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_literal = "\t\.const\n\t\.align 2\n";
336 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
337 $T_HDR_data = "\t\.data\n\t\.align 2\n";
338 $T_HDR_rodata = "\t\.const\n\t\.align 2\n";
339 $T_HDR_relrodata= "\t\.const_data\n\t\.align 2\n";
340 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
341 $T_HDR_info = "\t\.text\n\t\.align 2\n";
342 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
343 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
345 #--------------------------------------------------------#
346 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ ) {
348 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
349 $T_US = ''; # _ if symbols have an underscore on the front
350 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
351 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
354 $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)';
355 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
357 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
358 $T_DOT_GLOBAL = '\.globl';
359 $T_HDR_toc = "\.toc\n";
360 $T_HDR_literal = "\t\.section\t.rodata\n\t\.align 2\n";
361 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
362 $T_HDR_data = "\t\.data\n\t\.align 2\n";
363 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
364 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
365 $T_HDR_info = "\t\.text\n\t\.align 2\n";
366 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
367 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
369 #--------------------------------------------------------#
370 } elsif ( $TargetPlatform =~ /^powerpc64-.*-linux/ ) {
372 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
373 $T_US = '\.'; # _ if symbols have an underscore on the front
374 $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP
375 $T_CONST_LBL = '^\.LC\d+:'; # regexp for what such a lbl looks like
378 $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)';
379 $T_COPY_DIRVS = '^\s*\.(globl|type|size|local)';
381 $T_DOT_WORD = '\.(long|short|byte|fill|space)';
382 $T_DOT_GLOBAL = '\.globl';
383 $T_HDR_toc = "\.toc\n";
384 $T_HDR_literal = "\t\.section\t\".toc\",\"aw\"\n";
385 $T_HDR_misc = "\t\.text\n\t\.align 2\n";
386 $T_HDR_data = "\t\.data\n\t\.align 2\n";
387 $T_HDR_rodata = "\t\.section\t.rodata\n\t\.align 2\n";
388 $T_HDR_closure = "\t\.data\n\t\.align 2\n";
389 $T_HDR_info = "\t\.text\n\t\.align 2\n";
390 $T_HDR_entry = "\t\.text\n\t\.align 2\n";
391 $T_HDR_vector = "\t\.text\n\t\.align 2\n";
393 #--------------------------------------------------------#
394 } elsif ( $TargetPlatform =~ /^sparc-.*-(solaris2|openbsd)/ ) {
396 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
397 $T_US = ''; # _ if symbols have an underscore on the front
398 $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
399 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
402 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*|\s*\.section.*|\s+\.type.*|\s+\.size.*)\n)';
403 $T_COPY_DIRVS = '\.(global|proc|stab)';
405 $T_DOT_WORD = '\.(long|word|byte|half|skip|uahalf|uaword)';
406 $T_DOT_GLOBAL = '^\t\.global';
407 $T_HDR_literal = "\.text\n\t\.align 8\n";
408 $T_HDR_misc = "\.text\n\t\.align 4\n";
409 $T_HDR_data = "\.data\n\t\.align 8\n";
410 $T_HDR_rodata = "\.text\n\t\.align 4\n";
411 $T_HDR_closure = "\.data\n\t\.align 4\n";
412 $T_HDR_info = "\.text\n\t\.align 4\n";
413 $T_HDR_entry = "\.text\n\t\.align 4\n";
414 $T_HDR_vector = "\.text\n\t\.align 4\n";
416 #--------------------------------------------------------#
417 } elsif ( $TargetPlatform =~ /^sparc-.*-sunos4/ ) {
419 $T_STABBY = 1; # 1 iff .stab things (usually if a.out format)
420 $T_US = '_'; # _ if symbols have an underscore on the front
421 $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP
422 $T_CONST_LBL = '^LC(\d+):$';
425 $T_MOVE_DIRVS = '^((\s+\.align\s+\d+|\s+\.proc\s+\d+|\s+\.global\s+\S+|\.text|\.data|\.stab.*)\n)';
426 $T_COPY_DIRVS = '\.(global|proc|stab)';
428 $T_DOT_WORD = '\.word';
429 $T_DOT_GLOBAL = '^\t\.global';
430 $T_HDR_literal = "\.text\n\t\.align 8\n";
431 $T_HDR_misc = "\.text\n\t\.align 4\n";
432 $T_HDR_data = "\.data\n\t\.align 8\n";
433 $T_HDR_rodata = "\.text\n\t\.align 4\n";
434 $T_HDR_closure = "\.data\n\t\.align 4\n";
435 $T_HDR_info = "\.text\n\t\.align 4\n";
436 $T_HDR_entry = "\.text\n\t\.align 4\n";
437 $T_HDR_vector = "\.text\n\t\.align 4\n";
439 #--------------------------------------------------------#
440 } elsif ( $TargetPlatform =~ /^sparc-.*-linux/ ) {
441 $T_STABBY = 0; # 1 iff .stab things (usually if a.out format)
442 $T_US = ''; # _ if symbols have an underscore on the front
443 $T_PRE_APP = '#'; # regexp that says what comes before APP/NO_APP
444 # Probably doesn't apply anyway
445 $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like
448 $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)';
449 $T_COPY_DIRVS = '\.(global|globl|proc|stab)';
451 $T_DOT_WORD = '\.(long|word|nword|xword|byte|half|short|skip|uahalf|uaword)';
452 $T_DOT_GLOBAL = '^\t\.global';
453 $T_HDR_literal = "\.text\n\t\.align 8\n";
454 $T_HDR_misc = "\.text\n\t\.align 4\n";
455 $T_HDR_data = "\.data\n\t\.align 8\n";
456 $T_HDR_rodata = "\.text\n\t\.align 4\n";
457 $T_HDR_closure = "\.data\n\t\.align 4\n";
458 $T_HDR_info = "\.text\n\t\.align 4\n";
459 $T_HDR_entry = "\.text\n\t\.align 4\n";
460 $T_HDR_vector = "\.text\n\t\.align 4\n";
462 #--------------------------------------------------------#
464 print STDERR "$Pgm: don't know how to mangle assembly language for: $TargetPlatform\n";
468 if($T_HDR_relrodata eq "") {
470 # relrodata defaults to rodata.
471 $T_HDR_relrodata = $T_HDR_rodata;
475 print STDERR "T_STABBY: $T_STABBY\n";
476 print STDERR "T_US: $T_US\n";
477 print STDERR "T_PRE_APP: $T_PRE_APP\n";
478 print STDERR "T_CONST_LBL: $T_CONST_LBL\n";
479 print STDERR "T_POST_LBL: $T_POST_LBL\n";
480 if ( $TargetPlatform =~ /^i386-/ ) {
481 print STDERR "T_X86_PRE_LLBL_PAT: $T_X86_PRE_LLBL_PAT\n";
482 print STDERR "T_X86_PRE_LLBL: $T_X86_PRE_LLBL\n";
483 print STDERR "T_X86_BADJMP: $T_X86_BADJMP\n";
485 print STDERR "T_MOVE_DIRVS: $T_MOVE_DIRVS\n";
486 print STDERR "T_COPY_DIRVS: $T_COPY_DIRVS\n";
487 print STDERR "T_DOT_WORD: $T_DOT_WORD\n";
488 print STDERR "T_HDR_literal: $T_HDR_literal\n";
489 print STDERR "T_HDR_misc: $T_HDR_misc\n";
490 print STDERR "T_HDR_data: $T_HDR_data\n";
491 print STDERR "T_HDR_rodata: $T_HDR_rodata\n";
492 print STDERR "T_HDR_closure: $T_HDR_closure\n";
493 print STDERR "T_HDR_info: $T_HDR_info\n";
494 print STDERR "T_HDR_entry: $T_HDR_entry\n";
495 print STDERR "T_HDR_vector: $T_HDR_vector\n";
501 %************************************************************************
503 \subsection{Mangle away}
505 %************************************************************************
509 local($in_asmf, $out_asmf) = @_;
511 # multi-line regexp matching:
516 &init_TARGET_STUFF();
517 &init_FUNNY_THINGS();
519 open(INASM, "< $in_asmf")
520 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
521 open(OUTASM,"> $out_asmf")
522 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
524 # read whole file, divide into "chunks":
525 # record some info about what we've found...
527 @chk = (); # contents of the chunk
528 $numchks = 0; # number of them
529 @chkcat = (); # what category of thing in each chunk
530 @chksymb = (); # what symbol(base) is defined in this chunk
531 %entrychk = (); # ditto, its entry code
532 %closurechk = (); # ditto, the (static) closure
533 %srtchk = (); # ditto, its SRT (for top-level things)
534 %infochk = (); # given a symbol base, say what chunk its info tbl is in
535 %vectorchk = (); # ditto, return vector table
536 $EXTERN_DECLS = ''; # .globl <foo> .text (MIPS only)
538 $i = 0; $chkcat[0] = 'misc'; $chk[0] = '';
541 tr/\r//d if $TargetPlatform =~ /-mingw32$/; # In case Perl doesn't convert line endings
542 next if $T_STABBY && /^\.stab.*${T_US}__stg_split_marker/o;
543 next if $T_STABBY && /^\.stab.*ghc.*c_ID/;
544 next if /^\t\.def.*endef$/;
545 next if /${T_PRE_APP}(NO_)?APP/o;
546 next if /^;/ && $TargetPlatform =~ /^hppa/;
548 next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|ia64)-/;
550 if ( $TargetPlatform =~ /^mips-/
551 && /^\t\.(globl\S+\.text|comm\t)/ ) {
552 $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/;
553 # Treat .comm variables as data. These show up in two (known) places:
555 # - the module_registered variable used in the __stginit fragment.
556 # even though these are declared static and initialised, gcc 3.3
557 # likes to make them .comm, presumably to save space in the
560 # - global variables used to pass arguments from C to STG in
561 # a foreign export. (is this still true? --SDM)
563 } elsif ( /^\t\.comm.*$/ ) {
565 $chkcat[$i] = 'data';
568 # Labels ending "_str": these are literal strings.
569 } elsif ( /^${T_US}([A-Za-z0-9_]+)_str${T_POST_LBL}$/ ) {
571 $chkcat[$i] = 'rodata';
573 } elsif ( $TargetPlatform =~ /-darwin/
574 && (/^\s*\.subsections_via_symbols/
575 ||/^\s*\.no_dead_strip.*/)) {
576 # Don't allow Apple's linker to do any dead-stripping of symbols
577 # in this file, because it will mess up info-tables in mangled
579 # The .no_dead_strip directives are actually put there by
580 # the gcc3 "used" attribute on entry points.
582 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && (
583 /^\s*\.picsymbol_stub/
584 || /^\s*\.section __TEXT,__picsymbol_stub\d,.*/
585 || /^\s*\.section __TEXT,__picsymbolstub\d,.*/
586 || /^\s*\.symbol_stub/
587 || /^\s*\.section __TEXT,__symbol_stub\d,.*/
588 || /^\s*\.section __TEXT,__symbolstub\d,.*/
589 || /^\s*\.lazy_symbol_pointer/
590 || /^\s*\.non_lazy_symbol_pointer/ ))
593 $chkcat[$i] = 'dyld';
596 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' && /^\s*\.data/)
597 { # non_lazy_symbol_ptrs that point to local symbols
599 $chkcat[$i] = 'dyld';
602 } elsif ( /^\s+/ ) { # most common case first -- a simple line!
603 # duplicated from the bottom
607 } elsif ( /\.\.ng:$/ && $TargetPlatform =~ /^alpha-/ ) {
608 # Alphas: Local labels not to be confused with new chunks
610 # NB: all the rest start with a non-space
612 } elsif ( $TargetPlatform =~ /^mips-/
613 && /^\d+:/ ) { # a funny-looking very-local label
616 } elsif ( /$T_CONST_LBL/o ) {
618 $chkcat[$i] = 'literal';
621 } elsif ( /^${T_US}__stg_split_marker(\d*)${T_POST_LBL}$/o ) {
623 $chkcat[$i] = 'splitmarker';
626 } elsif ( /^${T_US}([A-Za-z0-9_]+)_info${T_POST_LBL}$/o ) {
629 $chkcat[$i] = 'infotbl';
630 $chksymb[$i] = $symb;
632 die "Info table already? $symb; $i\n" if defined($infochk{$symb});
634 $infochk{$symb} = $i;
636 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(entry|ret)${T_POST_LBL}$/o ) {
638 $chkcat[$i] = 'entry';
643 } elsif ( /^${T_US}([A-Za-z0-9_]+)_closure${T_POST_LBL}$/o ) {
645 $chkcat[$i] = 'closure';
648 $closurechk{$1} = $i;
650 } elsif ( /^${T_US}([A-Za-z0-9_]+)_srt${T_POST_LBL}$/o ) {
657 } elsif ( /^${T_US}([A-Za-z0-9_]+)_ct${T_POST_LBL}$/o ) {
659 $chkcat[$i] = 'data';
662 } elsif ( /^${T_US}(stg_ap_stack_entries|stg_stack_save_entries|stg_arg_bitmaps)${T_POST_LBL}$/o ) {
664 $chkcat[$i] = 'data';
667 } elsif ( /^(${T_US}__gnu_compiled_c|gcc2_compiled\.)${T_POST_LBL}/o ) {
670 } elsif ( /^${T_US}[A-Za-z0-9_]+\.\d+${T_POST_LBL}$/o
671 || /^${T_US}.*_CAT${T_POST_LBL}$/o # PROF: _entryname_CAT
672 || /^${T_US}.*_done${T_POST_LBL}$/o # PROF: _module_done
673 || /^${T_US}_module_registered${T_POST_LBL}$/o # PROF: _module_registered
676 $chkcat[$i] = 'data';
679 } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) {
684 } elsif ( /^${T_US}([A-Za-z0-9_]+)_cc(s)?${T_POST_LBL}$/o ) {
685 # all CC_ symbols go in the data section...
687 $chkcat[$i] = 'data';
690 } elsif ( /^${T_US}([A-Za-z0-9_]+)_(alt|dflt)${T_POST_LBL}$/o ) {
692 $chkcat[$i] = 'misc';
694 } elsif ( /^${T_US}([A-Za-z0-9_]+)_vtbl${T_POST_LBL}$/o ) {
696 $chkcat[$i] = 'vector';
701 } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/
702 && /^[A-Za-z0-9][A-Za-z0-9_]*:/ ) {
703 # Some Solaris system headers contain function definitions (as
704 # opposed to mere prototypes), which end up in the .hc file when
705 # a Haskell module foreign imports the corresponding system
706 # functions (most notably stat()). We put them into the text
707 # segment. Note that this currently does not extend to function
708 # names starting with an underscore.
711 $chkcat[$i] = 'misc';
714 } elsif ( $TargetPlatform =~ /^i386-apple-darwin/ && /^(___i686\.get_pc_thunk\.[abcd]x):/o) {
715 # To handle PIC on Darwin/x86, we need to appropriately pass through
716 # the get_pc_thunk functions. The need to be put into a special section
717 # marked as coalesced (otherwise the .weak_definition doesn't work
720 $chkcat[$i] = 'get_pc_thunk';
723 } elsif ( /^${T_US}[A-Za-z0-9_]/o
724 && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case
726 && ( $TargetPlatform !~ /^powerpc64/ # we need to avoid local labels in this case
727 || ! /^\.L\d+:$/ ) ) {
734 /^${T_US}stg_.*${T_POST_LBL}$/o # RTS internals
735 || /^${T_US}__stg_.*${T_POST_LBL}$/o # more RTS internals
736 || /^${T_US}__fexp_.*${T_POST_LBL}$/o # foreign export
737 || /^${T_US}.*_slow${T_POST_LBL}$/o # slow entry
738 || /^${T_US}__stginit.*${T_POST_LBL}$/o # __stginit<module>
739 || /^${T_US}.*_btm${T_POST_LBL}$/o # large bitmaps
740 || /^${T_US}.*_fast${T_POST_LBL}$/o # primops
741 || /^_uname:/o # x86/Solaris2
744 $chkcat[$i] = 'misc';
746 /^${T_US}.*_srtd${T_POST_LBL}$/o # large bitmaps
747 || /^${T_US}.*_closure_tbl${T_POST_LBL}$/o # closure tables
750 $chkcat[$i] = 'relrodata';
753 print STDERR "Warning: retaining unknown function \`$thing' in output from C compiler\n";
754 $chkcat[$i] = 'unknown';
757 } elsif ( $TargetPlatform =~ /^powerpc-.*-linux/ && /^\.LCTOC1 = /o ) {
758 # PowerPC Linux's large-model PIC (-fPIC) generates a gobal offset
759 # table "by hand". Be sure to copy it over.
760 # Note that this label and all entries in the table should actually
761 # go into the .got2 section, but it isn't easy to distinguish them
762 # from other constant literals (.LC\d+), so we just put everything
765 $chkcat[$i] = 'literal';
766 $chksymb[$i] = 'LCTOC1';
767 } else { # simple line (duplicated at the top)
772 $numchks = $#chk + 1;
774 # open CHUNKS, ">/tmp/chunks1" or die "Cannot open /tmp/chunks1: $!\n";
775 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
778 # the division into chunks is imperfect;
779 # we throw some things over the fence into the next
782 # also, there are things we would like to know
783 # about the whole module before we start spitting
786 local($FIRST_MANGLABLE) = ($TargetPlatform =~ /^(alpha-|hppa|mips-)/) ? 1 : 0;
787 local($FIRST_TOSSABLE ) = ($TargetPlatform =~ /^(hppa|mips-)/) ? 1 : 0;
789 # print STDERR "first chunk to mangle: $FIRST_MANGLABLE\n";
791 # Alphas: NB: we start meddling at chunk 1, not chunk 0
792 # The first ".rdata" is quite magical; as of GCC 2.7.x, it
793 # spits a ".quad 0" in after the very first ".rdata"; we
794 # detect this special case (tossing the ".quad 0")!
795 local($magic_rdata_seen) = 0;
797 # HPPAs, MIPSen: also start medding at chunk 1
799 for ($i = $FIRST_TOSSABLE; $i < $numchks; $i++) {
800 $c = $chk[$i]; # convenience copy
802 # print STDERR "\nCHK $i (BEFORE) (",$chkcat[$i],"):\n", $c;
804 # toss all prologue stuff; HPPA is pretty weird
806 $c = &hppa_mash_prologue($c) if $TargetPlatform =~ /^hppa-/;
808 # be slightly paranoid to make sure there's
809 # nothing surprising in there
810 if ( $c =~ /--- BEGIN ---/ ) {
811 if (($p, $r) = split(/--- BEGIN ---/, $c)) {
813 # remove junk whitespace around the split point
817 if ($TargetPlatform =~ /^i386-/) {
818 if ($p =~ /^\tsubl\s+\$(\d+),\s*\%esp\n/) {
820 die "Error: reserved stack space exceeded!\n Possible workarounds: compile with -fasm, or try another version of gcc.\n"
824 $p =~ s/^\tpushl\s+\%edi\n//;
825 $p =~ s/^\tpushl\s+\%esi\n//;
826 $p =~ s/^\tpushl\s+\%ebx\n//;
827 $p =~ s/^\tmovl\s+\%esi,\s*\d*\(\%esp\)\n//;
828 $p =~ s/^\tmovl\s+\%edi,\s*\d*\(\%esp\)\n//;
829 $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
830 $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
832 # GCC 3.1 is in the habit of adding spurious writes to the
833 # stack in the prologue. Just to be on the safe side,
834 # chuck these over the fence into the main code.
835 while ($p =~ /^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n/) {
836 # print "Spurious instruction: $&";
841 if ($TargetPlatform =~ /^i386-apple-darwin/) {
843 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
845 $pcrel_reg =~ s/(.|\n)*.*___i686\.get_pc_thunk\.([abcd]x)\n(.|\n)*/$2/ or $pcrel_reg = "";
846 $p =~ s/^\s+call\s+___i686\.get_pc_thunk\..x//;
847 $p =~ s/^\"?L\d+\$pb\"?:\n//;
849 if ($pcrel_reg eq "bx") {
850 # Bad gcc. Goes and uses %ebx, our BaseReg, for PIC. Bad gcc.
851 die "Darwin/x86: -fPIC -via-C doesn't work yet, use -fasm. Aborting."
855 } elsif ($TargetPlatform =~ /^x86_64-/) {
856 $p =~ s/^\tpushq\s+\%r(bx|bp|12|13|14)\n//g;
857 $p =~ s/^\tmovq\s+\%r(bx|bp|12|13|14),\s*\d*\(\%rsp\)\n//g;
858 $p =~ s/^\tsubq\s+\$\d+,\s*\%rsp\n//;
860 } elsif ($TargetPlatform =~ /^ia64-/) {
861 $p =~ s/^\t\.prologue .*\n//;
862 $p =~ s/^\t\.save ar\.pfs, r\d+\n\talloc r\d+ = ar\.pfs, 0, 3[12], \d+, 0\n//;
863 $p =~ s/^\t\.fframe \d+\n\tadds r12 = -\d+, r12\n//;
864 $p =~ s/^\t\.save rp, r\d+\n\tmov r\d+ = b0\n//;
865 $p =~ s/^\t\.(mii|mmi)\n//g; # bundling is no longer sensible
866 $p =~ s/^\t;;\n//g; # discard stops
867 $p =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
869 # GCC 3.3 saves r1 in the prologue, move this to the body
870 if ($p =~ /^\tmov r\d+ = r1\n/) {
874 } elsif ($TargetPlatform =~ /^m68k-/) {
875 $p =~ s/^\tlink a6,#-?\d.*\n//;
876 $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//;
877 # The above showed up in the asm code,
878 # so I added it here.
879 # I hope it's correct.
881 $p =~ s/^\tmovel d2,sp\@-\n//;
882 $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only?
883 $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only?
884 } elsif ($TargetPlatform =~ /^mips-/) {
885 # the .frame/.mask/.fmask that we use is the same
886 # as that produced by GCC for miniInterpret; this
887 # gives GDB some chance of figuring out what happened
888 $FRAME = "\t.frame\t\$sp,2168,\$31\n\t.mask\t0x90000000,-4\n\t.fmask\t0x00000000,0\n";
889 $p =~ s/^\t\.(frame).*\n/__FRAME__/g;
890 $p =~ s/^\t\.(mask|fmask).*\n//g;
891 $p =~ s/^\t\.cprestore.*\n/\t\.cprestore 416\n/; # 16 + 100 4-byte args
892 $p =~ s/^\tsubu\t\$sp,\$sp,\d+\n//;
893 $p =~ s/^\tsw\t\$31,\d+\(\$sp\)\n//;
894 $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//;
895 $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//;
896 $p =~ s/__FRAME__/$FRAME/;
897 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
899 $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";
901 $p =~ s/^\tmflr r0\n//;
902 $p =~ s/^\tbl saveFP # f\d+\n//;
903 $p =~ s/^\tbl saveFP ; save f\d+-f\d+\n//;
904 $p =~ s/^\"?L\d+\$pb\"?:\n//;
905 $p =~ s/^\tstmw r\d+,-\d+\(r1\)\n//;
906 $p =~ s/^\tstfd f\d+,-\d+\(r1\)\n//g;
907 $p =~ s/^\tstw r0,\d+\(r1\)\n//g;
908 $p =~ s/^\tstwu r1,-\d+\(r1\)\n//;
909 $p =~ s/^\tstw r\d+,-\d+\(r1\)\n//g;
910 $p =~ s/^\tbcl 20,31,L\d+\$pb\n//;
911 $p =~ s/^L\d+\$pb:\n//;
912 $p =~ s/^\tmflr r31\n//;
914 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
915 # under some circumstances, only when generating position dependent code.
916 # I have no idea why, and I don't think it is necessary, so let's toss it.
917 $p =~ s/^\tli r\d+,0\n//g;
918 $p =~ s/^\tstw r\d+,\d+\(r1\)\n//g;
919 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/) {
920 $p =~ s/^\tmflr 0\n//;
921 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
922 $p =~ s/^\tstfd \d+,\d+\(1\)\n//g;
923 $p =~ s/^\tstw r0,8\(1\)\n//;
924 $p =~ s/^\tstwu 1,-\d+\(1\)\n//;
925 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
927 # GCC's "large-model" PIC (-fPIC)
929 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
931 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
932 $p =~ s/^.LCF\d+:\n//;
933 $p =~ s/^\tmflr 30\n//;
934 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
935 $p =~ s/^\tadd 30,0,30\n//;
937 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
938 # under some circumstances, only when generating position dependent code.
939 # I have no idea why, and I don't think it is necessary, so let's toss it.
940 $p =~ s/^\tli \d+,0\n//g;
941 $p =~ s/^\tstw \d+,\d+\(1\)\n//g;
942 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
943 $p =~ s/^\tmr 31,1\n//;
944 $p =~ s/^\tmflr 0\n//;
945 $p =~ s/^\tstmw \d+,\d+\(1\)\n//;
946 $p =~ s/^\tstfd \d+,-?\d+\(1\)\n//g;
947 $p =~ s/^\tstd r0,8\(1\)\n//;
948 $p =~ s/^\tstdu 1,-\d+\(1\)\n//;
949 $p =~ s/^\tstd \d+,-?\d+\(1\)\n//g;
951 # GCC's "large-model" PIC (-fPIC)
953 $pcrel_label =~ s/(.|\n)*^.LCF(\d+):\n(.|\n)*/$2/ or $pcrel_label = "";
955 $p =~ s/^\tbcl 20,31,.LCF\d+\n//;
956 $p =~ s/^.LCF\d+:\n//;
957 $p =~ s/^\tmflr 30\n//;
958 $p =~ s/^\tlwz 0,\.LCL\d+-\.LCF\d+\(30\)\n//;
959 $p =~ s/^\tadd 30,0,30\n//;
961 # This is bad: GCC 3 seems to zero-fill some local variables in the prologue
962 # under some circumstances, only when generating position dependent code.
963 # I have no idea why, and I don't think it is necessary, so let's toss it.
964 $p =~ s/^\tli \d+,0\n//g;
965 $p =~ s/^\tstd \d+,\d+\(1\)\n//g;
967 print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n";
970 # HWL HACK: dont die, just print a warning
971 #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/;
972 die "Prologue junk?: $p\n" if $p =~ /^\s+[^\s\.]/;
974 # For PIC, we want to keep part of the prologue
975 if ($TargetPlatform =~ /^powerpc-apple-darwin.*/ && $pcrel_label ne "") {
976 # Darwin: load the current instruction pointer into register r31
977 $p .= "bcl 20,31,$pcrel_label\n";
978 $p .= "$pcrel_label:\n";
979 $p .= "\tmflr r31\n";
980 } elsif ($TargetPlatform =~ /^powerpc-.*-linux/ && $pcrel_label ne "") {
981 # Linux: load the GOT pointer into register 30
982 $p .= "\tbcl 20,31,.LCF$pcrel_label\n";
983 $p .= ".LCF$pcrel_label:\n";
985 $p .= "\tlwz 0,.LCL$pcrel_label-.LCF$pcrel_label(30)\n";
986 $p .= "\tadd 30,0,30\n";
987 } elsif ($TargetPlatform =~ /^i386-apple-darwin.*/ && $pcrel_label ne "") {
988 $p .= "\tcall ___i686.get_pc_thunk.$pcrel_reg\n";
989 $p .= "$pcrel_label:\n";
992 # glue together what's left
997 if ( $TargetPlatform =~ /^mips-/ ) {
998 # MIPS: first, this basic sequence may occur "--- END ---" or not
999 $c =~ s/^\tlw\t\$31,\d+\(\$sp\)\n\taddu\t\$sp,\$sp,\d+\n\tj\t\$31\n\t\.end/\t\.end/;
1002 # toss all epilogue stuff; again, paranoidly
1003 if ( $c =~ /--- END ---/ ) {
1004 if (($r, $e) = split(/--- END ---/, $c)) {
1005 if ($TargetPlatform =~ /^i386-/) {
1007 $e =~ s/^\tpopl\s+\%edi\n//;
1008 $e =~ s/^\tpopl\s+\%esi\n//;
1009 $e =~ s/^\tpopl\s+\%edx\n//;
1010 $e =~ s/^\tpopl\s+\%ecx\n//;
1011 $e =~ s/^\taddl\s+\$\d+,\s*\%esp\n//;
1012 $e =~ s/^\tsubl\s+\$-\d+,\s*\%esp\n//;
1013 } elsif ($TargetPlatform =~ /^ia64-/) {
1014 $e =~ s/^\tmov ar\.pfs = r\d+\n//;
1015 $e =~ s/^\tmov b0 = r\d+\n//;
1016 $e =~ s/^\t\.restore sp\n\tadds r12 = \d+, r12\n//;
1017 $e =~ s/^\tbr\.ret\.sptk\.many b0\n//;
1018 $e =~ s/^\t\.(mii|mmi|mib)\n//g; # bundling is no longer sensible
1019 $e =~ s/^\t;;\n//g; # discard stops - stop at end of body is sufficient
1020 $e =~ s/^\t\/\/.*\n//g; # gcc inserts timings in // comments
1021 } elsif ($TargetPlatform =~ /^m68k-/) {
1022 $e =~ s/^\tunlk a6\n//;
1024 } elsif ($TargetPlatform =~ /^mips-/) {
1025 $e =~ s/^\tlw\t\$31,\d+\(\$sp\)\n//;
1026 $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//;
1027 $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//;
1028 $e =~ s/^\tj\t\$31\n//;
1029 } elsif ($TargetPlatform =~ /^powerpc-apple-darwin.*/) {
1030 $e =~ s/^\taddi r1,r1,\d+\n//;
1031 $e =~ s/^\tlwz r\d+,\d+\(r1\)\n//;
1032 $e =~ s/^\tlmw r\d+,-\d+\(r1\)\n//;
1033 $e =~ s/^\tmtlr r0\n//;
1035 $e =~ s/^\tb restFP ;.*\n//;
1036 } elsif ($TargetPlatform =~ /^powerpc64-.*-linux/) {
1037 $e =~ s/^\tmr 3,0\n//;
1038 $e =~ s/^\taddi 1,1,\d+\n//;
1039 $e =~ s/^\tld 0,16\(1\)\n//;
1040 $e =~ s/^\tmtlr 0\n//;
1042 # callee-save registers
1043 $e =~ s/^\tld \d+,-?\d+\(1\)\n//g;
1044 $e =~ s/^\tlfd \d+,-?\d+\(1\)\n//g;
1046 # get rid of the debug junk along with the blr
1047 $e =~ s/^\tblr\n\t.long .*\n\t.byte .*\n//;
1049 # incase we missed it with the last one get the blr alone
1052 print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n";
1055 print STDERR "WARNING: Epilogue junk?: $e\n" if $e =~ /^\t\s*[^\.\s\n]/;
1057 # glue together what's left
1059 $c =~ s/\n\t\n/\n/; # junk blank line
1063 # On SPARCs, we don't do --- BEGIN/END ---, we just
1064 # toss the register-windowing save/restore/ret* instructions
1065 # directly unless they've been generated by function definitions in header
1067 if ( $TargetPlatform =~ /^sparc-/ ) {
1068 if ( ! ( $TargetPlatform =~ /solaris2$/ && $chkcat[$i] eq 'unknown' )) {
1069 $c =~ s/^\t(save.*|restore.*|ret|retl)\n//g;
1071 # throw away PROLOGUE comments
1072 $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//;
1075 # On Alphas, the prologue mangling is done a little later (below)
1077 # toss all calls to __DISCARD__
1078 $c =~ s/^\t(call|jbsr|jal)\s+${T_US}__DISCARD__\n//go;
1079 $c =~ s/^\tjsr\s+\$26\s*,\s*${T_US}__DISCARD__\n//go if $TargetPlatform =~ /^alpha-/;
1080 $c =~ s/^\tbl\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /^powerpc-apple-darwin.*/;
1081 $c =~ s/^\tbl\s+__DISCARD__(\@plt)?\n//go if $TargetPlatform =~ /^powerpc-.*-linux/;
1082 $c =~ s/^\tbl\s+\.__DISCARD__\n\s+nop\n//go if $TargetPlatform =~ /^powerpc64-.*-linux/;
1083 $c =~ s/^\tcall\s+L___DISCARD__\$stub\n//go if $TargetPlatform =~ /i386-apple-darwin.*/;
1085 # IA64: mangle tailcalls into jumps here
1086 if ($TargetPlatform =~ /^ia64-/) {
1087 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/) {
1088 # Eek, the gcc optimiser is getting smarter... if we see a jump to the --- TAILCALL ---
1089 # marker then we reapply the substitution at the source sites
1090 $c =~ s/^\tbr \.L$2\n/\t--- TAILCALL ---\n/g if ($2);
1094 # MIPS: that may leave some gratuitous asm macros around
1095 # (no harm done; but we get rid of them to be tidier)
1096 $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/
1097 if $TargetPlatform =~ /^mips-/;
1099 # toss stack adjustment after DoSparks
1100 $c =~ s/^(\tjbsr _DoSparks\n)\taddqw #8,sp/$1/g
1101 if $TargetPlatform =~ /^m68k-/; # this looks old...
1103 if ( $TargetPlatform =~ /^alpha-/ &&
1104 ! $magic_rdata_seen &&
1105 $c =~ /^\s*\.rdata\n\t\.quad 0\n\t\.align \d\n/ ) {
1106 $c =~ s/^\s*\.rdata\n\t\.quad 0\n\t\.align (\d)\n/\.rdata\n\t\.align $1\n/;
1107 $magic_rdata_seen = 1;
1110 # pick some end-things and move them to the next chunk
1112 # pin a funny end-thing on (for easier matching):
1113 $c .= 'FUNNY#END#THING';
1115 while ( $c =~ /${T_MOVE_DIRVS}FUNNY#END#THING/o ) {
1119 # on x86 we try not to copy any directives into a literal
1120 # chunk, rather we keep looking for the next real chunk. This
1121 # is because we get things like
1123 # .globl blah_closure
1129 if ( $TargetPlatform =~ /^(i386|sparc|powerpc)/ && $to_move =~ /${T_COPY_DIRVS}/ ) {
1131 while ( $j < $numchks && $chk[$j] =~ /$T_CONST_LBL/) {
1134 if ( $j < $numchks ) {
1135 $chk[$j] = $to_move . $chk[$j];
1139 elsif ( $i < ($numchks - 1)
1140 && ( $to_move =~ /${T_COPY_DIRVS}/
1141 || ($TargetPlatform =~ /^hppa/ && $to_move =~ /align/ && $chkcat[$i+1] eq 'literal') )) {
1142 $chk[$i + 1] = $to_move . $chk[$i + 1];
1143 # otherwise they're tossed
1146 $c =~ s/${T_MOVE_DIRVS}FUNNY#END#THING/FUNNY#END#THING/o;
1149 if ( $TargetPlatform =~ /^alpha-/ && $c =~ /^\t\.ent\s+(\S+)/ ) {
1151 # toss all prologue stuff, except for loading gp, and the ..ng address
1152 unless ($c =~ /\.ent.*\n\$.*\.\.ng:/) {
1153 if (($p, $r) = split(/^\t\.prologue/, $c)) {
1154 if (($keep, $junk) = split(/\.\.ng:/, $p)) {
1155 $keep =~ s/^\t\.frame.*\n/\t.frame \$30,0,\$26,0\n/;
1156 $keep =~ s/^\t\.(mask|fmask).*\n//g;
1157 $c = $keep . "..ng:\n";
1159 print STDERR "malformed code block ($ent)?\n"
1162 $c .= "\t.prologue" . $r;
1166 $c =~ s/FUNNY#END#THING//;
1168 # print STDERR "\nCHK $i (AFTER) (",$chkcat[$i],"):\n", $c;
1170 $chk[$i] = $c; # update w/ convenience copy
1173 # open CHUNKS, ">/tmp/chunks2" or die "Cannot open /tmp/chunks2: $!\n";
1174 # for (my $i = 0; $i < @chk; ++$i) { print CHUNKS "======= $i =======\n", $chk[$i] }
1177 if ( $TargetPlatform =~ /^alpha-/ ) {
1178 # print out the header stuff first
1179 $chk[0] =~ s/^(\t\.file.*)"(ghc\d+\.c)"/$1"$ifile_root.hc"/;
1180 print OUTASM $chk[0];
1182 } elsif ( $TargetPlatform =~ /^hppa/ ) {
1183 print OUTASM $chk[0];
1185 } elsif ( $TargetPlatform =~ /^mips-/ ) {
1186 $chk[0] = "\t\.file\t1 \"$ifile_root.hc\"\n" . $chk[0];
1188 # get rid of horrible "<dollar>Revision: .*$" strings
1189 local(@lines0) = split(/\n/, $chk[0]);
1191 while ( $z <= $#lines0 ) {
1192 if ( $lines0[$z] =~ /^\t\.byte\t0x24,0x52,0x65,0x76,0x69,0x73,0x69,0x6f$/ ) {
1195 while ( $z <= $#lines0 ) {
1197 last if $lines0[$z] =~ /[,\t]0x0$/;
1203 $chk[0] = join("\n", @lines0);
1204 $chk[0] =~ s/\n\n+/\n/;
1205 print OUTASM $chk[0];
1208 # print out all the literal strings next
1209 for ($i = 0; $i < $numchks; $i++) {
1210 if ( $chkcat[$i] eq 'literal' ) {
1212 # HACK: try to detect 16-byte constants and align them
1213 # on a 16-byte boundary. x86_64 sometimes needs 128-bit
1214 # aligned constants.
1215 if ( $TargetPlatform =~ /^x86_64/ ) {
1217 if ($z =~ /(\.long.*\n.*\.long.*\n.*\.long.*\n.*\.long|\.quad.*\n.*\.quad)/) {
1218 print OUTASM $T_HDR_literal16;
1220 print OUTASM $T_HDR_literal;
1223 print OUTASM $T_HDR_literal;
1226 print OUTASM $chk[$i];
1227 print OUTASM "; end literal\n" if $TargetPlatform =~ /^hppa/; # for the splitter
1229 $chkcat[$i] = 'DONE ALREADY';
1233 # on the HPPA, print out all the bss next
1234 if ( $TargetPlatform =~ /^hppa/ ) {
1235 for ($i = 1; $i < $numchks; $i++) {
1236 if ( $chkcat[$i] eq 'bss' ) {
1237 print OUTASM "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$BSS\$\n\t.align 4\n";
1238 print OUTASM $chk[$i];
1240 $chkcat[$i] = 'DONE ALREADY';
1245 for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) {
1246 # print STDERR "$i: cat $chkcat[$i], symb $chksymb[$i]\n";
1248 next if $chkcat[$i] eq 'DONE ALREADY';
1250 if ( $chkcat[$i] eq 'misc' || $chkcat[$i] eq 'unknown' ) {
1251 if ($chk[$i] ne '') {
1252 print OUTASM $T_HDR_misc;
1253 &print_doctored($chk[$i], 0);
1256 } elsif ( $chkcat[$i] eq 'toss' ) {
1257 print STDERR "*** NB: TOSSING code for $chksymb[$i] !!! ***\n";
1259 } elsif ( $chkcat[$i] eq 'data' ) {
1260 if ($chk[$i] ne '') {
1261 print OUTASM $T_HDR_data;
1262 print OUTASM $chk[$i];
1265 } elsif ( $chkcat[$i] eq 'splitmarker' ) {
1266 # we can just re-constitute this one...
1267 # NB: we emit _three_ underscores no matter what,
1268 # so ghc-split doesn't have to care.
1269 print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n";
1271 } elsif ( $chkcat[$i] eq 'closure'
1272 || $chkcat[$i] eq 'srt'
1273 || $chkcat[$i] eq 'infotbl'
1274 || $chkcat[$i] eq 'entry') { # do them in that order
1275 $symb = $chksymb[$i];
1278 if ( defined($closurechk{$symb}) ) {
1279 print OUTASM $T_HDR_closure;
1280 print OUTASM $chk[$closurechk{$symb}];
1281 $chkcat[$closurechk{$symb}] = 'DONE ALREADY';
1285 if ( defined($srtchk{$symb}) ) {
1286 print OUTASM $T_HDR_relrodata;
1287 print OUTASM $chk[$srtchk{$symb}];
1288 $chkcat[$srtchk{$symb}] = 'DONE ALREADY';
1292 if ( defined($infochk{$symb}) ) {
1294 print OUTASM $T_HDR_info;
1295 print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1);
1297 # entry code will be put here!
1299 $chkcat[$infochk{$symb}] = 'DONE ALREADY';
1303 if ( defined($entrychk{$symb}) ) {
1305 $c = $chk[$entrychk{$symb}];
1307 # If this is an entry point with an info table,
1308 # eliminate the entry symbol and all directives involving it.
1309 if (defined($infochk{$symb}) && $TargetPlatform !~ /^ia64-/) {
1311 foreach $l (split(/\n/,$c)) {
1312 next if $l =~ /^.*$symb_(entry|ret)${T_POST_LBL}/;
1314 # If we have .type/.size direrctives involving foo_entry,
1315 # then make them refer to foo_info instead. The information
1316 # in these directives is used by the cachegrind annotator,
1317 # so it is worthwhile keeping.
1318 if ($l =~ /^\s*\.(type|size).*$symb_(entry|ret)/) {
1319 $l =~ s/$symb(_entry|_ret)/${symb}_info/g;
1323 next if $l =~ /^\s*\..*$symb.*\n?/;
1326 $c = join("\n",@o) . "\n";
1329 print OUTASM $T_HDR_entry;
1331 &print_doctored($c, 1); # NB: the 1!!!
1333 $chkcat[$entrychk{$symb}] = 'DONE ALREADY';
1336 } elsif ( $chkcat[$i] eq 'vector' ) {
1337 $symb = $chksymb[$i];
1340 if ( defined($vectorchk{$symb}) ) {
1341 print OUTASM $T_HDR_vector;
1342 print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0);
1344 # direct return code will be put here!
1345 $chkcat[$vectorchk{$symb}] = 'DONE ALREADY';
1347 } elsif ( $TargetPlatform =~ /^alpha-/ ) {
1348 # Alphas: the commented nop is for the splitter, to ensure
1349 # that no module ends with a label as the very last
1350 # thing. (The linker will adjust the label to point
1351 # to the first code word of the next module linked in,
1352 # even if alignment constraints cause the label to move!)
1354 print OUTASM "\t# nop\n";
1357 } elsif ( $chkcat[$i] eq 'rodata' ) {
1358 print OUTASM $T_HDR_rodata;
1359 print OUTASM $chk[$i];
1360 $chkcat[$i] = 'DONE ALREADY';
1361 } elsif ( $chkcat[$i] eq 'relrodata' ) {
1362 print OUTASM $T_HDR_relrodata;
1363 print OUTASM $chk[$i];
1364 $chkcat[$i] = 'DONE ALREADY';
1365 } elsif ( $chkcat[$i] eq 'toc' ) {
1366 # silly optimisation to print tocs, since they come in groups...
1367 print OUTASM $T_HDR_toc;
1369 while ($chkcat[$j] eq 'toc')
1370 { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump.
1373 print OUTASM $chk[$j];
1375 $chkcat[$j] = 'DONE ALREADY';
1379 } elsif ( $TargetPlatform =~ /^.*-apple-darwin.*/ && $chkcat[$i] eq 'dyld' ) {
1380 # apple-darwin: dynamic linker stubs
1381 if($chk[$i] !~ /\.indirect_symbol ___DISCARD__/)
1382 { # print them out unchanged, but remove the stubs for __DISCARD__
1383 print OUTASM $chk[$i];
1385 } elsif ( $TargetPlatform =~ /^i386-apple-darwin.*/ && $chkcat[$i] eq 'get_pc_thunk' ) {
1386 # i386-apple-darwin: __i686.get_pc_thunk.[abcd]x
1387 print OUTASM ".section __TEXT,__textcoal_nt,coalesced,no_toc\n";
1388 print OUTASM $chk[$i];
1390 &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n");
1394 print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/;
1397 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1398 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1403 sub hppa_mash_prologue { # OK, epilogue, too
1406 # toss all prologue stuff
1407 s/^\s+\.ENTRY[^\0]*--- BEGIN ---/\t.ENTRY/;
1409 # Lie about our .CALLINFO
1410 s/^\s+\.CALLINFO.*$/\t.CALLINFO NO_CALLS,NO_UNWIND/;
1417 # toss all epilogue stuff
1418 s/^\s+--- END ---[^\0]*\.EXIT/\t.EXIT/;
1420 # Sorry; we moved the _info stuff to the code segment.
1421 s/_info,DATA/_info,CODE/g;
1428 sub print_doctored {
1429 local($_, $need_fallthru_patch) = @_;
1431 if ( $TargetPlatform !~ /^i386-/
1432 || ! /^\t[a-z]/ # no instructions in here, apparently
1433 || /^${T_US}__stginit_[A-Za-z0-9_]+${T_POST_LBL}/) {
1437 # OK, must do some x86 **HACKING**
1439 local($entry_patch) = '';
1440 local($exit_patch) = '';
1442 # gotta watch out for weird instructions that
1443 # invisibly smash various regs:
1444 # rep* %ecx used for counting
1445 # scas* %edi used for destination index
1446 # cmps* %e[sd]i used for indices
1447 # loop* %ecx used for counting
1452 # * use of STG reg [ nn(%ebx) ] where no machine reg avail
1454 # * GCC used an "STG reg" for its own purposes
1456 # * some secret uses of machine reg, requiring STG reg
1457 # to be saved/restored
1459 # The most dangerous "GCC uses" of an "STG reg" are when
1460 # the reg holds the target of a jmp -- it's tricky to
1461 # insert the patch-up code before we get to the target!
1462 # So here we change the jmps:
1464 # --------------------------------------------------------
1465 # it can happen that we have jumps of the form...
1466 # jmp *<something involving %esp>
1468 # jmp <something involving another naughty register...>
1470 # a reasonably-common case is:
1472 # movl $_blah,<bad-reg>
1475 s/^\tmovl\s+\$${T_US}(.*),\s*(\%e[acd]x)\n\tjmp\s+\*\2/\tjmp $T_US$1/g;
1479 # movl -4(%ebx), %eax
1484 s/^\tmovl\s+(-?\d*\(\%e(bx|si)\)),\s*(\%e[acd]x)\n\tjmp\s+\*\3/\tjmp\t\*$1/g;
1486 if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi?
1487 s/^\tmovl\s+(.*),\s*\%esi\n\tjmp\s+\*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1488 s/^\tjmp\s+\*(.*\(.*\%esi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1489 s/^\tjmp\s+\*\%esi\n/\tmovl \%esi,\%eax\n\tjmp \*\%eax\n/g;
1490 die "$Pgm: (mangler) still have jump involving \%esi!\n$_"
1491 if /(jmp|call)\s+.*\%esi/;
1493 if ($StolenX86Regs <= 3 ) { # spurious uses of edi?
1494 s/^\tmovl\s+(.*),\s*\%edi\n\tjmp\s+\*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1495 s/^\tjmp\s+\*(.*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g;
1496 s/^\tjmp\s+\*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g;
1497 die "$Pgm: (mangler) still have jump involving \%edi!\n$_"
1498 if /(jmp|call)\s+.*\%edi/;
1501 # OK, now we can decide what our patch-up code is going to
1504 # Offsets into register table - you'd better update these magic
1505 # numbers should you change its contents!
1506 # local($OFFSET_R1)=0; No offset for R1 in new RTS.
1507 local($OFFSET_Hp)=88;
1509 # Note funky ".=" stuff; we're *adding* to these _patch guys
1510 if ( $StolenX86Regs <= 2
1511 && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi)
1512 $entry_patch .= "\tmovl \%esi,(\%ebx)\n";
1513 $exit_patch .= "\tmovl (\%ebx),\%esi\n";
1515 # nothing for call_{entry,exit} because %esi is callee-save
1517 if ( $StolenX86Regs <= 3
1518 && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi)
1519 $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n";
1520 $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n";
1522 # nothing for call_{entry,exit} because %edi is callee-save
1525 # --------------------------------------------------------
1526 # next, here we go with non-%esp patching!
1528 s/^(\t[a-z])/$entry_patch$1/; # before first instruction
1530 # Before calling GC we must set up the exit condition before the call
1531 # and entry condition when we come back
1533 # fix _all_ non-local jumps:
1535 s/^\tjmp\s+\*${T_X86_PRE_LLBL_PAT}/\tJMP___SL/go;
1536 s/^\tjmp\s+${T_X86_PRE_LLBL_PAT}/\tJMP___L/go;
1538 s/^(\tjmp\s+.*\n)/$exit_patch$1/g; # here's the fix...
1540 s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go;
1541 s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go;
1543 if ($StolenX86Regs == 2 ) {
1544 die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_"
1545 if /^\t(jmp|call)\s+.*\%e(si|di)/;
1546 } elsif ($StolenX86Regs == 3 ) {
1547 die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_"
1548 if /^\t(jmp|call)\s+.*\%edi/;
1551 # --------------------------------------------------------
1552 # that's it -- print it
1554 #die "Funny jumps?\n$_" if /${T_X86_BADJMP}/o; # paranoia
1558 if ( $need_fallthru_patch ) { # exit patch for end of slow entry code
1559 print OUTASM $exit_patch;
1560 # ToDo: make it not print if there is a "jmp" at the end
1566 sub init_FUNNY_THINGS {
1567 %KNOWN_FUNNY_THING = (
1569 # "${T_US}stg_.*{T_POST_LBL}", 1,
1574 The following table reversal is used for both info tables and return
1575 vectors. In both cases, we remove the first entry from the table,
1576 reverse the table, put the label at the end, and paste some code
1577 (that which is normally referred to by the first entry in the table)
1578 right after the table itself. (The code pasting is done elsewhere.)
1582 local($symb, $tbl, $discard1) = @_;
1584 return ($tbl) if ($TargetPlatform =~ /^ia64-/);
1586 local($before) = '';
1588 local(@imports) = (); # hppa only
1591 local(@lines) = split(/\n/, $tbl);
1594 # Deal with the header...
1595 for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1596 $label .= $lines[$i] . "\n",
1597 next if $lines[$i] =~ /^[A-Za-z0-9_]+_info${T_POST_LBL}$/o
1598 || $lines[$i] =~ /${T_DOT_GLOBAL}/o
1599 || $lines[$i] =~ /^${T_US}\S+_vtbl${T_POST_LBL}$/o;
1601 $before .= $lines[$i] . "\n"; # otherwise...
1605 $infoname =~ s/(.|\n)*^([A-Za-z0-9_]+_info)${T_POST_LBL}$(.|\n)*/\2/;
1607 # Grab the table data...
1608 if ( $TargetPlatform !~ /^hppa/ ) {
1609 for ( ; $i <= $#lines && $lines[$i] =~ /^\t?${T_DOT_WORD}\s+/o; $i++) {
1611 # Convert addresses of SRTs, slow entrypoints and large bitmaps
1612 # to offsets (relative to the info label),
1613 # in order to support position independent code.
1614 $line =~ s/$infoname/0/
1615 || $line =~ s/([A-Za-z0-9_]+_srtd)$/\1 - $infoname/
1616 || $line =~ s/([A-Za-z0-9_]+_srt(\+\d+)?)$/\1 - $infoname/
1617 || $line =~ s/([A-Za-z0-9_]+_slow)$/\1 - $infoname/
1618 || $line =~ s/([A-Za-z0-9_]+_btm)$/\1 - $infoname/
1619 || $line =~ s/([A-Za-z0-9_]+_alt)$/\1 - $infoname/
1620 || $line =~ s/([A-Za-z0-9_]+_dflt)$/\1 - $infoname/
1621 || $line =~ s/([A-Za-z0-9_]+_ret)$/\1 - $infoname/;
1622 push(@words, $line);
1624 } else { # hppa weirdness
1625 for ( ; $i <= $#lines && $lines[$i] =~ /^\s+(${T_DOT_WORD}|\.IMPORT)/; $i++) {
1626 # FIXME: the RTS now expects offsets instead of addresses
1627 # for all labels in info tables.
1628 if ($lines[$i] =~ /^\s+\.IMPORT/) {
1629 push(@imports, $lines[$i]);
1631 # We don't use HP's ``function pointers''
1632 # We just use labels in code space, like normal people
1633 $lines[$i] =~ s/P%//;
1634 push(@words, $lines[$i]);
1639 # Now throw away any initial zero word from the table. This is a hack
1640 # that lets us reduce the size of info tables when the SRT field is not
1641 # needed: see comments StgFunInfoTable in InfoTables.h.
1643 # The .zero business is for Linux/ELF.
1644 # The .skip business is for Sparc/Solaris/ELF.
1645 # The .blockz business is for HPPA.
1647 # if ($words[0] =~ /^\t?(${T_DOT_WORD}\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) {
1652 for (; $i <= $#lines; $i++) {
1653 $after .= $lines[$i] . "\n";
1656 # Alphas: If we have anonymous text (not part of a procedure), the
1657 # linker may complain about missing exception information. Bleh.
1658 # To suppress this, we place a .ent/.end pair around the code.
1659 # At the same time, we have to be careful and not enclose any leading
1660 # .file/.loc directives.
1661 if ( $TargetPlatform =~ /^alpha-/ && $label =~ /^([A-Za-z0-9_]+):$/) {
1662 local ($ident) = $1;
1663 $before =~ s/^((\s*\.(file|loc)\s+[^\n]*\n)*)/$1\t.ent $ident\n/;
1664 $after .= "\t.end $ident\n";
1667 # Alphas: The heroic Simon Marlow found a bug in the Digital UNIX
1668 # assembler (!) wherein .quad constants inside .text sections are
1669 # first narrowed to 32 bits then sign-extended back to 64 bits.
1670 # This obviously screws up our 64-bit bitmaps, so we work around
1671 # the bug by replacing .quad with .align 3 + .long + .long [ccshan]
1672 if ( $TargetPlatform =~ /^alpha-/ ) {
1674 if (/^\s*\.quad\s+([-+0-9].*\S)\s*$/ && length $1 >= 10) {
1675 local ($number) = $1;
1676 if ($number =~ /^([-+])?(0x?)?([0-9]+)$/) {
1677 local ($sign, $base, $digits) = ($1, $2, $3);
1678 $base = (10, 8, 16)[length $base];
1679 local ($hi, $lo) = (0, 0);
1680 foreach $i (split(//, $digits)) {
1681 $j = $lo * $base + $i;
1682 $lo = $j % 4294967296;
1683 $hi = $hi * $base + ($j - $lo) / 4294967296;
1685 ($hi, $lo) = (4294967295 - $hi, 4294967296 - $lo)
1687 $_ = "\t.align 3\n\t.long $lo\n\t.long $hi\n";
1688 # printf STDERR "TURNING %s into 0x %08x %08x\n", $number, $hi, $lo;
1690 print STDERR "Cannot handle \".quad $number\" in info table\n";
1698 . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n")
1699 . join("\n", @words) . "\n"
1702 # print STDERR "before=$before\n";
1703 # print STDERR "label=$label\n";
1704 # print STDERR "words=",(reverse @words),"\n";
1705 # print STDERR "after=$after\n";
1711 The HP is a major nuisance. The threaded code mangler moved info
1712 tables from data space to code space, but unthreaded code in the RTS
1713 still has references to info tables in data space. Since the HP
1714 linker is very precise about where symbols live, we need to patch the
1715 references in the unthreaded RTS as well.
1718 sub mini_mangle_asm_hppa {
1719 local($in_asmf, $out_asmf) = @_;
1721 open(INASM, "< $in_asmf")
1722 || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n");
1723 open(OUTASM,"> $out_asmf")
1724 || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n");
1727 s/_info,DATA/_info,CODE/; # Move _info references to code space
1733 close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n");
1734 close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n");
1740 sub tidy_up_and_die {
1741 local($return_val, $msg) = @_;
1743 exit (($return_val == 0) ? 0 : 1);