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 ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|gnu|freebsd|netbsd|openbsd)$/ ) {
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|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/ ))
594 $chkcat[$i] = 'dyld';
597 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
598 { # non_lazy_symbol_ptrs that point to local symbols
600 $chkcat[$i] = 'dyld';
603 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
604 # duplicated from the bottom
608 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
609 # Alphas: Local labels not to be confused with new chunks
611 # NB: all the rest start with a non-space
613 } elsif ( $TargetPlatform =~ /^mips-/
614 && /^\d+:/ ) { # a funny-looking very-local label
617 } elsif ( /$T_CONST_LBL/o ) {
619 $chkcat[$i] = 'literal';
622 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
624 $chkcat[$i] = 'splitmarker';
627 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
630 $chkcat[$i] = 'infotbl';
631 $chksymb[$i] = $symb;
633 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
635 $infochk{$symb} = $i;
637 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
639 $chkcat[$i] = 'entry';
644 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
646 $chkcat[$i] = 'closure';
649 $closurechk{$1} = $i;
651 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
658 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
660 $chkcat[$i] = 'data';
663 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
665 $chkcat[$i] = 'data';
668 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
671 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
672 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
673 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
674 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
677 $chkcat[$i] = 'data';
680 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
685 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
686 # all CC_ symbols go in the data section...
688 $chkcat[$i] = 'data';
691 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
693 $chkcat[$i] = 'misc';
695 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
697 $chkcat[$i] = 'vector';
702 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
703 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
704 # Some Solaris system headers contain function definitions (as
705 # opposed to mere prototypes), which end up in the .hc file when
706 # a Haskell module foreign imports the corresponding system
707 # functions (most notably stat()). We put them into the text
708 # segment. Note that this currently does not extend to function
709 # names starting with an underscore.
712 $chkcat[$i] = 'misc';
715 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
716 # To handle PIC on Darwin/x86, we need to appropriately pass through
717 # the get_pc_thunk functions. The need to be put into a special section
718 # marked as coalesced (otherwise the .weak_definition doesn't work
721 $chkcat[$i] = 'get_pc_thunk';
724 } elsif ( /^${T_US}[A-Za-z0-9_]/o
725 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
727 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
728 || ! /^\.L\d+:$/ ) ) {
735 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
736 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
737 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
738 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
739 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
740 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
741 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
742 || /^_uname:/o # x86/Solaris2
745 $chkcat[$i] = 'misc';
747 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
748 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
751 $chkcat[$i] = 'relrodata';
754 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
755 $chkcat[$i] = 'unknown';
758 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
759 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
760 # table "by hand". Be sure to copy it over.
761 # Note that this label and all entries in the table should actually
762 # go into the .got2 section, but it isn't easy to distinguish them
763 # from other constant literals (.LC\d+), so we just put everything
766 $chkcat[$i] = 'literal';
767 $chksymb[$i] = 'LCTOC1';
768 } else { # simple line (duplicated at the top)
773 $numchks = $#chk + 1;
775 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
776 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
779 # the division into chunks is imperfect;
780 # we throw some things over the fence into the next
783 # also, there are things we would like to know
784 # about the whole module before we start spitting
787 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
788 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
790 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
792 # Alphas: NB: we start meddling at chunk 1, not chunk 0
793 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
794 # spits a ".quad 0" in after the very first ".rdata"; we
795 # detect this special case (tossing the ".quad 0")!
796 local($magic_rdata_seen) = 0;
798 # HPPAs, MIPSen: also start medding at chunk 1
800 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
801 $c = $chk[$i]; # convenience copy
803 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
805 # toss all prologue stuff; HPPA is pretty weird
807 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
809 # be slightly paranoid to make sure there's
810 # nothing surprising in there
811 if ( $c =~ /--- BEGIN ---/ ) {
812 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
814 # remove junk whitespace around the split point
818 if ($TargetPlatform =~ /^i386-/) {
819 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
821 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
825 # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
826 # when compiling PrimOps.cmm with -optc-O2:
829 # movl %ecx, 16(%esp)
830 # movl %edx, 20(%esp)
831 # but then the code of the function doesn't assume
832 # anything about the contnets of these stack locations.
833 # I think it's to do with the use of inline functions for
834 # PK_Word64() and friends, where gcc is initialising the
835 # contents of the struct to zero, and failing to optimise
836 # away the initialisation. Let's live dangerously and
837 # discard these initalisations.
839 $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
840 $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
841 $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
842 $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
843 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
844 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
846 if ($TargetPlatform =~ /^i386-apple-darwin/) {
848 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
850 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
851 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
852 $p =~ s/^\"?L\d+\$pb\"?:\n//;
854 if ($pcrel_reg eq "bx") {
855 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
856 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
860 } elsif ($TargetPlatform =~ /^x86_64-/) {
861 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
862 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
863 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
865 } elsif ($TargetPlatform =~ /^ia64-/) {
866 $p =~ s/^\t\.prologue .*\n//;
867 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, 3[12], \d+, 0\n//;
868 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
869 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
870 $p =~ s/^\t\.(mii|mmi)\n//g; # bundling is no longer sensible
871 $p =~ s/^\t;;\n//g; # discard stops
872 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
874 # GCC 3.3 saves r1 in the prologue, move this to the body
875 if ($p =~ /^\tmov r\d+ = r1\n/) {
879 } elsif ($TargetPlatform =~ /^m68k-/) {
880 $p =~ s/^\tlink a6,#-?\d.*\n//;
881 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
882 # The above showed up in the asm code,
883 # so I added it here.
884 # I hope it's correct.
886 $p =~ s/^\tmovel d2,sp\@-\n//;
887 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
888 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
889 } elsif ($TargetPlatform =~ /^mips-/) {
890 # the .frame/.mask/.fmask that we use is the same
891 # as that produced by GCC for miniInterpret; this
892 # gives GDB some chance of figuring out what happened
893 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
894 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
895 $p =~ s/^\t\.(mask|fmask).*\n//g;
896 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
897 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
898 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
899 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
900 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
901 $p =~ s/__FRAME__/$FRAME/;
902 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
904 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
906 $p =~ s/^\tmflr r0\n//;
907 $p =~ s/^\tbl saveFP # f\d+\n//;
908 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
909 $p =~ s/^\"?L\d+\$pb\"?:\n//;
910 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
911 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
912 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
913 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
914 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
915 $p =~ s/^\tbcl 20,31,L\d+\$pb\n//;
916 $p =~ s/^L\d+\$pb:\n//;
917 $p =~ s/^\tmflr r31\n//;
919 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
920 # under some circumstances, only when generating position dependent code.
921 # I have no idea why, and I don't think it is necessary, so let's toss it.
922 $p =~ s/^\tli r\d+,0\n//g;
923 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
924 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
925 $p =~ s/^\tmflr 0\n//;
926 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
927 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
928 $p =~ s/^\tstw r0,8\(1\)\n//;
929 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
930 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
932 # GCC's "large-model" PIC (-fPIC)
934 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
936 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
937 $p =~ s/^.LCF\d+:\n//;
938 $p =~ s/^\tmflr 30\n//;
939 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
940 $p =~ s/^\tadd 30,0,30\n//;
942 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
943 # under some circumstances, only when generating position dependent code.
944 # I have no idea why, and I don't think it is necessary, so let's toss it.
945 $p =~ s/^\tli \d+,0\n//g;
946 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
947 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
948 $p =~ s/^\tmr 31,1\n//;
949 $p =~ s/^\tmflr 0\n//;
950 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
951 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
952 $p =~ s/^\tstd r0,8\(1\)\n//;
953 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
954 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
956 # GCC's "large-model" PIC (-fPIC)
958 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
960 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
961 $p =~ s/^.LCF\d+:\n//;
962 $p =~ s/^\tmflr 30\n//;
963 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
964 $p =~ s/^\tadd 30,0,30\n//;
966 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
967 # under some circumstances, only when generating position dependent code.
968 # I have no idea why, and I don't think it is necessary, so let's toss it.
969 $p =~ s/^\tli \d+,0\n//g;
970 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
972 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
975 # HWL HACK: dont die, just print a warning
976 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
977 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
979 # For PIC, we want to keep part of the prologue
980 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
981 # Darwin: load the current instruction pointer into register r31
982 $p .= "bcl 20,31,$pcrel_label\n";
983 $p .= "$pcrel_label:\n";
984 $p .= "\tmflr r31\n";
985 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
986 # Linux: load the GOT pointer into register 30
987 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
988 $p .= ".LCF$pcrel_label:\n";
990 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
991 $p .= "\tadd 30,0,30\n";
992 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
993 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
994 $p .= "$pcrel_label:\n";
997 # glue together what's left
1002 if ( $TargetPlatform =~ /^mips-/ ) {
1003 # MIPS: first, this basic sequence may occur "--- END ---" or not
1004 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1007 # toss all epilogue stuff; again, paranoidly
1008 if ( $c =~ /--- END ---/ ) {
1009 if (($r, $e) = split(/--- END ---/, $c)) {
1010 if ($TargetPlatform =~ /^i386-/) {
1012 $e =~ s/^\tpopl\s+\%edi\n//;
1013 $e =~ s/^\tpopl\s+\%esi\n//;
1014 $e =~ s/^\tpopl\s+\%edx\n//;
1015 $e =~ s/^\tpopl\s+\%ecx\n//;
1016 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1017 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1018 } elsif ($TargetPlatform =~ /^ia64-/) {
1019 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1020 $e =~ s/^\tmov b0 = r\d+\n//;
1021 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1022 $e =~ s/^\tbr\.ret\.sptk\.many b0\n//;
1023 $e =~ s/^\t\.(mii|mmi|mib)\n//g; # bundling is no longer sensible
1024 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1025 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1026 } elsif ($TargetPlatform =~ /^m68k-/) {
1027 $e =~ s/^\tunlk a6\n//;
1029 } elsif ($TargetPlatform =~ /^mips-/) {
1030 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1031 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1032 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1033 $e =~ s/^\tj\t\$31\n//;
1034 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1035 $e =~ s/^\taddi r1,r1,\d+\n//;
1036 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1037 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1038 $e =~ s/^\tmtlr r0\n//;
1040 $e =~ s/^\tb restFP ;.*\n//;
1041 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1042 $e =~ s/^\tmr 3,0\n//;
1043 $e =~ s/^\taddi 1,1,\d+\n//;
1044 $e =~ s/^\tld 0,16\(1\)\n//;
1045 $e =~ s/^\tmtlr 0\n//;
1047 # callee-save registers
1048 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1049 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1051 # get rid of the debug junk along with the blr
1052 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1054 # incase we missed it with the last one get the blr alone
1057 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1060 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1062 # glue together what's left
1064 $c =~ s/\n\t\n/\n/; # junk blank line
1068 # On SPARCs, we don't do --- BEGIN/END ---, we just
1069 # toss the register-windowing save/restore/ret* instructions
1070 # directly unless they've been generated by function definitions in header
1072 if ( $TargetPlatform =~ /^sparc-/ ) {
1073 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1074 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1076 # throw away PROLOGUE comments
1077 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1080 # On Alphas, the prologue mangling is done a little later (below)
1082 # toss all calls to __DISCARD__
1083 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1084 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1085 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1086 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1087 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1088 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1090 # IA64: mangle tailcalls into jumps here
1091 if ($TargetPlatform =~ /^ia64-/) {
1092 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/) {
1093 # Eek, the gcc optimiser is getting smarter... if we see a jump to the --- TAILCALL ---
1094 # marker then we reapply the substitution at the source sites
1095 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1099 # MIPS: that may leave some gratuitous asm macros around
1100 # (no harm done; but we get rid of them to be tidier)
1101 $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/
1102 if $TargetPlatform =~ /^mips-/;
1104 # toss stack adjustment after DoSparks
1105 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1106 if $TargetPlatform =~ /^m68k-/; # this looks old...
1108 if ( $TargetPlatform =~ /^alpha-/ &&
1109 ! $magic_rdata_seen &&
1110 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1111 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1112 $magic_rdata_seen = 1;
1115 # pick some end-things and move them to the next chunk
1117 # pin a funny end-thing on (for easier matching):
1118 $c .= 'FUNNY#END#THING';
1120 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1124 # on x86 we try not to copy any directives into a literal
1125 # chunk, rather we keep looking for the next real chunk. This
1126 # is because we get things like
1128 # .globl blah_closure
1134 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1136 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1139 if ( $j < $numchks ) {
1140 $chk[$j] = $to_move . $chk[$j];
1144 elsif ( $i < ($numchks - 1)
1145 && ( $to_move =~ /${T_COPY_DIRVS}/
1146 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
1147 $chk[$i + 1] = $to_move . $chk[$i + 1];
1148 # otherwise they're tossed
1151 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1154 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1156 # toss all prologue stuff, except for loading gp, and the ..ng address
1157 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1158 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1159 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1160 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1161 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1162 $c = $keep . "..ng:\n";
1164 print STDERR "malformed code block ($ent)?\n"
1167 $c .= "\t.prologue" . $r;
1171 $c =~ s/FUNNY#END#THING//;
1173 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1175 $chk[$i] = $c; # update w/ convenience copy
1178 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1179 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1182 if ( $TargetPlatform =~ /^alpha-/ ) {
1183 # print out the header stuff first
1184 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1185 print OUTASM $chk[0];
1187 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1188 print OUTASM $chk[0];
1190 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1191 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1193 # get rid of horrible "<dollar>Revision: .*$" strings
1194 local(@lines0) = split(/\n/, $chk[0]);
1196 while ( $z <= $#lines0 ) {
1197 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1200 while ( $z <= $#lines0 ) {
1202 last if $lines0[$z] =~ /[,\t]0x0$/;
1208 $chk[0] = join("\n", @lines0);
1209 $chk[0] =~ s/\n\n+/\n/;
1210 print OUTASM $chk[0];
1213 # print out all the literal strings next
1214 for ($i = 0; $i < $numchks; $i++) {
1215 if ( $chkcat[$i] eq 'literal' ) {
1217 # HACK: try to detect 16-byte constants and align them
1218 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1219 # aligned constants, and so does Darwin/x86.
1220 if ( $TargetPlatform =~ /^x86_64/
1221 || $TargetPlatform =~ /^i386-apple-darwin/ ) {
1223 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1224 print OUTASM $T_HDR_literal16;
1226 print OUTASM $T_HDR_literal;
1229 print OUTASM $T_HDR_literal;
1232 print OUTASM $chk[$i];
1233 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1235 $chkcat[$i] = 'DONE ALREADY';
1239 # on the HPPA, print out all the bss next
1240 if ( $TargetPlatform =~ /^hppa/ ) {
1241 for ($i = 1; $i < $numchks; $i++) {
1242 if ( $chkcat[$i] eq 'bss' ) {
1243 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1244 print OUTASM $chk[$i];
1246 $chkcat[$i] = 'DONE ALREADY';
1251 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
1252 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1254 next if $chkcat[$i] eq 'DONE ALREADY';
1256 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1257 if ($chk[$i] ne '') {
1258 print OUTASM $T_HDR_misc;
1259 &print_doctored($chk[$i], 0);
1262 } elsif ( $chkcat[$i] eq 'toss' ) {
1263 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1265 } elsif ( $chkcat[$i] eq 'data' ) {
1266 if ($chk[$i] ne '') {
1267 print OUTASM $T_HDR_data;
1268 print OUTASM $chk[$i];
1271 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1272 # we can just re-constitute this one...
1273 # NB: we emit _three_ underscores no matter what,
1274 # so ghc-split doesn't have to care.
1275 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1277 } elsif ( $chkcat[$i] eq 'closure'
1278 || $chkcat[$i] eq 'srt'
1279 || $chkcat[$i] eq 'infotbl'
1280 || $chkcat[$i] eq 'entry') { # do them in that order
1281 $symb = $chksymb[$i];
1284 if ( defined($closurechk{$symb}) ) {
1285 print OUTASM $T_HDR_closure;
1286 print OUTASM $chk[$closurechk{$symb}];
1287 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1291 if ( defined($srtchk{$symb}) ) {
1292 print OUTASM $T_HDR_relrodata;
1293 print OUTASM $chk[$srtchk{$symb}];
1294 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1298 if ( defined($infochk{$symb}) ) {
1300 print OUTASM $T_HDR_info;
1301 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1303 # entry code will be put here!
1305 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1309 if ( defined($entrychk{$symb}) ) {
1311 $c = $chk[$entrychk{$symb}];
1313 # If this is an entry point with an info table,
1314 # eliminate the entry symbol and all directives involving it.
1315 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1317 foreach $l (split(/\n/,$c)) {
1318 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1320 # If we have .type/.size direrctives involving foo_entry,
1321 # then make them refer to foo_info instead. The information
1322 # in these directives is used by the cachegrind annotator,
1323 # so it is worthwhile keeping.
1324 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1325 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1329 next if $l =~ /^\s*\..*$symb.*\n?/;
1332 $c = join("\n",@o) . "\n";
1335 print OUTASM $T_HDR_entry;
1337 &print_doctored($c, 1); # NB: the 1!!!
1339 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1342 } elsif ( $chkcat[$i] eq 'vector' ) {
1343 $symb = $chksymb[$i];
1346 if ( defined($vectorchk{$symb}) ) {
1347 print OUTASM $T_HDR_vector;
1348 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1350 # direct return code will be put here!
1351 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1353 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1354 # Alphas: the commented nop is for the splitter, to ensure
1355 # that no module ends with a label as the very last
1356 # thing. (The linker will adjust the label to point
1357 # to the first code word of the next module linked in,
1358 # even if alignment constraints cause the label to move!)
1360 print OUTASM "\t# nop\n";
1363 } elsif ( $chkcat[$i] eq 'rodata' ) {
1364 print OUTASM $T_HDR_rodata;
1365 print OUTASM $chk[$i];
1366 $chkcat[$i] = 'DONE ALREADY';
1367 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1368 print OUTASM $T_HDR_relrodata;
1369 print OUTASM $chk[$i];
1370 $chkcat[$i] = 'DONE ALREADY';
1371 } elsif ( $chkcat[$i] eq 'toc' ) {
1372 # silly optimisation to print tocs, since they come in groups...
1373 print OUTASM $T_HDR_toc;
1375 while ($chkcat[$j] eq 'toc')
1376 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1379 print OUTASM $chk[$j];
1381 $chkcat[$j] = 'DONE ALREADY';
1385 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1386 # apple-darwin: dynamic linker stubs
1387 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1388 { # print them out unchanged, but remove the stubs for __DISCARD__
1389 print OUTASM $chk[$i];
1391 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1392 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1393 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1394 print OUTASM $chk[$i];
1396 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1400 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1403 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1404 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1409 sub hppa_mash_prologue { # OK, epilogue, too
1412 # toss all prologue stuff
1413 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1415 # Lie about our .CALLINFO
1416 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1423 # toss all epilogue stuff
1424 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1426 # Sorry; we moved the _info stuff to the code segment.
1427 s/_info,DATA/_info,CODE/g;
1434 sub print_doctored {
1435 local($_, $need_fallthru_patch) = @_;
1437 if ( $TargetPlatform =~ /^x86_64-/ ) {
1440 # movq -4(%ebp), %rax
1445 s/^\tmovq\s+(-?\d*\(\%r(bx|bp|13)\)),\s*(\%r(ax|cx|dx|10|11))\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1446 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e(ax|cx|si|di))\n\tjmp\s+\*\%r\3/\tjmp\t$T_US$1/g;
1449 if ( $TargetPlatform !~ /^i386-/
1450 || ! /^\t[a-z]/ # no instructions in here, apparently
1451 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1456 # OK, must do some x86 **HACKING**
1458 local($entry_patch) = '';
1459 local($exit_patch) = '';
1461 # gotta watch out for weird instructions that
1462 # invisibly smash various regs:
1463 # rep* %ecx used for counting
1464 # scas* %edi used for destination index
1465 # cmps* %e[sd]i used for indices
1466 # loop* %ecx used for counting
1471 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1473 # * GCC used an "STG reg" for its own purposes
1475 # * some secret uses of machine reg, requiring STG reg
1476 # to be saved/restored
1478 # The most dangerous "GCC uses" of an "STG reg" are when
1479 # the reg holds the target of a jmp -- it's tricky to
1480 # insert the patch-up code before we get to the target!
1481 # So here we change the jmps:
1483 # --------------------------------------------------------
1484 # it can happen that we have jumps of the form...
1485 # jmp *<something involving %esp>
1487 # jmp <something involving another naughty register...>
1489 # a reasonably-common case is:
1491 # movl $_blah,<bad-reg>
1494 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1498 # movl -4(%ebx), %eax
1503 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1505 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1506 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1507 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1508 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1509 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1510 if /(jmp|call)\s+.*\%esi/;
1512 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1513 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1514 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1515 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1516 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1517 if /(jmp|call)\s+.*\%edi/;
1520 # OK, now we can decide what our patch-up code is going to
1523 # Offsets into register table - you'd better update these magic
1524 # numbers should you change its contents!
1525 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1526 local($OFFSET_Hp)=88;
1528 # Note funky ".=" stuff; we're *adding* to these _patch guys
1529 if ( $StolenX86Regs <= 2
1530 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1531 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1532 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1534 # nothing for call_{entry,exit} because %esi is callee-save
1536 if ( $StolenX86Regs <= 3
1537 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1538 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1539 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1541 # nothing for call_{entry,exit} because %edi is callee-save
1544 # --------------------------------------------------------
1545 # next, here we go with non-%esp patching!
1547 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1549 # Before calling GC we must set up the exit condition before the call
1550 # and entry condition when we come back
1552 # fix _all_ non-local jumps:
1554 if ( $TargetPlatform =~ /^.*-apple-darwin.*/ ) {
1555 # On Darwin, we've got local-looking jumps that are
1556 # actually global (i.e. jumps to Lfoo$stub or via
1557 # Lfoo$non_lazy_ptr), so we fix those first.
1558 # In fact, we just fix everything that contains a dollar
1559 # because false positives don't hurt here.
1561 s/^(\tjmp\s+\*?L.*\$.*\n)/$exit_patch$1/g;
1564 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1565 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1567 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1569 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1570 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1572 if ($StolenX86Regs == 2 ) {
1573 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1574 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1575 } elsif ($StolenX86Regs == 3 ) {
1576 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1577 if /^\t(jmp|call)\s+.*\%edi/;
1580 # --------------------------------------------------------
1581 # that's it -- print it
1583 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1587 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1588 print OUTASM $exit_patch;
1589 # ToDo: make it not print if there is a "jmp" at the end
1595 sub init_FUNNY_THINGS {
1596 %KNOWN_FUNNY_THING = (
1598 # "${T_US}stg_.*{T_POST_LBL}", 1,
1603 The following table reversal is used for both info tables and return
1604 vectors. In both cases, we remove the first entry from the table,
1605 reverse the table, put the label at the end, and paste some code
1606 (that which is normally referred to by the first entry in the table)
1607 right after the table itself. (The code pasting is done elsewhere.)
1611 local($symb, $tbl, $discard1) = @_;
1613 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1615 local($before) = '';
1617 local(@imports) = (); # hppa only
1620 local(@lines) = split(/\n/, $tbl);
1623 # Deal with the header...
1624 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1625 $label .= $lines[$i] . "\n",
1626 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1627 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1628 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1630 $before .= $lines[$i] . "\n"; # otherwise...
1634 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1636 # Grab the table data...
1637 if ( $TargetPlatform !~ /^hppa/ ) {
1638 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1640 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1641 # to offsets (relative to the info label),
1642 # in order to support position independent code.
1643 $line =~ s/$infoname/0/
1644 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1645 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1646 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1647 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1648 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1649 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1650 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1651 push(@words, $line);
1653 } else { # hppa weirdness
1654 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1655 # FIXME: the RTS now expects offsets instead of addresses
1656 # for all labels in info tables.
1657 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1658 push(@imports, $lines[$i]);
1660 # We don't use HP's ``function pointers''
1661 # We just use labels in code space, like normal people
1662 $lines[$i] =~ s/P%//;
1663 push(@words, $lines[$i]);
1668 # Now throw away any initial zero word from the table. This is a hack
1669 # that lets us reduce the size of info tables when the SRT field is not
1670 # needed: see comments StgFunInfoTable in InfoTables.h.
1672 # The .zero business is for Linux/ELF.
1673 # The .skip business is for Sparc/Solaris/ELF.
1674 # The .blockz business is for HPPA.
1676 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1681 for (; $i <= $#lines; $i++) {
1682 $after .= $lines[$i] . "\n";
1685 # Alphas: If we have anonymous text (not part of a procedure), the
1686 # linker may complain about missing exception information. Bleh.
1687 # To suppress this, we place a .ent/.end pair around the code.
1688 # At the same time, we have to be careful and not enclose any leading
1689 # .file/.loc directives.
1690 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1691 local ($ident) = $1;
1692 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1693 $after .= "\t.end $ident\n";
1696 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1697 # assembler (!) wherein .quad constants inside .text sections are
1698 # first narrowed to 32 bits then sign-extended back to 64 bits.
1699 # This obviously screws up our 64-bit bitmaps, so we work around
1700 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1701 if ( $TargetPlatform =~ /^alpha-/ ) {
1703 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1704 local ($number) = $1;
1705 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1706 local ($sign, $base, $digits) = ($1, $2, $3);
1707 $base = (10, 8, 16)[length $base];
1708 local ($hi, $lo) = (0, 0);
1709 foreach $i (split(//, $digits)) {
1710 $j = $lo * $base + $i;
1711 $lo = $j % 4294967296;
1712 $hi = $hi * $base + ($j - $lo) / 4294967296;
1714 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1716 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1717 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1719 print STDERR "Cannot handle \".quad $number\" in info table\n";
1727 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1728 . join("\n", @words) . "\n"
1731 # print STDERR "before=$before\n";
1732 # print STDERR "label=$label\n";
1733 # print STDERR "words=",(reverse @words),"\n";
1734 # print STDERR "after=$after\n";
1740 The HP is a major nuisance. The threaded code mangler moved info
1741 tables from data space to code space, but unthreaded code in the RTS
1742 still has references to info tables in data space. Since the HP
1743 linker is very precise about where symbols live, we need to patch the
1744 references in the unthreaded RTS as well.
1747 sub mini_mangle_asm_hppa {
1748 local($in_asmf, $out_asmf) = @_;
1750 open(INASM, "< $in_asmf")
1751 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1752 open(OUTASM,"> $out_asmf")
1753 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1756 s/_info,DATA/_info,CODE/; # Move _info references to code space
1762 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1763 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1769 sub tidy_up_and_die {
1770 local($return_val, $msg) = @_;
1772 exit (($return_val == 0) ? 0 : 1);