1 %************************************************************************
3 \section[Driver-asm-fiddling]{Fiddling with assembler files}
5 %************************************************************************
10 Utterly stomp out C functions' prologues and epilogues; i.e., the
11 stuff to do with the C stack.
13 Any other required tidying up.
16 General note [chak]: Many regexps are very fragile because they rely on white
17 space being in the right place. This caused trouble with gcc 2.95 (at least
18 on Linux), where the use of white space in .s files generated by gcc suddenly
19 changed. To guarantee compatibility across different versions of gcc, make
20 sure (at least on i386-.*-linux) that regexps tolerate varying amounts of white
21 space between an assembler statement and its arguments as well as after a the
22 comma separating multiple arguments.
24 \emph{For the time being, I have corrected the regexps for i386-.*-linux. I
25 didn't touch all the regexps for other i386 platforms, as I don't have
26 a box to test these changes.}
31 The HP linker is very picky about symbols being in the appropriate
32 space (code vs. data). When we mangle the threaded code to put the
33 info tables just prior to the code, they wind up in code space
34 rather than data space. This means that references to *_info from
35 un-mangled parts of the RTS (e.g. unthreaded GC code) get
36 unresolved symbols. Solution: mini-mangler for .c files on HP. I
37 think this should really be triggered in the driver by a new -rts
38 option, so that user code doesn't get mangled inappropriately.
40 With reversed tables, jumps are to the _info label rather than to
41 the _entry label. The _info label is just an address in code
42 space, rather than an entry point with the descriptive blob we
43 talked about yesterday. As a result, you can't use the call-style
44 JMP_ macro. However, some JMP_ macros take _info labels as targets
45 and some take code entry points within the RTS. The latter won't
46 work with the goto-style JMP_ macro. Sigh. Solution: Use the goto
47 style JMP_ macro, and mangle some more assembly, changing all
48 "RP'literal" and "LP'literal" references to "R'literal" and
49 "L'literal," so that you get the real address of the code, rather
50 than the descriptive blob. Also change all ".word P%literal"
51 entries in info tables and vector tables to just ".word literal,"
52 for the same reason. Advantage: No more ridiculous call sequences.
55 %************************************************************************
57 \subsection{Top-level code}
59 %************************************************************************
62 $TargetPlatform = $TARGETPLATFORM;
64 ($Pgm = $0) =~ s|.*/||;
68 if ( $TargetPlatform =~ /^i386-/ ) {
72 $StolenX86Regs = $ARGV[2];
76 &mangle_asm($ifile,$ofile);
81 %************************************************************************
83 \subsection{Constants for various architectures}
85 %************************************************************************
88 sub init_TARGET_STUFF {
90 #--------------------------------------------------------#
91 if ( $TargetPlatform =~ /^alpha-.*-.*/ ) {
93 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
94 $T_US = ''; # _ if symbols have an underscore on the front
95 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
96 $T_CONST_LBL = '^\$L?C(\d+):$'; # regexp for what such a lbl looks like
99 $T_MOVE_DIRVS = '^(\s*(\$.*\.\.ng:|\.align\s+\d+|\.(globl|ent)\s+\S+|\#.*|\.(file|loc)\s+\S+\s+\S+|\.text|\.r?data)\n)';
100 $T_COPY_DIRVS = '^\s*(\$.*\.\.ng:|\#|\.(file|globl|ent|loc))';
102 $T_DOT_WORD = '\.(long|quad|byte|word)';
103 $T_DOT_GLOBAL = '^\t\.globl';
104 $T_HDR_literal = "\.rdata\n\t\.align 3\n";
105 $T_HDR_misc = "\.text\n\t\.align 3\n";
106 $T_HDR_data = "\.data\n\t\.align 3\n";
107 $T_HDR_rodata = "\.rdata\n\t\.align 3\n";
108 $T_HDR_closure = "\.data\n\t\.align 3\n";
109 $T_HDR_info = "\.text\n\t\.align 3\n";
110 $T_HDR_entry = "\.text\n\t\.align 3\n";
111 $T_HDR_vector = "\.text\n\t\.align 3\n";
113 #--------------------------------------------------------#
114 } elsif ( $TargetPlatform =~ /^hppa/ ) {
116 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
117 $T_US = ''; # _ if symbols have an underscore on the front
118 $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP
119 $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like
122 $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)';
123 $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)';
125 $T_DOT_WORD = '\.(blockz|word|half|byte)';
126 $T_DOT_GLOBAL = '^\s+\.EXPORT';
127 $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n";
128 $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
129 $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
130 $T_HDR_rodata = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
131 $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n";
132 $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
133 $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
134 $T_HDR_vector = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n";
136 #--------------------------------------------------------#
137 } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|nextstep3|cygwin32|mingw32)$/ ) {
138 # NeXT added but not tested. CaS
140 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
141 $T_US = '_'; # _ if symbols have an underscore on the front
142 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
143 $T_CONST_LBL = '^LC(\d+):$';
145 $T_X86_PRE_LLBL_PAT = 'L';
146 $T_X86_PRE_LLBL = 'L';
147 $T_X86_BADJMP = '^\tjmp [^L\*]';
149 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*|\.lcomm.*)\n)';
150 $T_COPY_DIRVS = '\.(globl|stab|lcomm)';
151 $T_DOT_WORD = '\.(long|word|value|byte|space)';
152 $T_DOT_GLOBAL = '\.globl';
153 $T_HDR_literal = "\.text\n\t\.align 2\n";
154 $T_HDR_misc = "\.text\n\t\.align 2,0x90\n";
155 $T_HDR_data = "\.data\n\t\.align 2\n";
156 $T_HDR_rodata = "\.text\n\t\.align 2\n";
157 $T_HDR_closure = "\.data\n\t\.align 2\n";
158 $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding
159 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
160 $T_HDR_vector = "\.text\n\t\.align 2\n"; # NB: requires padding
162 #--------------------------------------------------------#
163 } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|gnu|freebsd|netbsd|openbsd|kfreebsdgnu)$/ ) {
165 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
166 $T_US = ''; # _ if symbols have an underscore on the front
167 $T_PRE_APP = # regexp that says what comes before APP/NO_APP
168 ($TargetPlatform =~ /-(linux|gnu|freebsd|netbsd|openbsd)$/) ? '#' : '/' ;
169 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
171 $T_X86_PRE_LLBL_PAT = '\.L';
172 $T_X86_PRE_LLBL = '.L';
173 $T_X86_BADJMP = '^\tjmp\s+[^\.\*]';
175 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s.*|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
176 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
178 $T_DOT_WORD = '\.(long|value|word|byte|zero)';
179 $T_DOT_GLOBAL = '\.globl';
180 $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11)
181 $T_HDR_misc = "\.text\n\t\.align 4\n";
182 $T_HDR_data = "\.data\n\t\.align 4\n";
183 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 4\n";
184 $T_HDR_closure = "\.data\n\t\.align 4\n";
185 $T_HDR_info = "\.text\n\t\.align 4\n";
186 $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?)
187 $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding
189 #--------------------------------------------------------#
190 } elsif ( $TargetPlatform =~ /^ia64-.*-linux$/ ) {
192 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
193 $T_US = ''; # _ if symbols have an underscore on the front
195 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
198 $T_MOVE_DIRVS = '^(\s*\.(global|proc|pred\.safe_across_calls|text|data|section|subsection|align|size|type|ident)\s+.*\n)';
199 $T_COPY_DIRVS = '\.(global|proc)';
201 $T_DOT_WORD = '\.(long|value|byte|zero)';
202 $T_DOT_GLOBAL = '\.global';
203 $T_HDR_literal = "\.section\t\.rodata\n";
204 $T_HDR_misc = "\.text\n\t\.align 8\n";
205 $T_HDR_data = "\.data\n\t\.align 8\n";
206 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
207 $T_HDR_closure = "\.data\n\t\.align 8\n";
208 $T_HDR_info = "\.text\n\t\.align 8\n";
209 $T_HDR_entry = "\.text\n\t\.align 16\n";
210 $T_HDR_vector = "\.text\n\t\.align 8\n";
212 #--------------------------------------------------------#
213 } elsif ( $TargetPlatform =~ /^x86_64-.*-(linux|openbsd)$/ ) {
215 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
216 $T_US = ''; # _ if symbols have an underscore on the front
218 $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like
221 $T_MOVE_DIRVS = '^(\s*\.(globl|text|data|section|align|size|type|ident|local)\s+.*\n)';
222 $T_COPY_DIRVS = '\.(globl|type|size|local)';
224 $T_DOT_WORD = '\.(quad|long|value|byte|zero)';
225 $T_DOT_GLOBAL = '\.global';
227 $T_HDR_literal16 = "\.section\t\.rodata.cst16\n\t.align 16\n";
228 $T_HDR_literal = "\.section\t\.rodata\n";
230 $T_HDR_misc = "\.text\n\t\.align 8\n";
231 $T_HDR_data = "\.data\n\t\.align 8\n";
232 $T_HDR_rodata = "\.section\t\.rodata\n\t\.align 8\n";
234 # the assembler on x86_64/Linux refuses to generate code for
236 # where x is in the text section and y in the rodata section.
237 # It works if y is in the text section, though. This is probably
238 # going to cause difficulties for PIC, I imagine.
239 $T_HDR_relrodata= "\.text\n\t\.align 8\n";
241 $T_HDR_closure = "\.data\n\t\.align 8\n";
242 $T_HDR_info = "\.text\n\t\.align 8\n";
243 $T_HDR_entry = "\.text\n\t\.align 8\n";
244 $T_HDR_vector = "\.text\n\t\.align 8\n";
246 #--------------------------------------------------------#
247 } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) {
249 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
250 $T_US = '_'; # _ if symbols have an underscore on the front
251 $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP
252 $T_CONST_LBL = '^LC(\d+):$';
255 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.proc\s+\d+|\.const|\.cstring|\.globl\s+\S+|\.text|\.data|\.even|\.stab[^n].*)\n)';
256 $T_COPY_DIRVS = '\.(globl|proc|stab)';
258 $T_DOT_WORD = '\.long';
259 $T_DOT_GLOBAL = '\.globl';
260 $T_HDR_literal = "\.text\n\t\.even\n";
261 $T_HDR_misc = "\.text\n\t\.even\n";
262 $T_HDR_data = "\.data\n\t\.even\n";
263 $T_HDR_rodata = "\.text\n\t\.even\n";
264 $T_HDR_closure = "\.data\n\t\.even\n";
265 $T_HDR_info = "\.text\n\t\.even\n";
266 $T_HDR_entry = "\.text\n\t\.even\n";
267 $T_HDR_vector = "\.text\n\t\.even\n";
269 #--------------------------------------------------------#
270 } elsif ( $TargetPlatform =~ /^mips-.*/ ) {
272 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
273 $T_US = ''; # _ if symbols have an underscore on the front
274 $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP
275 $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like
278 $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+|\.(globl|ent)\s+\S+|\.text|\.r?data)\n)';
279 $T_COPY_DIRVS = '\.(globl|ent)';
281 $T_DOT_WORD = '\.word';
282 $T_DOT_GLOBAL = '^\t\.globl';
283 $T_HDR_literal = "\t\.rdata\n\t\.align 2\n";
284 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
285 $T_HDR_data = "\t\.data\n\t\.align 2\n";
286 $T_HDR_rodata = "\t\.rdata\n\t\.align 2\n";
287 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
288 $T_HDR_info = "\t\.text\n\t\.align 2\n";
289 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
290 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
292 #--------------------------------------------------------#
293 } elsif ( $TargetPlatform =~ /^powerpc-apple-darwin.*/ ) {
294 # Apple PowerPC Darwin/MacOS X.
295 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
296 $T_US = '_'; # _ if symbols have an underscore on the front
297 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
298 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
301 $T_MOVE_DIRVS = '^(\s*(\.align \d+|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*|\.lcomm.*)\n)';
302 $T_COPY_DIRVS = '\.(globl|lcomm)';
304 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
305 $T_DOT_GLOBAL = '\.globl';
306 $T_HDR_toc = "\.toc\n";
307 $T_HDR_literal = "\t\.const\n\t\.align 2\n";
308 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
309 $T_HDR_data = "\t\.data\n\t\.align 2\n";
310 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
311 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
312 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
313 $T_HDR_info = "\t\.text\n\t\.align 2\n";
314 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
315 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
317 #--------------------------------------------------------#
318 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ ) {
319 # Apple PowerPC Darwin/MacOS X.
320 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
321 $T_US = '_'; # _ if symbols have an underscore on the front
322 $T_PRE_APP = 'DOESNT APPLY'; # regexp that says what comes before APP/NO_APP
323 $T_CONST_LBL = '^\LC\d+:'; # regexp for what such a lbl looks like
325 $T_X86_PRE_LLBL_PAT = 'L';
326 $T_X86_PRE_LLBL = 'L';
327 $T_X86_BADJMP = '^\tjmp [^L\*]';
329 $T_MOVE_DIRVS = '^(\s*(\.align \d+|\.text|\.data|\.const_data|\.cstring|\.non_lazy_symbol_pointer|\.const|\.static_const|\.literal4|\.literal8|\.static_data|\.globl \S+|\.section .*|\.lcomm.*)\n)';
330 $T_COPY_DIRVS = '\.(globl|lcomm)';
332 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
333 $T_DOT_GLOBAL = '\.globl';
334 $T_HDR_toc = "\.toc\n";
335 $T_HDR_literal16= "\t\.literal8\n\t\.align 4\n";
336 $T_HDR_literal = "\t\.const\n\t\.align 4\n";
337 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
338 $T_HDR_data = "\t\.data\n\t\.align 2\n";
339 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
340 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
341 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
342 $T_HDR_info = "\t\.text\n\t\.align 2\n";
343 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
344 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
346 #--------------------------------------------------------#
347 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ ) {
349 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
350 $T_US = ''; # _ if symbols have an underscore on the front
351 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
352 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
355 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,\s*0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
356 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
358 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
359 $T_DOT_GLOBAL = '\.globl';
360 $T_HDR_toc = "\.toc\n";
361 $T_HDR_literal = "\t\.section\t.rodata\n\t\.align 2\n";
362 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
363 $T_HDR_data = "\t\.data\n\t\.align 2\n";
364 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
365 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
366 $T_HDR_info = "\t\.text\n\t\.align 2\n";
367 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
368 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
370 #--------------------------------------------------------#
371 } elsif ( $TargetPlatform =~ /^powerpc64-.*-linux/ ) {
373 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
374 $T_US = '\.'; # _ if symbols have an underscore on the front
375 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
376 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
379 $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,\s*0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.size\s+\S+\s*,\s*\d+|\.ident.*|\.local.*)\n)';
380 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
382 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
383 $T_DOT_GLOBAL = '\.globl';
384 $T_HDR_toc = "\.toc\n";
385 $T_HDR_literal = "\t\.section\t\".toc\",\"aw\"\n";
386 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
387 $T_HDR_data = "\t\.data\n\t\.align 2\n";
388 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
389 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
390 $T_HDR_info = "\t\.text\n\t\.align 2\n";
391 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
392 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
394 #--------------------------------------------------------#
395 } elsif ( $TargetPlatform =~ /^sparc-.*-(solaris2|openbsd)/ ) {
397 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
398 $T_US = ''; # _ if symbols have an underscore on the front
399 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
400 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
403 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\s+\.local\s+\S+|\.text|\.data|\.stab.*|\s*\.section.*|\s+\.type.*|\s+\.size.*)\n)';
404 $T_COPY_DIRVS = '\.(global|local|proc|stab)';
406 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
407 $T_DOT_GLOBAL = '^\t\.global';
408 $T_HDR_literal = "\.text\n\t\.align 8\n";
409 $T_HDR_misc = "\.text\n\t\.align 4\n";
410 $T_HDR_data = "\.data\n\t\.align 8\n";
411 $T_HDR_rodata = "\.text\n\t\.align 4\n";
412 $T_HDR_closure = "\.data\n\t\.align 4\n";
413 $T_HDR_info = "\.text\n\t\.align 4\n";
414 $T_HDR_entry = "\.text\n\t\.align 4\n";
415 $T_HDR_vector = "\.text\n\t\.align 4\n";
417 #--------------------------------------------------------#
418 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
420 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
421 $T_US = '_'; # _ if symbols have an underscore on the front
422 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
423 $T_CONST_LBL = '^LC(\d+):$';
426 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
427 $T_COPY_DIRVS = '\.(global|proc|stab)';
429 $T_DOT_WORD = '\.word';
430 $T_DOT_GLOBAL = '^\t\.global';
431 $T_HDR_literal = "\.text\n\t\.align 8\n";
432 $T_HDR_misc = "\.text\n\t\.align 4\n";
433 $T_HDR_data = "\.data\n\t\.align 8\n";
434 $T_HDR_rodata = "\.text\n\t\.align 4\n";
435 $T_HDR_closure = "\.data\n\t\.align 4\n";
436 $T_HDR_info = "\.text\n\t\.align 4\n";
437 $T_HDR_entry = "\.text\n\t\.align 4\n";
438 $T_HDR_vector = "\.text\n\t\.align 4\n";
440 #--------------------------------------------------------#
441 } elsif ( $TargetPlatform =~ /^sparc-.*-linux/ ) {
442 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
443 $T_US = ''; # _ if symbols have an underscore on the front
444 $T_PRE_APP = '#'; # regexp that says what comes before APP/NO_APP
445 # Probably doesn't apply anyway
446 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
449 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.seg|\.stab.*|\s+?\.section.*|\s+\.type.*|\s+\.size.*)\n)';
450 $T_COPY_DIRVS = '\.(global|globl|proc|stab)';
452 $T_DOT_WORD = '\.(long|word|nword|xword|byte|half|short|skip|uahalf|uaword)';
453 $T_DOT_GLOBAL = '^\t\.global';
454 $T_HDR_literal = "\.text\n\t\.align 8\n";
455 $T_HDR_misc = "\.text\n\t\.align 4\n";
456 $T_HDR_data = "\.data\n\t\.align 8\n";
457 $T_HDR_rodata = "\.text\n\t\.align 4\n";
458 $T_HDR_closure = "\.data\n\t\.align 4\n";
459 $T_HDR_info = "\.text\n\t\.align 4\n";
460 $T_HDR_entry = "\.text\n\t\.align 4\n";
461 $T_HDR_vector = "\.text\n\t\.align 4\n";
463 #--------------------------------------------------------#
465 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
469 if($T_HDR_relrodata eq "") {
471 # relrodata defaults to rodata.
472 $T_HDR_relrodata = $T_HDR_rodata;
476 print STDERR "T_STABBY: $T_STABBY\n";
477 print STDERR "T_US: $T_US\n";
478 print STDERR "T_PRE_APP: $T_PRE_APP\n";
479 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
480 print STDERR "T_POST_LBL: $T_POST_LBL\n";
481 if ( $TargetPlatform =~ /^i386-/ ) {
482 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
483 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
484 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
486 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
487 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
488 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
489 print STDERR "T_HDR_literal: $T_HDR_literal\n";
490 print STDERR "T_HDR_misc: $T_HDR_misc\n";
491 print STDERR "T_HDR_data: $T_HDR_data\n";
492 print STDERR "T_HDR_rodata: $T_HDR_rodata\n";
493 print STDERR "T_HDR_closure: $T_HDR_closure\n";
494 print STDERR "T_HDR_info: $T_HDR_info\n";
495 print STDERR "T_HDR_entry: $T_HDR_entry\n";
496 print STDERR "T_HDR_vector: $T_HDR_vector\n";
502 %************************************************************************
504 \subsection{Mangle away}
506 %************************************************************************
510 local($in_asmf, $out_asmf) = @_;
512 # multi-line regexp matching:
517 &init_TARGET_STUFF();
518 &init_FUNNY_THINGS();
520 open(INASM, "< $in_asmf")
521 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
522 open(OUTASM,"> $out_asmf")
523 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
525 # read whole file, divide into "chunks":
526 # record some info about what we've found...
528 @chk = (); # contents of the chunk
529 $numchks = 0; # number of them
530 @chkcat = (); # what category of thing in each chunk
531 @chksymb = (); # what symbol(base) is defined in this chunk
532 %entrychk = (); # ditto, its entry code
533 %closurechk = (); # ditto, the (static) closure
534 %srtchk = (); # ditto, its SRT (for top-level things)
535 %infochk = (); # given a symbol base, say what chunk its info tbl is in
536 %vectorchk = (); # ditto, return vector table
537 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
539 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
542 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
543 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
544 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
545 next if /^\t\.def.*endef$/;
546 next if /${T_PRE_APP}(NO_)?APP/o;
547 next if /^;/ && $TargetPlatform =~ /^hppa/;
549 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
551 if ( $TargetPlatform =~ /^mips-/
552 && /^\t\.(globl\S+\.text|comm\t)/ ) {
553 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
554 # Treat .comm variables as data. These show up in two (known) places:
556 # - the module_registered variable used in the __stginit fragment.
557 # even though these are declared static and initialised, gcc 3.3
558 # likes to make them .comm, presumably to save space in the
561 # - global variables used to pass arguments from C to STG in
562 # a foreign export. (is this still true? --SDM)
564 } elsif ( /^\t\.comm.*$/ ) {
566 $chkcat[$i] = 'data';
569 # Labels ending "_str": these are literal strings.
570 } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
572 $chkcat[$i] = 'rodata';
574 } elsif ( $TargetPlatform =~ /-darwin/
575 && (/^\s*\.subsections_via_symbols/
576 ||/^\s*\.no_dead_strip.*/)) {
577 # Don't allow Apple's linker to do any dead-stripping of symbols
578 # in this file, because it will mess up info-tables in mangled
580 # The .no_dead_strip directives are actually put there by
581 # the gcc3 "used" attribute on entry points.
583 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
584 /^\s*\.picsymbol_stub/
585 || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
586 || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
587 || /^\s*\.symbol_stub/
588 || /^\s*\.section __TEXT,__symbol_stub\d,.*/
589 || /^\s*\.section __TEXT,__symbolstub\d,.*/
590 || /^\s*\.lazy_symbol_pointer/
591 || /^\s*\.non_lazy_symbol_pointer/
592 || /^\s*\.section __IMPORT.*/))
595 $chkcat[$i] = 'dyld';
599 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
600 { # non_lazy_symbol_ptrs that point to local symbols
602 $chkcat[$i] = 'dyld';
605 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^L_.*:$/)
606 { # non_lazy_symbol_ptrs that point to local symbols
607 $chk[++$i] = $dyld_section . $_;
608 $chkcat[$i] = 'dyld';
611 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
612 # duplicated from the bottom
616 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
617 # Alphas: Local labels not to be confused with new chunks
619 # NB: all the rest start with a non-space
621 } elsif ( $TargetPlatform =~ /^mips-/
622 && /^\d+:/ ) { # a funny-looking very-local label
625 } elsif ( /$T_CONST_LBL/o ) {
627 $chkcat[$i] = 'literal';
630 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
632 $chkcat[$i] = 'splitmarker';
635 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
638 $chkcat[$i] = 'infotbl';
639 $chksymb[$i] = $symb;
641 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
643 $infochk{$symb} = $i;
645 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
647 $chkcat[$i] = 'entry';
652 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
654 $chkcat[$i] = 'closure';
657 $closurechk{$1} = $i;
659 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
666 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
668 $chkcat[$i] = 'data';
671 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
673 $chkcat[$i] = 'data';
676 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
679 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
680 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
681 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
682 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
685 $chkcat[$i] = 'data';
688 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
693 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
694 # all CC_ symbols go in the data section...
696 $chkcat[$i] = 'data';
699 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
701 $chkcat[$i] = 'misc';
703 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
705 $chkcat[$i] = 'vector';
710 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
711 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
712 # Some Solaris system headers contain function definitions (as
713 # opposed to mere prototypes), which end up in the .hc file when
714 # a Haskell module foreign imports the corresponding system
715 # functions (most notably stat()). We put them into the text
716 # segment. Note that this currently does not extend to function
717 # names starting with an underscore.
720 $chkcat[$i] = 'misc';
723 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
724 # To handle PIC on Darwin/x86, we need to appropriately pass through
725 # the get_pc_thunk functions. The need to be put into a special section
726 # marked as coalesced (otherwise the .weak_definition doesn't work
729 $chkcat[$i] = 'get_pc_thunk';
732 } elsif ( /^${T_US}[A-Za-z0-9_]/o
733 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
735 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
736 || ! /^\.L\d+:$/ ) ) {
743 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
744 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
745 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
746 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
747 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
748 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
749 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
750 || /^_uname:/o # x86/Solaris2
753 $chkcat[$i] = 'misc';
755 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
756 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
759 $chkcat[$i] = 'relrodata';
762 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
763 $chkcat[$i] = 'unknown';
766 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
767 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
768 # table "by hand". Be sure to copy it over.
769 # Note that this label and all entries in the table should actually
770 # go into the .got2 section, but it isn't easy to distinguish them
771 # from other constant literals (.LC\d+), so we just put everything
774 $chkcat[$i] = 'literal';
775 $chksymb[$i] = 'LCTOC1';
776 } else { # simple line (duplicated at the top)
781 $numchks = $#chk + 1;
783 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
784 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
787 # the division into chunks is imperfect;
788 # we throw some things over the fence into the next
791 # also, there are things we would like to know
792 # about the whole module before we start spitting
795 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
796 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
798 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
800 # Alphas: NB: we start meddling at chunk 1, not chunk 0
801 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
802 # spits a ".quad 0" in after the very first ".rdata"; we
803 # detect this special case (tossing the ".quad 0")!
804 local($magic_rdata_seen) = 0;
806 # HPPAs, MIPSen: also start medding at chunk 1
808 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
809 $c = $chk[$i]; # convenience copy
811 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
813 # toss all prologue stuff; HPPA is pretty weird
815 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
817 # be slightly paranoid to make sure there's
818 # nothing surprising in there
819 if ( $c =~ /--- BEGIN ---/ ) {
820 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
822 # remove junk whitespace around the split point
826 if ($TargetPlatform =~ /^i386-/) {
827 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
829 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
833 # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
834 # when compiling PrimOps.cmm with -optc-O2:
837 # movl %ecx, 16(%esp)
838 # movl %edx, 20(%esp)
839 # but then the code of the function doesn't assume
840 # anything about the contnets of these stack locations.
841 # I think it's to do with the use of inline functions for
842 # PK_Word64() and friends, where gcc is initialising the
843 # contents of the struct to zero, and failing to optimise
844 # away the initialisation. Let's live dangerously and
845 # discard these initalisations.
847 $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
848 $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
849 $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
850 $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
851 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
852 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
854 if ($TargetPlatform =~ /^i386-apple-darwin/) {
856 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
858 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
859 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
860 $p =~ s/^\"?L\d+\$pb\"?:\n//;
862 if ($pcrel_reg eq "bx") {
863 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
864 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
868 } elsif ($TargetPlatform =~ /^x86_64-/) {
869 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
870 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
871 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
873 } elsif ($TargetPlatform =~ /^ia64-/) {
874 $p =~ s/^\t\.prologue .*\n//;
875 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, 3[12], \d+, 0\n//;
876 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
877 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
878 $p =~ s/^\t\.(mii|mmi)\n//g; # bundling is no longer sensible
879 $p =~ s/^\t;;\n//g; # discard stops
880 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
882 # GCC 3.3 saves r1 in the prologue, move this to the body
883 if ($p =~ /^\tmov r\d+ = r1\n/) {
887 } elsif ($TargetPlatform =~ /^m68k-/) {
888 $p =~ s/^\tlink a6,#-?\d.*\n//;
889 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
890 # The above showed up in the asm code,
891 # so I added it here.
892 # I hope it's correct.
894 $p =~ s/^\tmovel d2,sp\@-\n//;
895 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
896 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
897 } elsif ($TargetPlatform =~ /^mips-/) {
898 # the .frame/.mask/.fmask that we use is the same
899 # as that produced by GCC for miniInterpret; this
900 # gives GDB some chance of figuring out what happened
901 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
902 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
903 $p =~ s/^\t\.(mask|fmask).*\n//g;
904 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
905 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
906 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
907 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
908 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
909 $p =~ s/__FRAME__/$FRAME/;
910 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
912 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
914 $p =~ s/^\tmflr r0\n//;
915 $p =~ s/^\tbl saveFP # f\d+\n//;
916 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
917 $p =~ s/^\"?L\d+\$pb\"?:\n//;
918 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
919 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
920 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
921 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
922 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
923 $p =~ s/^\tbcl 20,31,L\d+\$pb\n//;
924 $p =~ s/^L\d+\$pb:\n//;
925 $p =~ s/^\tmflr r31\n//;
927 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
928 # under some circumstances, only when generating position dependent code.
929 # I have no idea why, and I don't think it is necessary, so let's toss it.
930 $p =~ s/^\tli r\d+,0\n//g;
931 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
932 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
933 $p =~ s/^\tmflr 0\n//;
934 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
935 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
936 $p =~ s/^\tstw r0,8\(1\)\n//;
937 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
938 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
940 # GCC's "large-model" PIC (-fPIC)
942 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
944 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
945 $p =~ s/^.LCF\d+:\n//;
946 $p =~ s/^\tmflr 30\n//;
947 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
948 $p =~ s/^\tadd 30,0,30\n//;
950 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
951 # under some circumstances, only when generating position dependent code.
952 # I have no idea why, and I don't think it is necessary, so let's toss it.
953 $p =~ s/^\tli \d+,0\n//g;
954 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
955 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
956 $p =~ s/^\tmr 31,1\n//;
957 $p =~ s/^\tmflr 0\n//;
958 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
959 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
960 $p =~ s/^\tstd r0,8\(1\)\n//;
961 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
962 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
964 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
965 # under some circumstances, only when generating position dependent code.
966 # I have no idea why, and I don't think it is necessary, so let's toss it.
967 $p =~ s/^\tli \d+,0\n//g;
968 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
970 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
973 # HWL HACK: dont die, just print a warning
974 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
975 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
977 # For PIC, we want to keep part of the prologue
978 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
979 # Darwin: load the current instruction pointer into register r31
980 $p .= "bcl 20,31,$pcrel_label\n";
981 $p .= "$pcrel_label:\n";
982 $p .= "\tmflr r31\n";
983 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
984 # Linux: load the GOT pointer into register 30
985 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
986 $p .= ".LCF$pcrel_label:\n";
988 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
989 $p .= "\tadd 30,0,30\n";
990 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
991 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
992 $p .= "$pcrel_label:\n";
995 # glue together what's left
1000 if ( $TargetPlatform =~ /^mips-/ ) {
1001 # MIPS: first, this basic sequence may occur "--- END ---" or not
1002 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1005 # toss all epilogue stuff; again, paranoidly
1006 if ( $c =~ /--- END ---/ ) {
1007 if (($r, $e) = split(/--- END ---/, $c)) {
1008 if ($TargetPlatform =~ /^i386-/) {
1010 $e =~ s/^\tpopl\s+\%edi\n//;
1011 $e =~ s/^\tpopl\s+\%esi\n//;
1012 $e =~ s/^\tpopl\s+\%edx\n//;
1013 $e =~ s/^\tpopl\s+\%ecx\n//;
1014 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1015 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1016 } elsif ($TargetPlatform =~ /^ia64-/) {
1017 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1018 $e =~ s/^\tmov b0 = r\d+\n//;
1019 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1020 $e =~ s/^\tbr\.ret\.sptk\.many b0\n//;
1021 $e =~ s/^\t\.(mii|mmi|mib)\n//g; # bundling is no longer sensible
1022 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1023 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1024 } elsif ($TargetPlatform =~ /^m68k-/) {
1025 $e =~ s/^\tunlk a6\n//;
1027 } elsif ($TargetPlatform =~ /^mips-/) {
1028 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1029 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1030 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1031 $e =~ s/^\tj\t\$31\n//;
1032 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1033 $e =~ s/^\taddi r1,r1,\d+\n//;
1034 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1035 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1036 $e =~ s/^\tmtlr r0\n//;
1038 $e =~ s/^\tb restFP ;.*\n//;
1039 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1040 $e =~ s/^\tmr 3,0\n//;
1041 $e =~ s/^\taddi 1,1,\d+\n//;
1042 $e =~ s/^\tld 0,16\(1\)\n//;
1043 $e =~ s/^\tmtlr 0\n//;
1045 # callee-save registers
1046 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1047 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1049 # get rid of the debug junk along with the blr
1050 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1052 # incase we missed it with the last one get the blr alone
1055 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1058 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1060 # glue together what's left
1062 $c =~ s/\n\t\n/\n/; # junk blank line
1066 # On SPARCs, we don't do --- BEGIN/END ---, we just
1067 # toss the register-windowing save/restore/ret* instructions
1068 # directly unless they've been generated by function definitions in header
1070 if ( $TargetPlatform =~ /^sparc-/ ) {
1071 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1072 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1074 # throw away PROLOGUE comments
1075 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1078 # On Alphas, the prologue mangling is done a little later (below)
1080 # toss all calls to __DISCARD__
1081 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1082 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1083 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1084 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1085 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1086 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1088 # IA64: mangle tailcalls into jumps here
1089 if ($TargetPlatform =~ /^ia64-/) {
1090 while ($c =~ s/^\tbr\.call\.sptk\.many b0 = (.*)\n(?:^\.L([0-9]*):\n)?(?:\t;;\n)?(?:\tmov r1 = r\d+\n)?(?:\t;;\n)?\t--- TAILCALL ---\n(?:\t;;\n\tbr \.L\d+\n)?/\tbr\.few $1\n/) {
1091 # Eek, the gcc optimiser is getting smarter... if we see a jump to the --- TAILCALL ---
1092 # marker then we reapply the substitution at the source sites
1093 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1097 # MIPS: that may leave some gratuitous asm macros around
1098 # (no harm done; but we get rid of them to be tidier)
1099 $c =~ s/^\t\.set\tnoreorder\n\t\.set\tnomacro\n\taddu\t(\S+)\n\t\.set\tmacro\n\t\.set\treorder\n/\taddu\t$1\n/
1100 if $TargetPlatform =~ /^mips-/;
1102 # toss stack adjustment after DoSparks
1103 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1104 if $TargetPlatform =~ /^m68k-/; # this looks old...
1106 if ( $TargetPlatform =~ /^alpha-/ &&
1107 ! $magic_rdata_seen &&
1108 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1109 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1110 $magic_rdata_seen = 1;
1113 # pick some end-things and move them to the next chunk
1115 # pin a funny end-thing on (for easier matching):
1116 $c .= 'FUNNY#END#THING';
1118 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1122 # on x86 we try not to copy any directives into a literal
1123 # chunk, rather we keep looking for the next real chunk. This
1124 # is because we get things like
1126 # .globl blah_closure
1132 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1134 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1137 if ( $j < $numchks ) {
1138 $chk[$j] = $to_move . $chk[$j];
1142 elsif ( $i < ($numchks - 1)
1143 && ( $to_move =~ /${T_COPY_DIRVS}/
1144 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
1145 $chk[$i + 1] = $to_move . $chk[$i + 1];
1146 # otherwise they're tossed
1149 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1152 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1154 # toss all prologue stuff, except for loading gp, and the ..ng address
1155 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1156 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1157 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1158 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1159 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1160 $c = $keep . "..ng:\n";
1162 print STDERR "malformed code block ($ent)?\n"
1165 $c .= "\t.prologue" . $r;
1169 $c =~ s/FUNNY#END#THING//;
1171 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1173 $chk[$i] = $c; # update w/ convenience copy
1176 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1177 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1180 if ( $TargetPlatform =~ /^alpha-/ ) {
1181 # print out the header stuff first
1182 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1183 print OUTASM $chk[0];
1185 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1186 print OUTASM $chk[0];
1188 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1189 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1191 # get rid of horrible "<dollar>Revision: .*$" strings
1192 local(@lines0) = split(/\n/, $chk[0]);
1194 while ( $z <= $#lines0 ) {
1195 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1198 while ( $z <= $#lines0 ) {
1200 last if $lines0[$z] =~ /[,\t]0x0$/;
1206 $chk[0] = join("\n", @lines0);
1207 $chk[0] =~ s/\n\n+/\n/;
1208 print OUTASM $chk[0];
1211 # print out all the literal strings next
1212 for ($i = 0; $i < $numchks; $i++) {
1213 if ( $chkcat[$i] eq 'literal' ) {
1215 # HACK: try to detect 16-byte constants and align them
1216 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1217 # aligned constants, and so does Darwin/x86.
1218 if ( $TargetPlatform =~ /^x86_64/
1219 || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1221 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1222 print OUTASM $T_HDR_literal16;
1224 print OUTASM $T_HDR_literal;
1227 print OUTASM $T_HDR_literal;
1230 print OUTASM $chk[$i];
1231 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1233 $chkcat[$i] = 'DONE ALREADY';
1237 # on the HPPA, print out all the bss next
1238 if ( $TargetPlatform =~ /^hppa/ ) {
1239 for ($i = 1; $i < $numchks; $i++) {
1240 if ( $chkcat[$i] eq 'bss' ) {
1241 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1242 print OUTASM $chk[$i];
1244 $chkcat[$i] = 'DONE ALREADY';
1249 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
1250 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1252 next if $chkcat[$i] eq 'DONE ALREADY';
1254 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1255 if ($chk[$i] ne '') {
1256 print OUTASM $T_HDR_misc;
1257 &print_doctored($chk[$i], 0);
1260 } elsif ( $chkcat[$i] eq 'toss' ) {
1261 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1263 } elsif ( $chkcat[$i] eq 'data' ) {
1264 if ($chk[$i] ne '') {
1265 print OUTASM $T_HDR_data;
1266 print OUTASM $chk[$i];
1269 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1270 # we can just re-constitute this one...
1271 # NB: we emit _three_ underscores no matter what,
1272 # so ghc-split doesn't have to care.
1273 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1275 } elsif ( $chkcat[$i] eq 'closure'
1276 || $chkcat[$i] eq 'srt'
1277 || $chkcat[$i] eq 'infotbl'
1278 || $chkcat[$i] eq 'entry') { # do them in that order
1279 $symb = $chksymb[$i];
1282 if ( defined($closurechk{$symb}) ) {
1283 print OUTASM $T_HDR_closure;
1284 print OUTASM $chk[$closurechk{$symb}];
1285 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1289 if ( defined($srtchk{$symb}) ) {
1290 print OUTASM $T_HDR_relrodata;
1291 print OUTASM $chk[$srtchk{$symb}];
1292 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1296 if ( defined($infochk{$symb}) ) {
1298 print OUTASM $T_HDR_info;
1299 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1301 # entry code will be put here!
1303 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1307 if ( defined($entrychk{$symb}) ) {
1309 $c = $chk[$entrychk{$symb}];
1311 # If this is an entry point with an info table,
1312 # eliminate the entry symbol and all directives involving it.
1313 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1315 foreach $l (split(/\n/,$c)) {
1316 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1318 # If we have .type/.size direrctives involving foo_entry,
1319 # then make them refer to foo_info instead. The information
1320 # in these directives is used by the cachegrind annotator,
1321 # so it is worthwhile keeping.
1322 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1323 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1327 next if $l =~ /^\s*\..*$symb.*\n?/;
1330 $c = join("\n",@o) . "\n";
1333 print OUTASM $T_HDR_entry;
1335 &print_doctored($c, 1); # NB: the 1!!!
1337 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1340 } elsif ( $chkcat[$i] eq 'vector' ) {
1341 $symb = $chksymb[$i];
1344 if ( defined($vectorchk{$symb}) ) {
1345 print OUTASM $T_HDR_vector;
1346 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1348 # direct return code will be put here!
1349 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1351 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1352 # Alphas: the commented nop is for the splitter, to ensure
1353 # that no module ends with a label as the very last
1354 # thing. (The linker will adjust the label to point
1355 # to the first code word of the next module linked in,
1356 # even if alignment constraints cause the label to move!)
1358 print OUTASM "\t# nop\n";
1361 } elsif ( $chkcat[$i] eq 'rodata' ) {
1362 print OUTASM $T_HDR_rodata;
1363 print OUTASM $chk[$i];
1364 $chkcat[$i] = 'DONE ALREADY';
1365 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1366 print OUTASM $T_HDR_relrodata;
1367 print OUTASM $chk[$i];
1368 $chkcat[$i] = 'DONE ALREADY';
1369 } elsif ( $chkcat[$i] eq 'toc' ) {
1370 # silly optimisation to print tocs, since they come in groups...
1371 print OUTASM $T_HDR_toc;
1373 while ($chkcat[$j] eq 'toc')
1374 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1377 print OUTASM $chk[$j];
1379 $chkcat[$j] = 'DONE ALREADY';
1383 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1384 # apple-darwin: dynamic linker stubs
1385 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1386 { # print them out unchanged, but remove the stubs for __DISCARD__
1387 print OUTASM $chk[$i];
1389 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1390 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1391 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1392 print OUTASM $chk[$i];
1394 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1398 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1401 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1402 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1407 sub hppa_mash_prologue { # OK, epilogue, too
1410 # toss all prologue stuff
1411 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1413 # Lie about our .CALLINFO
1414 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1421 # toss all epilogue stuff
1422 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1424 # Sorry; we moved the _info stuff to the code segment.
1425 s/_info,DATA/_info,CODE/g;
1432 sub print_doctored {
1433 local($_, $need_fallthru_patch) = @_;
1435 if ( $TargetPlatform =~ /^x86_64-/ ) {
1438 # movq -4(%ebp), %rax
1443 s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1444 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1447 if ( $TargetPlatform !~ /^i386-/
1448 || ! /^\t[a-z]/ # no instructions in here, apparently
1449 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1454 # OK, must do some x86 **HACKING**
1456 local($entry_patch) = '';
1457 local($exit_patch) = '';
1459 # gotta watch out for weird instructions that
1460 # invisibly smash various regs:
1461 # rep* %ecx used for counting
1462 # scas* %edi used for destination index
1463 # cmps* %e[sd]i used for indices
1464 # loop* %ecx used for counting
1469 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1471 # * GCC used an "STG reg" for its own purposes
1473 # * some secret uses of machine reg, requiring STG reg
1474 # to be saved/restored
1476 # The most dangerous "GCC uses" of an "STG reg" are when
1477 # the reg holds the target of a jmp -- it's tricky to
1478 # insert the patch-up code before we get to the target!
1479 # So here we change the jmps:
1481 # --------------------------------------------------------
1482 # it can happen that we have jumps of the form...
1483 # jmp *<something involving %esp>
1485 # jmp <something involving another naughty register...>
1487 # a reasonably-common case is:
1489 # movl $_blah,<bad-reg>
1492 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1496 # movl -4(%ebx), %eax
1501 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1503 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1504 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1505 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1506 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1507 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1508 if /(jmp|call)\s+.*\%esi/;
1510 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1511 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1512 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1513 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1514 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1515 if /(jmp|call)\s+.*\%edi/;
1518 # OK, now we can decide what our patch-up code is going to
1521 # Offsets into register table - you'd better update these magic
1522 # numbers should you change its contents!
1523 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1524 local($OFFSET_Hp)=88;
1526 # Note funky ".=" stuff; we're *adding* to these _patch guys
1527 if ( $StolenX86Regs <= 2
1528 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1529 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1530 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1532 # nothing for call_{entry,exit} because %esi is callee-save
1534 if ( $StolenX86Regs <= 3
1535 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1536 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1537 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1539 # nothing for call_{entry,exit} because %edi is callee-save
1542 # --------------------------------------------------------
1543 # next, here we go with non-%esp patching!
1545 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1547 # Before calling GC we must set up the exit condition before the call
1548 # and entry condition when we come back
1550 # fix _all_ non-local jumps:
1552 if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1553 # On Darwin, we've got local-looking jumps that are
1554 # actually global (i.e. jumps to Lfoo$stub or via
1555 # Lfoo$non_lazy_ptr), so we fix those first.
1556 # In fact, we just fix everything that contains a dollar
1557 # because false positives don't hurt here.
1559 s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1562 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1563 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1565 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1567 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1568 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1570 if ($StolenX86Regs == 2 ) {
1571 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1572 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1573 } elsif ($StolenX86Regs == 3 ) {
1574 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1575 if /^\t(jmp|call)\s+.*\%edi/;
1578 # --------------------------------------------------------
1579 # that's it -- print it
1581 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1585 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1586 print OUTASM $exit_patch;
1587 # ToDo: make it not print if there is a "jmp" at the end
1593 sub init_FUNNY_THINGS {
1594 %KNOWN_FUNNY_THING = (
1596 # "${T_US}stg_.*{T_POST_LBL}", 1,
1601 The following table reversal is used for both info tables and return
1602 vectors. In both cases, we remove the first entry from the table,
1603 reverse the table, put the label at the end, and paste some code
1604 (that which is normally referred to by the first entry in the table)
1605 right after the table itself. (The code pasting is done elsewhere.)
1609 local($symb, $tbl, $discard1) = @_;
1611 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1613 local($before) = '';
1615 local(@imports) = (); # hppa only
1618 local(@lines) = split(/\n/, $tbl);
1621 # Deal with the header...
1622 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1623 $label .= $lines[$i] . "\n",
1624 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1625 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1626 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1628 $before .= $lines[$i] . "\n"; # otherwise...
1632 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1634 # Grab the table data...
1635 if ( $TargetPlatform !~ /^hppa/ ) {
1636 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1638 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1639 # to offsets (relative to the info label),
1640 # in order to support position independent code.
1641 $line =~ s/$infoname/0/
1642 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1643 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1644 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1645 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1646 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1647 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1648 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1649 push(@words, $line);
1651 } else { # hppa weirdness
1652 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1653 # FIXME: the RTS now expects offsets instead of addresses
1654 # for all labels in info tables.
1655 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1656 push(@imports, $lines[$i]);
1658 # We don't use HP's ``function pointers''
1659 # We just use labels in code space, like normal people
1660 $lines[$i] =~ s/P%//;
1661 push(@words, $lines[$i]);
1666 # Now throw away any initial zero word from the table. This is a hack
1667 # that lets us reduce the size of info tables when the SRT field is not
1668 # needed: see comments StgFunInfoTable in InfoTables.h.
1670 # The .zero business is for Linux/ELF.
1671 # The .skip business is for Sparc/Solaris/ELF.
1672 # The .blockz business is for HPPA.
1674 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1679 for (; $i <= $#lines; $i++) {
1680 $after .= $lines[$i] . "\n";
1683 # Alphas: If we have anonymous text (not part of a procedure), the
1684 # linker may complain about missing exception information. Bleh.
1685 # To suppress this, we place a .ent/.end pair around the code.
1686 # At the same time, we have to be careful and not enclose any leading
1687 # .file/.loc directives.
1688 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1689 local ($ident) = $1;
1690 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1691 $after .= "\t.end $ident\n";
1694 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1695 # assembler (!) wherein .quad constants inside .text sections are
1696 # first narrowed to 32 bits then sign-extended back to 64 bits.
1697 # This obviously screws up our 64-bit bitmaps, so we work around
1698 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1699 if ( $TargetPlatform =~ /^alpha-/ ) {
1701 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1702 local ($number) = $1;
1703 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1704 local ($sign, $base, $digits) = ($1, $2, $3);
1705 $base = (10, 8, 16)[length $base];
1706 local ($hi, $lo) = (0, 0);
1707 foreach $i (split(//, $digits)) {
1708 $j = $lo * $base + $i;
1709 $lo = $j % 4294967296;
1710 $hi = $hi * $base + ($j - $lo) / 4294967296;
1712 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1714 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1715 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1717 print STDERR "Cannot handle \".quad $number\" in info table\n";
1725 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1726 . join("\n", @words) . "\n"
1729 # print STDERR "before=$before\n";
1730 # print STDERR "label=$label\n";
1731 # print STDERR "words=",(reverse @words),"\n";
1732 # print STDERR "after=$after\n";
1738 The HP is a major nuisance. The threaded code mangler moved info
1739 tables from data space to code space, but unthreaded code in the RTS
1740 still has references to info tables in data space. Since the HP
1741 linker is very precise about where symbols live, we need to patch the
1742 references in the unthreaded RTS as well.
1745 sub mini_mangle_asm_hppa {
1746 local($in_asmf, $out_asmf) = @_;
1748 open(INASM, "< $in_asmf")
1749 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1750 open(OUTASM,"> $out_asmf")
1751 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1754 s/_info,DATA/_info,CODE/; # Move _info references to code space
1760 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1761 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1767 sub tidy_up_and_die {
1768 local($return_val, $msg) = @_;
1770 exit (($return_val == 0) ? 0 : 1);