1 module PositionIndependentCode (
2 cmmMakeDynamicReference,
10 This module handles generation of position independent code and
11 dynamic-linking related issues for the native code generator.
13 Things outside this module which are related to this:
16 - PIC base label (pretty printed as local label 1)
17 - DynamicLinkerLabels - several kinds:
18 CodeStub, SymbolPtr, GotSymbolPtr, GotSymbolOffset
19 - labelDynamic predicate
21 - The CmmExpr datatype has a CmmPicBaseReg constructor
22 - The CmmLit datatype has a CmmLabelDiffOff constructor
24 - When tablesNextToCode, no absolute addresses are stored in info tables
25 any more. Instead, offsets from the info label are used.
26 - For Win32 only, SRTs might contain addresses of __imp_ symbol pointers
27 because Win32 doesn't support external references in data sections.
28 TODO: make sure this still works, it might be bitrotted
30 - The cmmToCmm pass in AsmCodeGen calls cmmMakeDynamicReference for all
32 - nativeCodeGen calls pprImportedSymbol and pprGotDeclaration to output
33 all the necessary stuff for imported symbols.
34 - The NCG monad keeps track of a list of imported symbols.
35 - MachCodeGen invokes initializePicBase to generate code to initialize
36 the PIC base register when needed.
37 - MachCodeGen calls cmmMakeDynamicReference whenever it uses a CLabel
38 that wasn't in the original Cmm code (e.g. floating point literals).
40 - The mangler converts absolure refs to relative refs in info tables
41 - Symbol pointers, stub code and PIC calculations that are generated
42 by GCC are left intact by the mangler (so far only on ppc-darwin
46 #include "HsVersions.h"
47 #include "nativeGen/NCG.h"
50 import MachOp ( MachOp(MO_Add), wordRep )
51 import CLabel ( CLabel, pprCLabel,
52 mkDynamicLinkerLabel, DynamicLinkerLabelInfo(..),
53 dynamicLinkerLabelInfo, mkPicBaseLabel,
54 labelDynamic, externallyVisibleCLabel )
57 import CLabel ( mkForeignLabel )
62 import NCGMonad ( NatM, getNewRegNat, getNewLabelNat )
64 import CmdLineOpts ( opt_PIC, opt_Static )
67 import qualified Outputable
69 import Panic ( panic )
72 -- The most important function here is cmmMakeDynamicReference.
74 -- It gets called by the cmmToCmm pass for every CmmLabel in the Cmm
75 -- code. It does The Right Thing(tm) to convert the CmmLabel into a
76 -- position-independent, dynamic-linking-aware reference to the thing
78 -- Note that this also has to be called from MachCodeGen in order to
79 -- access static data like floating point literals (labels that were
80 -- created after the cmmToCmm pass).
81 -- The function must run in a monad that can keep track of imported symbols
82 -- A function for recording an imported symbol must be passed in:
83 -- - addImportCmmOpt for the CmmOptM monad
84 -- - addImportNat for the NatM monad.
86 cmmMakeDynamicReference
87 :: Monad m => (CLabel -> m ()) -- a monad & a function
88 -- used for recording imported symbols
89 -> Bool -- whether this is the target of a jump
90 -> CLabel -- the label
93 cmmMakeDynamicReference addImport isJumpTarget lbl
94 | Just _ <- dynamicLinkerLabelInfo lbl
95 = return $ CmmLit $ CmmLabel lbl -- already processed it, pass through
96 | otherwise = case howToAccessLabel isJumpTarget lbl of
98 let stub = mkDynamicLinkerLabel CodeStub lbl
100 return $ CmmLit $ CmmLabel stub
101 AccessViaSymbolPtr -> do
102 let symbolPtr = mkDynamicLinkerLabel SymbolPtr lbl
104 return $ CmmLoad (cmmMakePicReference symbolPtr) wordRep
106 -- all currently supported processors support
107 -- a PC-relative branch instruction, so just jump there
108 | isJumpTarget -> return $ CmmLit $ CmmLabel lbl
109 -- for data, we might have to make some calculations:
110 | otherwise -> return $ cmmMakePicReference lbl
112 -- -------------------------------------------------------------------
114 -- Create a position independent reference to a label.
115 -- (but do not bother with dynamic linking).
116 -- We calculate the label's address by adding some (platform-dependent)
117 -- offset to our base register; this offset is calculated by
118 -- the function picRelative in the platform-dependent part below.
120 cmmMakePicReference :: CLabel -> CmmExpr
122 #if !mingw32_TARGET_OS
123 -- Windows doesn't need PIC,
124 -- everything gets relocated at runtime
126 cmmMakePicReference lbl
127 | opt_PIC && absoluteLabel lbl = CmmMachOp (MO_Add wordRep) [
129 CmmLit $ picRelative lbl
132 absoluteLabel lbl = case dynamicLinkerLabelInfo lbl of
133 Just (GotSymbolPtr, _) -> False
134 Just (GotSymbolOffset, _) -> False
138 cmmMakePicReference lbl = CmmLit $ CmmLabel lbl
140 -- ===================================================================
141 -- Platform dependent stuff
142 -- ===================================================================
144 -- Knowledge about how special dynamic linker labels like symbol
145 -- pointers, code stubs and GOT offsets look like is located in the
148 -- -------------------------------------------------------------------
150 -- We have to decide which labels need to be accessed
151 -- indirectly or via a piece of stub code.
153 data LabelAccessStyle = AccessViaStub
157 howToAccessLabel :: Bool -> CLabel -> LabelAccessStyle
159 #if mingw32_TARGET_OS
162 -- We need to use access *exactly* those things that
163 -- are imported from a DLL via an __imp_* label.
164 -- There are no stubs for imported code.
166 howToAccessLabel _ lbl | labelDynamic lbl = AccessViaSymbolPtr
167 | otherwise = AccessDirectly
169 #elif darwin_TARGET_OS
170 -- Mach-O (Darwin, Mac OS X)
172 -- Indirect access is required in the following cases:
173 -- * things imported from a dynamic library
174 -- * things from a different module, if we're generating PIC code
175 -- It is always possible to access something indirectly,
176 -- even when it's not necessary.
178 howToAccessLabel True lbl
179 -- jumps to a dynamic library go via a symbol stub
180 | labelDynamic lbl = AccessViaStub
181 -- when generating PIC code, all cross-module references must
182 -- must go via a symbol pointer, too.
183 -- Unfortunately, we don't know whether it's cross-module,
184 -- so we do it for all externally visible labels.
185 -- This is a slight waste of time and space, but otherwise
186 -- we'd need to pass the current Module all the way in to
188 | opt_PIC && externallyVisibleCLabel lbl = AccessViaStub
189 howToAccessLabel False lbl
190 -- data access to a dynamic library goes via a symbol pointer
191 | labelDynamic lbl = AccessViaSymbolPtr
192 -- cross-module PIC references: same as above
193 | opt_PIC && externallyVisibleCLabel lbl = AccessViaSymbolPtr
194 howToAccessLabel _ _ = AccessDirectly
196 #elif linux_TARGET_OS && powerpc64_TARGET_ARCH
197 -- ELF PPC64 (powerpc64-linux), AIX, MacOS 9, BeOS/PPC
199 howToAccessLabel True lbl = AccessDirectly -- actually, .label instead of label
200 howToAccessLabel _ lbl = AccessViaSymbolPtr
202 #elif linux_TARGET_OS
205 -- ELF tries to pretend to the main application code that dynamic linking does
206 -- not exist. While this may sound convenient, it tends to mess things up in
207 -- very bad ways, so we have to be careful when we generate code for the main
208 -- program (-dynamic but no -fPIC).
210 -- Indirect access is required for references to imported symbols
211 -- from position independent code. It is also required from the main program
212 -- when dynamic libraries containing Haskell code are used.
214 howToAccessLabel isJump lbl
215 -- no PIC -> the dynamic linker does everything for us;
216 -- if we don't dynamically link to Haskell code,
217 -- it actually manages to do so without messing thins up.
218 | not opt_PIC && opt_Static = AccessDirectly
220 #if !i386_TARGET_ARCH
221 -- for Intel, we temporarily disable the use of the
222 -- Procedure Linkage Table, because PLTs on intel require the
223 -- address of the GOT to be loaded into register %ebx before
224 -- a jump through the PLT is made.
225 -- TODO: make the i386 NCG ensure this before jumping to a
226 -- CodeStub label, so we can remove this special case.
228 -- As long as we're in a shared library ourselves,
229 -- we can use the plt.
230 -- NOTE: We might want to disable this, because this
231 -- prevents -fPIC code from being linked statically.
232 | isJump && labelDynamic lbl && opt_PIC = AccessViaStub
234 -- TODO: it would be OK to access non-Haskell code via a stub
235 -- | isJump && labelDynamic lbl && not isHaskellCode lbl = AccessViaStub
237 -- Using code stubs for jumps from the main program to an entry
238 -- label in a dynamic library is deadly; this will cause the dynamic
239 -- linker to replace all references (even data references) to that
240 -- label by references to the stub, so we won't find our info tables
244 -- A dynamic label needs to be accessed via a symbol pointer.
245 -- NOTE: It would be OK to jump to foreign code via a PLT stub.
246 | labelDynamic lbl = AccessViaSymbolPtr
248 #if powerpc_TARGET_ARCH
249 -- For PowerPC32 -fPIC, we have to access even static data
250 -- via a symbol pointer (see below for an explanation why
251 -- PowerPC32 Linux is especially broken).
252 | opt_PIC && not isJump = AccessViaSymbolPtr
255 | otherwise = AccessDirectly
259 -- all other platforms
262 | not opt_PIC = AccessDirectly
263 | otherwise = panic "howToAccessLabel: PIC not defined for this platform"
266 -- -------------------------------------------------------------------
268 -- What do we have to add to our 'PIC base register' in order to
269 -- get the address of a label?
271 picRelative :: CLabel -> CmmLit
274 -- The PIC base register points to the PIC base label at the beginning
275 -- of the current CmmTop. We just have to use a label difference to
277 -- We have already made sure that all labels that are not from the current
278 -- module are accessed indirectly ('as' can't calculate differences between
279 -- undefined labels).
282 = CmmLabelDiffOff lbl mkPicBaseLabel 0
284 #elif powerpc_TARGET_ARCH && linux_TARGET_OS
286 -- The PIC base register points to our fake GOT. Use a label difference
287 -- to get the offset.
288 -- We have made sure that *everything* is accessed indirectly, so this
289 -- is only used for offsets from the GOT to symbol pointers inside the
292 = CmmLabelDiffOff lbl gotLabel 0
294 #elif linux_TARGET_OS
295 -- Other Linux versions:
296 -- The PIC base register points to the GOT. Use foo@got for symbol
297 -- pointers, and foo@gotoff for everything else.
300 | Just (SymbolPtr, lbl') <- dynamicLinkerLabelInfo lbl
301 = CmmLabel $ mkDynamicLinkerLabel GotSymbolPtr lbl'
303 = CmmLabel $ mkDynamicLinkerLabel GotSymbolOffset lbl
306 picRelative lbl = panic "PositionIndependentCode.picRelative"
309 -- -------------------------------------------------------------------
311 -- What do we have to add to every assembly file we generate?
313 -- utility function for pretty-printing asm-labels,
314 -- copied from PprMach
315 asmSDoc d = Outputable.withPprStyleDoc (
316 Outputable.mkCodeStyle Outputable.AsmStyle) d
317 pprCLabel_asm l = asmSDoc (pprCLabel l)
322 needImportedSymbols = True
324 -- We don't need to declare any offset tables
325 pprGotDeclaration = Pretty.empty
327 -- On Darwin, we have to generate our own stub code for lazy binding..
328 -- There are two versions, one for PIC and one for non-PIC.
329 pprImportedSymbol importedLbl
330 | Just (CodeStub, lbl) <- dynamicLinkerLabelInfo importedLbl
334 ptext SLIT(".symbol_stub"),
335 ptext SLIT("L") <> pprCLabel_asm lbl <> ptext SLIT("$stub:"),
336 ptext SLIT("\t.indirect_symbol") <+> pprCLabel_asm lbl,
337 ptext SLIT("\tlis r11,ha16(L") <> pprCLabel_asm lbl
338 <> ptext SLIT("$lazy_ptr)"),
339 ptext SLIT("\tlwz r12,lo16(L") <> pprCLabel_asm lbl
340 <> ptext SLIT("$lazy_ptr)(r11)"),
341 ptext SLIT("\tmtctr r12"),
342 ptext SLIT("\taddi r11,r11,lo16(L") <> pprCLabel_asm lbl
343 <> ptext SLIT("$lazy_ptr)"),
348 ptext SLIT(".section __TEXT,__picsymbolstub1,")
349 <> ptext SLIT("symbol_stubs,pure_instructions,32"),
350 ptext SLIT("\t.align 2"),
351 ptext SLIT("L") <> pprCLabel_asm lbl <> ptext SLIT("$stub:"),
352 ptext SLIT("\t.indirect_symbol") <+> pprCLabel_asm lbl,
353 ptext SLIT("\tmflr r0"),
354 ptext SLIT("\tbcl 20,31,L0$") <> pprCLabel_asm lbl,
355 ptext SLIT("L0$") <> pprCLabel_asm lbl <> char ':',
356 ptext SLIT("\tmflr r11"),
357 ptext SLIT("\taddis r11,r11,ha16(L") <> pprCLabel_asm lbl
358 <> ptext SLIT("$lazy_ptr-L0$") <> pprCLabel_asm lbl <> char ')',
359 ptext SLIT("\tmtlr r0"),
360 ptext SLIT("\tlwzu r12,lo16(L") <> pprCLabel_asm lbl
361 <> ptext SLIT("$lazy_ptr-L0$") <> pprCLabel_asm lbl
362 <> ptext SLIT(")(r11)"),
363 ptext SLIT("\tmtctr r12"),
367 ptext SLIT(".lazy_symbol_pointer"),
368 ptext SLIT("L") <> pprCLabel_asm lbl <> ptext SLIT("$lazy_ptr:"),
369 ptext SLIT("\t.indirect_symbol") <+> pprCLabel_asm lbl,
370 ptext SLIT("\t.long dyld_stub_binding_helper")
373 -- We also have to declare our symbol pointers ourselves:
374 | Just (SymbolPtr, lbl) <- dynamicLinkerLabelInfo importedLbl
376 ptext SLIT(".non_lazy_symbol_pointer"),
377 char 'L' <> pprCLabel_asm lbl <> ptext SLIT("$non_lazy_ptr:"),
378 ptext SLIT("\t.indirect_symbol") <+> pprCLabel_asm lbl,
379 ptext SLIT("\t.long\t0")
384 #elif linux_TARGET_OS && !powerpc32_TARGET_ARCH
388 -- In theory, we don't need to generate any stubs or symbol pointers
389 -- by hand for Linux.
391 -- Reality differs from this in two areas.
393 -- 1) If we just use a dynamically imported symbol directly in a read-only
394 -- section of the main executable (as GCC does), ld generates R_*_COPY
395 -- relocations, which are fundamentally incompatible with reversed info
396 -- tables. Therefore, we need a table of imported addresses in a writable
398 -- The "official" GOT mechanism (label@got) isn't intended to be used
399 -- in position dependent code, so we have to create our own "fake GOT"
400 -- when not opt_PCI && not opt_Static.
402 -- 2) PowerPC Linux is just plain broken.
403 -- While it's theoretically possible to use GOT offsets larger
404 -- than 16 bit, the standard crt*.o files don't, which leads to
405 -- linker errors as soon as the GOT size exceeds 16 bit.
406 -- Also, the assembler doesn't support @gotoff labels.
407 -- In order to be able to use a larger GOT, we have to circumvent the
408 -- entire GOT mechanism and do it ourselves (this is also what GCC does).
411 -- When needImportedSymbols is defined,
412 -- the NCG will keep track of all DynamicLinkerLabels it uses
413 -- and output each of them using pprImportedSymbol.
414 #if powerpc_TARGET_ARCH
415 -- PowerPC Linux: -fPIC or -dynamic
416 needImportedSymbols = opt_PIC || not opt_Static
418 -- i386 (and others?): -dynamic but not -fPIC
419 needImportedSymbols = not opt_Static && not opt_PIC
423 -- The label used to refer to our "fake GOT" from
424 -- position-independent code.
425 gotLabel = mkForeignLabel -- HACK: it's not really foreign
426 FSLIT(".LCTOC1") Nothing False
429 -- Output whatever needs to be output once per .s file.
430 -- The .LCTOC1 label is defined to point 32768 bytes into the table,
431 -- to make the most of the PPC's 16-bit displacements.
432 -- Only needed for PIC.
435 | not opt_PIC = Pretty.empty
437 ptext SLIT(".section \".got2\",\"aw\""),
438 ptext SLIT(".LCTOC1 = .+32768")
441 -- We generate one .long literal for every symbol we import;
442 -- the dynamic linker will relocate those addresses.
444 pprImportedSymbol importedLbl
445 | Just (SymbolPtr, lbl) <- dynamicLinkerLabelInfo importedLbl
447 ptext SLIT(".section \".got2\", \"aw\""),
448 ptext SLIT(".LC_") <> pprCLabel_asm lbl <> char ':',
449 ptext SLIT("\t.long") <+> pprCLabel_asm lbl
452 -- PLT code stubs are generated automatically be the dynamic linker.
457 -- For all other currently supported platforms, we don't need to do
460 needImportedSymbols = False
461 pprGotDeclaration = Pretty.empty
462 pprImportedSymbol _ = empty
465 -- -------------------------------------------------------------------
467 -- Generate code to calculate the address that should be put in the
468 -- PIC base register.
469 -- This is called by MachCodeGen for every CmmProc that accessed the
470 -- PIC base register. It adds the appropriate instructions to the
471 -- top of the CmmProc.
473 -- It is assumed that the first NatCmmTop in the input list is a Proc
474 -- and the rest are CmmDatas.
476 initializePicBase :: Reg -> [NatCmmTop] -> NatM [NatCmmTop]
478 #if powerpc_TARGET_ARCH && darwin_TARGET_OS
480 -- Darwin is simple: just fetch the address of a local label.
481 initializePicBase picReg (CmmProc info lab params blocks : statics)
482 = return (CmmProc info lab params (b':tail blocks) : statics)
483 where BasicBlock bID insns = head blocks
484 b' = BasicBlock bID (FETCHPC picReg : insns)
486 #elif powerpc_TARGET_ARCH && linux_TARGET_OS
488 -- Get a pointer to our own fake GOT, which is defined on a per-module basis.
489 -- This is exactly how GCC does it, and it's quite horrible:
490 -- We first fetch the address of a local label (mkPicBaseLabel).
491 -- Then we add a 16-bit offset to that to get the address of a .long that we
492 -- define in .text space right next to the proc. This .long literal contains
493 -- the (32-bit) offset from our local label to our global offset table
494 -- (.LCTOC1 aka gotOffLabel).
495 initializePicBase picReg
496 (CmmProc info lab params blocks : statics)
498 gotOffLabel <- getNewLabelNat
499 tmp <- getNewRegNat wordRep
501 gotOffset = CmmData Text [
502 CmmDataLabel gotOffLabel,
503 CmmStaticLit (CmmLabelDiffOff gotLabel
507 offsetToOffset = ImmConstantDiff (ImmCLbl gotOffLabel)
508 (ImmCLbl mkPicBaseLabel)
509 BasicBlock bID insns = head blocks
510 b' = BasicBlock bID (FETCHPC picReg
512 (AddrRegImm picReg offsetToOffset)
513 : ADD picReg picReg (RIReg tmp)
515 return (CmmProc info lab params (b' : tail blocks) : gotOffset : statics)
516 #elif i386_TARGET_ARCH && linux_TARGET_OS
518 -- We cheat a bit here by defining a pseudo-instruction named FETCHGOT
519 -- which pretty-prints as:
522 -- addl __GLOBAL_OFFSET_TABLE__+.-1b, %picReg
525 initializePicBase picReg (CmmProc info lab params blocks : statics)
526 = return (CmmProc info lab params (b':tail blocks) : statics)
527 where BasicBlock bID insns = head blocks
528 b' = BasicBlock bID (FETCHGOT picReg : insns)
531 initializePicBase picReg proc = panic "initializePicBase"
533 -- mingw32_TARGET_OS: not needed, won't be called
535 -- i386_TARGET_ARCH && darwin_TARGET_OS:
536 -- (just for completeness ;-)