slot
600 let off_w = (off-delta) `div` IF_ARCH_i386(4,8)
601 in case targetClassOfReg reg of
602 RcInteger -> MOV IF_ARCH_i386(II32,II64)
603 (OpReg reg) (OpAddr (spRel off_w))
604 RcDouble -> GST FF80 reg (spRel off_w) {- RcFloat/RcDouble -}
605 RcDoubleSSE -> MOV FF64 (OpReg reg) (OpAddr (spRel off_w))
606 _ -> panic "X86.mkSpillInstr: no match"
609 -- | Make a spill reload instruction.
611 :: Reg -- register to load
612 -> Int -- current stack delta
613 -> Int -- spill slot to use
616 x86_mkLoadInstr reg delta slot
617 = let off = spillSlotToOffset slot
619 let off_w = (off-delta) `div` IF_ARCH_i386(4,8)
620 in case targetClassOfReg reg of
621 RcInteger -> MOV IF_ARCH_i386(II32,II64)
622 (OpAddr (spRel off_w)) (OpReg reg)
623 RcDouble -> GLD FF80 (spRel off_w) reg {- RcFloat/RcDouble -}
624 RcDoubleSSE -> MOV FF64 (OpAddr (spRel off_w)) (OpReg reg)
625 _ -> panic "X86.x86_mkLoadInstr"
628 spillSlotSize = IF_ARCH_i386(12, 8)
631 maxSpillSlots = ((rESERVED_C_STACK_BYTES - 64) `div` spillSlotSize) - 1
633 -- convert a spill slot number to a *byte* offset, with no sign:
634 -- decide on a per arch basis whether you are spilling above or below
635 -- the C stack pointer.
636 spillSlotToOffset :: Int -> Int
637 spillSlotToOffset slot
638 | slot >= 0 && slot < maxSpillSlots
639 = 64 + spillSlotSize * slot
641 = pprPanic "spillSlotToOffset:"
642 ( text "invalid spill location: " <> int slot
643 $$ text "maxSpillSlots: " <> int maxSpillSlots)
645 --------------------------------------------------------------------------------
647 -- | See if this instruction is telling us the current C stack delta
652 x86_takeDeltaInstr instr
662 x86_isMetaInstr instr
672 -- | Make a reg-reg move instruction.
673 -- On SPARC v8 there are no instructions to move directly between
674 -- floating point and integer regs. If we need to do that then we
675 -- have to go via memory.
677 x86_mkRegRegMoveInstr
682 x86_mkRegRegMoveInstr src dst
683 = case targetClassOfReg src of
685 RcInteger -> MOV II32 (OpReg src) (OpReg dst)
687 RcInteger -> MOV II64 (OpReg src) (OpReg dst)
689 RcDouble -> GMOV src dst
690 RcDoubleSSE -> MOV FF64 (OpReg src) (OpReg dst)
691 _ -> panic "X86.RegInfo.mkRegRegMoveInstr: no match"
693 -- | Check whether an instruction represents a reg-reg move.
694 -- The register allocator attempts to eliminate reg->reg moves whenever it can,
695 -- by assigning the src and dest temporaries to the same real register.
697 x86_takeRegRegMoveInstr
701 x86_takeRegRegMoveInstr (MOV _ (OpReg r1) (OpReg r2))
704 x86_takeRegRegMoveInstr _ = Nothing
707 -- | Make an unconditional branch instruction.
720 :: [GenBasicBlock Instr]
721 -> [GenBasicBlock Instr]
723 i386_insert_ffrees blocks
724 | or (map (any is_G_instr) [ instrs | BasicBlock _ instrs <- blocks ])
725 = map ffree_before_nonlocal_transfers blocks
730 ffree_before_nonlocal_transfers (BasicBlock id insns)
731 = BasicBlock id (foldr p [] insns)
732 where p insn r = case insn of
733 CALL _ _ -> GFREE : insn : r
734 JMP _ -> GFREE : insn : r
737 -- if you ever add a new FP insn to the fake x86 FP insn set,
738 -- you must update this too
739 is_G_instr :: Instr -> Bool
762 GFREE -> panic "is_G_instr: GFREE (!)"
766 data JumpDest = DestBlockId BlockId | DestImm Imm
769 canShortcut :: Instr -> Maybe JumpDest
770 canShortcut (JXX ALWAYS id) = Just (DestBlockId id)
771 canShortcut (JMP (OpImm imm)) = Just (DestImm imm)
772 canShortcut _ = Nothing
775 -- This helper shortcuts a sequence of branches.
776 -- The blockset helps avoid following cycles.
777 shortcutJump :: (BlockId -> Maybe JumpDest) -> Instr -> Instr
778 shortcutJump fn insn = shortcutJump' fn emptyBlockSet insn
779 where shortcutJump' fn seen insn@(JXX cc id) =
780 if elemBlockSet id seen then insn
783 Just (DestBlockId id') -> shortcutJump' fn seen' (JXX cc id')
784 Just (DestImm imm) -> shortcutJump' fn seen' (JXX_GBL cc imm)
785 where seen' = extendBlockSet seen id
786 shortcutJump' _ _ other = other
788 -- Here because it knows about JumpDest
789 shortcutStatic :: (BlockId -> Maybe JumpDest) -> CmmStatic -> CmmStatic
790 shortcutStatic fn (CmmStaticLit (CmmLabel lab))
791 | Just uq <- maybeAsmTemp lab
792 = CmmStaticLit (CmmLabel (shortBlockId fn emptyUniqSet (BlockId uq)))
793 shortcutStatic fn (CmmStaticLit (CmmLabelDiffOff lbl1 lbl2 off))
794 | Just uq <- maybeAsmTemp lbl1
795 = CmmStaticLit (CmmLabelDiffOff (shortBlockId fn emptyUniqSet (BlockId uq)) lbl2 off)
796 -- slightly dodgy, we're ignoring the second label, but this
797 -- works with the way we use CmmLabelDiffOff for jump tables now.
799 shortcutStatic _ other_static
803 :: (BlockId -> Maybe JumpDest)
808 shortBlockId fn seen blockid@(BlockId uq) =
809 case (elementOfUniqSet uq seen, fn blockid) of
810 (True, _) -> mkAsmTempLabel uq
811 (_, Nothing) -> mkAsmTempLabel uq
812 (_, Just (DestBlockId blockid')) -> shortBlockId fn (addOneToUniqSet seen uq) blockid'
813 (_, Just (DestImm (ImmCLbl lbl))) -> lbl
814 (_, _other) -> panic "shortBlockId"