From: simonpj Date: Tue, 13 Jun 2000 15:35:29 +0000 (+0000) Subject: [project @ 2000-06-13 15:35:29 by simonpj] X-Git-Tag: Approximately_9120_patches~4222 X-Git-Url: http://git.megacz.com/?a=commitdiff_plain;h=77bdc0d10fa1508ef9dcd6c02c5a00a590fa6272;p=ghc-hetmet.git [project @ 2000-06-13 15:35:29 by simonpj] More small changes towards 2.02 --- diff --git a/ghc/driver/mangler/ghc-asm.lprl b/ghc/driver/mangler/ghc-asm.lprl index 6139b3c..89cc4b1 100644 --- a/ghc/driver/mangler/ghc-asm.lprl +++ b/ghc/driver/mangler/ghc-asm.lprl @@ -54,6 +54,7 @@ sub init_TARGET_STUFF { $T_STABBY = 0; # 1 iff .stab things (usually if a.out format) $T_US = ''; # _ if symbols have an underscore on the front + $T_DO_GC = 'PerformGC_wrapper'; $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP $T_CONST_LBL = '^\$C(\d+):$'; # regexp for what such a lbl looks like $T_POST_LBL = ':'; @@ -69,7 +70,6 @@ sub init_TARGET_STUFF { $T_HDR_data = "\.data\n\t\.align 3\n"; $T_HDR_consist = "\.text\n"; $T_HDR_closure = "\.data\n\t\.align 3\n"; - $T_HDR_srt = "\.text\n\t\.align 3\n"; $T_HDR_info = "\.text\n\t\.align 3\n"; $T_HDR_entry = "\.text\n\t\.align 3\n"; $T_HDR_fast = "\.text\n\t\.align 3\n"; @@ -81,6 +81,7 @@ sub init_TARGET_STUFF { $T_STABBY = 0; # 1 iff .stab things (usually if a.out format) $T_US = ''; # _ if symbols have an underscore on the front + $T_DO_GC = 'PerformGC_wrapper'; $T_PRE_APP = 'DONT THINK THIS APPLIES'; # regexp that says what comes before APP/NO_APP $T_CONST_LBL = '^L\$C(\d+)$'; # regexp for what such a lbl looks like $T_POST_LBL = ''; @@ -88,15 +89,14 @@ sub init_TARGET_STUFF { $T_MOVE_DIRVS = '^((\s+\.(IMPORT|EXPORT|PARAM).*|\s+\.align\s+\d+|\s+\.(SPACE|SUBSPA)\s+\S+|\s*)\n)'; $T_COPY_DIRVS = '^\s+\.(IMPORT|EXPORT)'; - $T_hsc_cc_PAT = '\.STRING.*\)(hsc|cc) (.*)\\\\x09(.*)\\\\x00"'; - $T_DOT_WORD = '\.(blockz|word|half|byte)'; + $T_hsc_cc_PAT = '\.STRING.*\)(hsc|cc) (.*)\\\\x09(.*)\\\\x00'; + $T_DOT_WORD = '\.word'; $T_DOT_GLOBAL = '^\s+\.EXPORT'; $T_HDR_literal = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n"; $T_HDR_misc = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n"; $T_HDR_data = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n"; $T_HDR_consist = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$LIT\$\n"; $T_HDR_closure = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n"; - $T_HDR_srt = "\t.SPACE \$PRIVATE\$\n\t.SUBSPA \$DATA\$\n\t\.align 4\n"; $T_HDR_info = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n"; $T_HDR_entry = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n"; $T_HDR_fast = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n"; @@ -104,11 +104,11 @@ sub init_TARGET_STUFF { $T_HDR_direct = "\t.SPACE \$TEXT\$\n\t.SUBSPA \$CODE\$\n\t\.align 4\n"; #--------------------------------------------------------# - } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd2|netbsd|nextstep3|cygwin32|mingw32)$/ ) { - # NeXT added but not tested. CaS + } elsif ( $TargetPlatform =~ /^i386-.*-(linuxaout|freebsd|cygwin32)/ ) { $T_STABBY = 1; # 1 iff .stab things (usually if a.out format) $T_US = '_'; # _ if symbols have an underscore on the front + $T_DO_GC = '_PerformGC_wrapper'; $T_PRE_APP = '^#'; # regexp that says what comes before APP/NO_APP $T_CONST_LBL = '^LC(\d+):$'; $T_POST_LBL = ':'; @@ -116,18 +116,16 @@ sub init_TARGET_STUFF { $T_X86_PRE_LLBL = 'L'; $T_X86_BADJMP = '^\tjmp [^L\*]'; - $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,0x90)?|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*)\n)'; + $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+(,0x90)?|\.globl\s+\S+|\.text|\.data|\.stab[^n].*|\.type\s+.*|\.size\s+.*)\n)'; $T_COPY_DIRVS = '\.(globl|stab)'; $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"'; - $T_DOT_WORD = '\.(long|word|value|byte|space)'; + $T_DOT_WORD = '\.long'; $T_DOT_GLOBAL = '\.globl'; $T_HDR_literal = "\.text\n\t\.align 2\n"; $T_HDR_misc = "\.text\n\t\.align 2,0x90\n"; $T_HDR_data = "\.data\n\t\.align 2\n"; $T_HDR_consist = "\.text\n"; $T_HDR_closure = "\.data\n\t\.align 2\n"; - $T_HDR_closure = "\.data\n\t\.align 2\n\t.long 0\n" if ( $TargetPlatform =~ /.*-mingw32$/ ); - $T_HDR_srt = "\.text\n\t\.align 2\n"; $T_HDR_info = "\.text\n\t\.align 2\n"; # NB: requires padding $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?) $T_HDR_fast = "\.text\n\t\.align 2,0x90\n"; @@ -135,46 +133,42 @@ sub init_TARGET_STUFF { $T_HDR_direct = "\.text\n\t\.align 2,0x90\n"; #--------------------------------------------------------# - } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux|freebsd|netbsd_elf)$/ ) { + } elsif ( $TargetPlatform =~ /^i386-.*-(solaris2|linux)$/ ) { $T_STABBY = 0; # 1 iff .stab things (usually if a.out format) $T_US = ''; # _ if symbols have an underscore on the front + $T_DO_GC = 'PerformGC_wrapper'; $T_PRE_APP = # regexp that says what comes before APP/NO_APP - ($TargetPlatform =~ /-(linux|freebsd|netbsd_elf)$/) ? '#' : '/' ; + ($TargetPlatform =~ /-linux$/) ? '#' : '/' ; $T_CONST_LBL = '^\.LC(\d+):$'; # regexp for what such a lbl looks like $T_POST_LBL = ':'; $T_X86_PRE_LLBL_PAT = '\.L'; $T_X86_PRE_LLBL = '.L'; $T_X86_BADJMP = '^\tjmp [^\.\*]'; - $T_MOVE_DIRVS = '^(\s*(\.(p2)?align\s+\d+(,0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.Lfe.*\n\t\.size\s+.*|\.size\s+.*|\.ident.*)\n)'; + $T_MOVE_DIRVS = '^(\s*(\.align\s+\d+(,0x90)?|\.globl\s+\S+|\.text|\.data|\.section\s+.*|\.type\s+.*|\.Lfe.*\n\t\.size\s+.*|\.size\s+.*|\.ident.*)\n)'; $T_COPY_DIRVS = '\.(globl)'; - if ( $TargetPlatform =~ /freebsd|netbsd_elf/ ) { - $T_hsc_cc_PAT = '\.ascii.*\)(hsc|cc) (.*)\\\\11"\n\t\.ascii\s+"(.*)\\\\0"'; - } else { - $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"'; - } - - $T_DOT_WORD = '\.(long|value|byte|zero)'; + $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"'; + $T_DOT_WORD = '\.long'; $T_DOT_GLOBAL = '\.globl'; $T_HDR_literal = "\.section\t\.rodata\n"; # or just use .text??? (WDP 95/11) - $T_HDR_misc = "\.text\n\t\.align 4\n"; + $T_HDR_misc = "\.text\n\t\.align 16\n"; $T_HDR_data = "\.data\n\t\.align 4\n"; # ToDo: change align?? $T_HDR_consist = "\.text\n"; $T_HDR_closure = "\.data\n\t\.align 4\n"; # ToDo: change align? - $T_HDR_srt = "\.text\n\t\.align 4\n"; # ToDo: change align? - $T_HDR_info = "\.text\n\t\.align 4\n"; # NB: requires padding + $T_HDR_info = "\.text\n\t\.align 16\n"; # NB: requires padding $T_HDR_entry = "\.text\n"; # no .align so we're right next to _info (arguably wrong...?) - $T_HDR_fast = "\.text\n\t\.align 4\n"; - $T_HDR_vector = "\.text\n\t\.align 4\n"; # NB: requires padding - $T_HDR_direct = "\.text\n\t\.align 4\n"; + $T_HDR_fast = "\.text\n\t\.align 16\n"; + $T_HDR_vector = "\.text\n\t\.align 16\n"; # NB: requires padding + $T_HDR_direct = "\.text\n\t\.align 16\n"; #--------------------------------------------------------# } elsif ( $TargetPlatform =~ /^m68k-.*-sunos4/ ) { $T_STABBY = 1; # 1 iff .stab things (usually if a.out format) $T_US = '_'; # _ if symbols have an underscore on the front + $T_DO_GC = '_PerformGC_wrapper'; $T_PRE_APP = '^# MAY NOT APPLY'; # regexp that says what comes before APP/NO_APP $T_CONST_LBL = '^LC(\d+):$'; $T_POST_LBL = ':'; @@ -190,7 +184,6 @@ sub init_TARGET_STUFF { $T_HDR_data = "\.data\n\t\.even\n"; $T_HDR_consist = "\.text\n"; $T_HDR_closure = "\.data\n\t\.even\n"; - $T_HDR_srt = "\.text\n\t\.even\n"; $T_HDR_info = "\.text\n\t\.even\n"; $T_HDR_entry = "\.text\n\t\.even\n"; $T_HDR_fast = "\.text\n\t\.even\n"; @@ -202,6 +195,7 @@ sub init_TARGET_STUFF { $T_STABBY = 0; # 1 iff .stab things (usually if a.out format) $T_US = ''; # _ if symbols have an underscore on the front + $T_DO_GC = 'PerformGC_wrapper'; $T_PRE_APP = '^\s*#'; # regexp that says what comes before APP/NO_APP $T_CONST_LBL = '^\$LC(\d+):$'; # regexp for what such a lbl looks like $T_POST_LBL = ':'; @@ -217,7 +211,6 @@ sub init_TARGET_STUFF { $T_HDR_data = "\t\.data\n\t\.align 2\n"; $T_HDR_consist = 'TOO LAZY TO DO THIS TOO'; $T_HDR_closure = "\t\.data\n\t\.align 2\n"; - $T_HDR_srt = "\t\.text\n\t\.align 2\n"; $T_HDR_info = "\t\.text\n\t\.align 2\n"; $T_HDR_entry = "\t\.text\n\t\.align 2\n"; $T_HDR_fast = "\t\.text\n\t\.align 2\n"; @@ -225,38 +218,38 @@ sub init_TARGET_STUFF { $T_HDR_direct = "\t\.text\n\t\.align 2\n"; #--------------------------------------------------------# - } elsif ( $TargetPlatform =~ /^powerpc-.*|^rs6000-.*/ ) { + } elsif ( $TargetPlatform =~ /^powerpc-.*/ ) { $T_STABBY = 0; # 1 iff .stab things (usually if a.out format) $T_US = ''; # _ if symbols have an underscore on the front + $T_DO_GC = 'PerformGC_wrapper'; $T_PRE_APP = 'NOT APPLICABLE'; # regexp that says what comes before APP/NO_APP - $T_CONST_LBL = 'NOT APPLICABLE'; # regexp for what such a lbl looks like + $T_CONST_LBL = '^LC\.\.(\d+):$'; # regexp for what such a lbl looks like $T_POST_LBL = ':'; - $T_MOVE_DIRVS = '^(\s*(\.toc|\.align \d+|\.csect \S+|\t\.?l?globl \S+)\n)'; + $T_MOVE_DIRVS = '^(\s*(\.toc|.csect \S+|\.l?globl \S+|\.align \d+)\n)'; $T_COPY_DIRVS = '\.(l?globl)'; - $T_hsc_cc_PAT = '\.byte.*\)(hsc|cc) (.*)"\n\t\.byte \d+\n\t\.byte "(.*)"\n\t\.byte \d+'; + $T_hsc_cc_PAT = '\.string.*\)(hsc|cc) (.*)\\\\t(.*)"'; $T_DOT_WORD = '\.long'; $T_DOT_GLOBAL = '\.globl'; - $T_HDR_toc = "\.toc\n"; - $T_HDR_literal = "\.csect .data[RW]\n\t\.align 2\n"; #not RO!? - $T_HDR_misc = "# misc\n\.csect \.text[PR]\n\t\.align 2\n"; - $T_HDR_data = "# data\n\.csect \.data[RW]\n\t\.align 2\n"; - $T_HDR_consist = "# consist\n\.csect \.data[RW]\n\t\.align 2\n"; - $T_HDR_closure = "# closure\n\.csect \.data[RW]\n\t\.align 2\n"; - $T_HDR_srt = "# closure\n\.csect \.data[RW]\n\t\.align 2\n"; - $T_HDR_info = "# info\n\.csect \.data[RW]\n\t\.align 2\n"; #not RO!? - $T_HDR_entry = "# entry\n\.csect \.text[PR]\n\t\.align 2\n"; - $T_HDR_fast = "# fast\n\.csect \.text[PR]\n\t\.align 2\n"; - $T_HDR_vector = "# vector\n\.csect \.data[RW]\n\t\.align 2\n"; #not RO!? - $T_HDR_direct = "# direct\n"; + $T_HDR_literal = "\.section\t\.rodata\n"; + $T_HDR_misc = "\.text\n\t\.align 2\n"; + $T_HDR_data = "\.data\n\t\.align 2\n"; + $T_HDR_consist = "\.text\n"; + $T_HDR_closure = "\.data\n\t\.align 2\n"; + $T_HDR_info = "\.text\n\t\.align 2\n"; + $T_HDR_entry = "\.text\n"; + $T_HDR_fast = "\.text\n\t\.align 2\n"; + $T_HDR_vector = "\.text\n\t\.align 2\n"; + $T_HDR_direct = "\.text\n\t\.align 2\n"; #--------------------------------------------------------# } elsif ( $TargetPlatform =~ /^sparc-.*-solaris2/ ) { $T_STABBY = 0; # 1 iff .stab things (usually if a.out format) $T_US = ''; # _ if symbols have an underscore on the front + $T_DO_GC = 'PerformGC_wrapper'; $T_PRE_APP = 'DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP $T_CONST_LBL = '^\.LLC(\d+):$'; # regexp for what such a lbl looks like $T_POST_LBL = ':'; @@ -265,14 +258,13 @@ sub init_TARGET_STUFF { $T_COPY_DIRVS = '\.(global|proc|stab)'; $T_hsc_cc_PAT = '\.asciz.*\)(hsc|cc) (.*)\\\\t(.*)"'; - $T_DOT_WORD = '\.(word|byte|half|skip|uahalf|uaword)'; + $T_DOT_WORD = '\.word'; $T_DOT_GLOBAL = '^\t\.global'; $T_HDR_literal = "\.text\n\t\.align 8\n"; $T_HDR_misc = "\.text\n\t\.align 4\n"; $T_HDR_data = "\.data\n\t\.align 8\n"; $T_HDR_consist = "\.text\n"; $T_HDR_closure = "\.data\n\t\.align 4\n"; - $T_HDR_srt = "\.data\n\t\.align 4\n"; $T_HDR_info = "\.text\n\t\.align 4\n"; $T_HDR_entry = "\.text\n\t\.align 4\n"; $T_HDR_fast = "\.text\n\t\.align 4\n"; @@ -284,6 +276,7 @@ sub init_TARGET_STUFF { $T_STABBY = 1; # 1 iff .stab things (usually if a.out format) $T_US = '_'; # _ if symbols have an underscore on the front + $T_DO_GC = '_PerformGC_wrapper'; $T_PRE_APP = '^# DOES NOT SEEM TO APPLY'; # regexp that says what comes before APP/NO_APP $T_CONST_LBL = '^LC(\d+):$'; $T_POST_LBL = ':'; @@ -299,7 +292,6 @@ sub init_TARGET_STUFF { $T_HDR_data = "\.data\n\t\.align 8\n"; $T_HDR_consist = "\.text\n"; $T_HDR_closure = "\.data\n\t\.align 4\n"; - $T_HDR_srt = "\.data\n\t\.align 4\n"; $T_HDR_info = "\.text\n\t\.align 4\n"; $T_HDR_entry = "\.text\n\t\.align 4\n"; $T_HDR_fast = "\.text\n\t\.align 4\n"; @@ -315,6 +307,7 @@ sub init_TARGET_STUFF { if ( 0 ) { print STDERR "T_STABBY: $T_STABBY\n"; print STDERR "T_US: $T_US\n"; +print STDERR "T_DO_GC: $T_DO_GC\n"; print STDERR "T_PRE_APP: $T_PRE_APP\n"; print STDERR "T_CONST_LBL: $T_CONST_LBL\n"; print STDERR "T_POST_LBL: $T_POST_LBL\n"; @@ -384,9 +377,9 @@ sub mangle_asm { %slowchk = (); # ditto, its regular "slow" entry code %fastchk = (); # ditto, fast entry code %closurechk = (); # ditto, the (static) closure - %srtchk = (); # ditto, its SRT (for top-level things) %infochk = (); # given a symbol base, say what chunk its info tbl is in %vectorchk = (); # ditto, return vector table + %directchk = (); # ditto, direct return code $EXTERN_DECLS = ''; # .globl .text (MIPS only) $i = 0; $chkcat[0] = 'misc'; $chk[0] = ''; @@ -394,27 +387,17 @@ sub mangle_asm { while () { next if $T_STABBY && /^\.stab.*$TUS[@]?__stg_split_marker/o; next if $T_STABBY && /^\.stab.*ghc.*c_ID/; - next if /^\t\.def.*endef$/; next if /$TPREAPP(NO_)?APP/o; next if /^;/ && $TargetPlatform =~ /^hppa/; - next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|powerpc|rs6000)-/; + next if /(^$|^\t\.file\t|^ # )/ && $TargetPlatform =~ /^(mips|powerpc)-/; - last if /^_section_\.text:$/ && $TargetPlatform =~ /^powerpc-|^rs6000-/; + last if /^_section_\.text:$/ && $TargetPlatform =~ /^powerpc-/; if ( $TargetPlatform =~ /^mips-/ && /^\t\.(globl \S+ \.text|comm\t)/ ) { $EXTERN_DECLS .= $_ unless /(__DISCARD__|\b(PK_|ASSIGN_)(FLT|DBL)\b)/; - # As a temporary solution for compiling "foreign export" declarations, - # we use global variables to pass arguments from C to STG land. - # These declarations live in the .hc file and not in the generated C - # stub file, so we let them pass through here. - } elsif ( /^\t\.comm\t__fexp_.*$/ ) { - $chk[++$i] = $_; - $chkcat[$i] = 'data'; - $chksymb[$i] = ''; - } elsif ( /^\s+/ ) { # most common case first -- a simple line! # duplicated from the bottom @@ -450,14 +433,14 @@ sub mangle_asm { $infochk{$symb} = $i; - } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)_(entry|ret)$TPOSTLBL[@]?$/o ) { + } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)_entry$TPOSTLBL[@]?$/o ) { $chk[++$i] = $_; $chkcat[$i] = 'slow'; $chksymb[$i] = $1; $slowchk{$1} = $i; - } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)_fast\d*$TPOSTLBL[@]?$/o ) { + } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)_fast\d+$TPOSTLBL[@]?$/o ) { $chk[++$i] = $_; $chkcat[$i] = 'fast'; $chksymb[$i] = $1; @@ -471,18 +454,6 @@ sub mangle_asm { $closurechk{$1} = $i; - } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)_srt$TPOSTLBL[@]?$/o ) { - $chk[++$i] = $_; - $chkcat[$i] = 'srt'; - $chksymb[$i] = $1; - - $srtchk{$1} = $i; - - } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)_ct$TPOSTLBL[@]?$/o ) { - $chk[++$i] = $_; - $chkcat[$i] = 'data'; - $chksymb[$i] = ''; - } elsif ( /^$TUS[@]?ghc.*c_ID$TPOSTLBL/o ) { $chk[++$i] = $_; $chkcat[$i] = 'consist'; @@ -490,8 +461,10 @@ sub mangle_asm { } elsif ( /^($TUS[@]?__gnu_compiled_c|gcc2_compiled\.)$TPOSTLBL/o ) { ; # toss it - } elsif ( /^$TUS[A-Za-z0-9_]+\.\d+$TPOSTLBL[@]?$/o + } elsif ( /^$TUS[@]?ErrorIO_call_count$TPOSTLBL[@]?$/o # HACK!!!! + || /^$TUS[A-Za-z0-9_]+\.\d+$TPOSTLBL[@]?$/o || /^$TUS[@]?.*_CAT$TPOSTLBL[@]?$/o # PROF: _entryname_CAT + || /^$TUS[@]?CC_.*_struct$TPOSTLBL[@]?$/o # PROF: _CC_ccident_struct || /^$TUS[@]?.*_done$TPOSTLBL[@]?$/o # PROF: _module_done || /^$TUS[@]?_module_registered$TPOSTLBL[@]?$/o # PROF: _module_registered ) { @@ -502,40 +475,30 @@ sub mangle_asm { } elsif ( /^([A-Za-z0-9_]+)\s+\.comm/ && $TargetPlatform =~ /^hppa/ ) { $chk[++$i] = $_; $chkcat[$i] = 'bss'; - $chksymb[$i] = ''; - - } elsif ( $TargetPlatform =~ /^powerpc-|^rs6000-/ && /^LC\.\.([0-9]+)/ ) { - $chk[++$i] = $_; - $chkcat[$i] = 'toc'; - $chksymb[$i] = $1; - - } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)_cc(s)?$TPOSTLBL[@]?$/o ) { - # all CC_ symbols go in the data section... - $chk[++$i] = $_; - $chkcat[$i] = 'data'; - $chksymb[$i] = ''; + $chksymb[$i] = $1; - } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)_(alt|dflt)$TPOSTLBL[@]?$/o ) { + } elsif ( /^$TUS[@]?(ret_|djn_)/o ) { $chk[++$i] = $_; $chkcat[$i] = 'misc'; $chksymb[$i] = ''; - #$symbtmp = $1; - #$chksymb[$i] = $symbtmp if ($TargetPlatform =~ /^powerpc-|^rs6000-/) ; #rm andre - } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)_vtbl$TPOSTLBL[@]?$/o ) { + } elsif ( /^$TUS[@]?vtbl_([A-Za-z0-9_]+)$TPOSTLBL[@]?$/o ) { $chk[++$i] = $_; $chkcat[$i] = 'vector'; $chksymb[$i] = $1; $vectorchk{$1} = $i; - # As a temporary solution for compiling "foreign export" declarations, - # we use global variables to pass arguments from C to STG land. - # These declarations live in the .hc file and not in the generated C - # stub file, so we let them pass through here. - } elsif ( /^[\t ]+\.comm[\t ]+__fexp_.*$/ ) { + } elsif ( /^$TUS[@]?([A-Za-z0-9_]+)DirectReturn$TPOSTLBL[@]?$/o ) { $chk[++$i] = $_; - $chkcat[$i] = 'data'; + $chkcat[$i] = 'direct'; + $chksymb[$i] = $1; + + $directchk{$1} = $i; + + } elsif ( /^$TUS[@]?[A-Za-z0-9_]+_upd$TPOSTLBL[@]?$/o ) { + $chk[++$i] = $_; + $chkcat[$i] = 'misc'; $chksymb[$i] = ''; } elsif ( $TargetPlatform =~ /^i386-.*-solaris2/ @@ -556,22 +519,18 @@ sub mangle_asm { } elsif ( /^$TUS[@]?[A-Za-z0-9_]/o && ( $TargetPlatform !~ /^hppa/ # need to avoid local labels in this case || ! /^L\$\d+$/ ) - && ( $TargetPlatform !~ /^powerpc|^rs6000/ # ditto + && ( $TargetPlatform !~ /^powerpc/ # ditto || ! /^(L\.\.\d+|LT\.\..*):$/ ) ) { local($thing); chop($thing = $_); - print "Funny global thing?: $_" + print STDERR "Funny global thing?: $_" unless $KNOWN_FUNNY_THING{$thing} - || /^$TUS[@]?stg_.*$TPOSTLBL[@]?$/o # RTS internals - || /^$TUS[@]__fexp_.*$TPOSTLBL$/o # foreign export - || /^$TUS[@]?_reg.*$TPOSTLBL$/o # PROF: __reg - || /^$TUS[@]?.*_btm$TPOSTLBL$/o # large bitmaps - || /^$TUS[@]?.*_closure_tbl$TPOSTLBL$/o; # closure tables + || /^$TUS[@]?_(PRIn|PRStart).*$TPOSTLBL[@]?$/o # pointer reversal GC routines + || /^$TUS[@]?CC_.*$TPOSTLBL$/o # PROF: _CC_ccident ([@]? is a silly hack (see above)) + || /^$TUS[@]?_reg.*$TPOSTLBL$/o; # PROF: __reg $chk[++$i] = $_; $chkcat[$i] = 'misc'; - if ($TargetPlatform =~ /^powerpc-|^rs6000-/) - { $chksymb[$i] = $thing; } - else { $chksymb[$i] = ''; }; + $chksymb[$i] = ''; } else { # simple line (duplicated at the top) @@ -600,37 +559,6 @@ sub mangle_asm { # HPPAs, MIPSen: also start medding at chunk 1 - if ($TargetPlatform =~ /^powerpc|^rs6000/) { - print OUTASM $T_HDR_toc; # yes, we have to put a .toc - # in the beginning of every file! - %tocequiv = (); # maps toc symbol number to toc symbol - %revtocequiv = (); # maps toc symbol to toc symbol number - for ($i = 1; $i < $numchks; $i++) { - $chk[$i] =~ s/\[RW\]//g; - $chk[$i] =~ s/\[DS\]//g; - $chk[$i] =~ s/^\.csect .*\[DS\]$//g; - - if ( $chkcat[$i] eq 'toc' && $chk[$i] !~ /\.byte/ ) -#ToDo: instead of all these changes, just fix mangle_powerpc_tailjump and delete/ignore these tocs? - { $chk[$i] =~ s/$T_MOVE_DIRVS//g; - $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(\S+_fast\d+)/\t\.tc \1\[TC\],\.\2/; - $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(\S+_entry)\n/\t\.tc \1\[TC\],\.\2\n/; - $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(ret_\S+)/\t\.tc \1\[TC\],\.\2/; - $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(alt_\S+)/\t\.tc \1\[TC\],\.\2/; - $chk[$i] =~ s/\t\.tc (\S+)\[TC\],(vtbl_\S+)/\t\.tc \1\[TC\],\.\2/; - - $tocnumber = $chksymb[$i]; - $tocsymb = $chk[$i]; - $tocsymb =~ s/^LC\.\.\d+:\n//; - $tocsymb =~ s/^\t\.tc \S+,(\S+)\n/\1/; - $tocequiv{$tocnumber} = $tocsymb; - - } elsif ( $chkcat[$i] eq 'toc' && $chk[$i] =~ /\.byte/ ) { - $chkcat[$i] = 'literal'; - } - } - }; - for ($i = $FIRST_MANGLABLE; $i < $numchks; $i++) { $c = $chk[$i]; # convenience copy @@ -648,16 +576,10 @@ sub mangle_asm { if ($TargetPlatform =~ /^i386-/) { $p =~ s/^\tpushl \%edi\n//; $p =~ s/^\tpushl \%esi\n//; - $p =~ s/^\tpushl \%ebx\n//; $p =~ s/^\tsubl \$\d+,\%esp\n//; $p =~ s/^\tmovl \$\d+,\%eax\n\tcall __alloca\n// if ($TargetPlatform =~ /^.*-cygwin32/); } elsif ($TargetPlatform =~ /^m68k-/) { $p =~ s/^\tlink a6,#-?\d.*\n//; - $p =~ s/^\tpea a6@\n\tmovel sp,a6\n//; - # The above showed up in the asm code, - # so I added it here. - # I hope it's correct. - # CaS $p =~ s/^\tmovel d2,sp\@-\n//; $p =~ s/^\tmovel d5,sp\@-\n//; # SMmark.* only? $p =~ s/^\tmoveml \#0x[0-9a-f]+,sp\@-\n//; # SMmark.* only? @@ -674,22 +596,10 @@ sub mangle_asm { $p =~ s/^\tsw\t\$fp,\d+\(\$sp\)\n//; $p =~ s/^\tsw\t\$28,\d+\(\$sp\)\n//; $p =~ s/__FRAME__/$FRAME/; - } elsif ($TargetPlatform =~ /^powerpc-|^rs6000/) { - $p =~ s/^\tmflr 0\n//; - $p =~ s/^\tstm \d+,-\d+\(1\)\n//; - $p =~ s/^\tstw? 0,\d+\(1\)\n//g; - $p =~ s/^\tstw? 1,\d+\(1\)\n//g; #mc - $p =~ s/^\tlw?z 0,0\(1\)\n//g; #mc - $p =~ s/^\tstw?u 1,-\d+\(1\)\n//; - $p =~ s/^\tstw? \d+,-\d+\(1\)\n//g; - $p =~ s/^\tstfd \d+,-\d+\(1\)\n//g; } else { print STDERR "$Pgm: unknown prologue mangling? $TargetPlatform\n"; } - # HWL HACK: dont die, just print a warning - #print stderr "HWL: this should die! Prologue junk?: $p\n" if $p =~ /^\t[^\.]/ - # && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test die "Prologue junk?: $p\n" if $p =~ /^\t[^\.]/ && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test @@ -711,7 +621,6 @@ sub mangle_asm { $e =~ s/^\tret\n//; $e =~ s/^\tpopl \%edi\n//; $e =~ s/^\tpopl \%esi\n//; - $e =~ s/^\tpopl \%ecx\n//; $e =~ s/^\taddl \$\d+,\%esp\n//; } elsif ($TargetPlatform =~ /^m68k-/) { $e =~ s/^\tunlk a6\n//; @@ -721,18 +630,9 @@ sub mangle_asm { $e =~ s/^\tlw\t\$fp,\d+\(\$sp\)\n//; $e =~ s/^\taddu\t\$sp,\$sp,\d+\n//; $e =~ s/^\tj\t\$31\n//; - } elsif ($TargetPlatform =~ /^powerpc-|^rs6000-/) { - $e =~ s/^\taddi 1,1,\d+\n//; - $e =~ s/^\tcal 1,\d+\(1\)\n//; - $e =~ s/^\tlw?z? \d+,\d+\(1\)\n//; - $e =~ s/^\tmtlr 0\n//; - $e =~ s/^\tbl?r\n//; } else { print STDERR "$Pgm: unknown epilogue mangling? $TargetPlatform\n"; } - # HWL HACK: dont die, just print a warning - #print stderr "HWL: this should die! Epilogue junk?: $e\n" if $e =~ /^\t[^\.]/ - # && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test die "Epilogue junk?: $e\n" if $e =~ /^\t[^\.]/ && $TargetPlatform !~ /^powerpc-/; #ToDo: remove test @@ -746,7 +646,7 @@ sub mangle_asm { # toss the register-windowing save/restore/ret* instructions # directly: if ( $TargetPlatform =~ /^sparc-/ ) { - $c =~ s/^\t(save.*|restore|ret|retl)\n//g; + $c =~ s/^\t(save .*|restore|ret|retl)\n//g; # throw away PROLOGUE comments $c =~ s/^\t!#PROLOGUE# 0\n\t!#PROLOGUE# 1\n//; } @@ -872,26 +772,7 @@ sub mangle_asm { if ( $chkcat[$i] eq 'misc' ) { if ($chk[$i] ne '') { print OUTASM $T_HDR_misc; - if ($TargetPlatform =~ /^powerpc-|^rs6000/) { - $chksymb[$i] =~ s/://; -#new if ($chksymb[$i] =~ /ret.*upd/ || $KNOWN_FUNNY_THING{$chksymb[$i]} -#new || $chksymb[$i] =~ /^$.{T_US}_(PRIn|PRStart).*${T_POST_LBL}$/o ) -#new { print OUTASM "\t\.globl $chksymb[$i]\n"; } -# if ($chksymb[$i] ne '' && $chksymb[$i] !~ /ret_[a-z]/ && $chksymb[$i] !~ /djn_[a-z]/) - if ($chksymb[$i] ne '') - { print OUTASM "\t\.globl \.$chksymb[$i]\n"; }; - if ($chk[$i] =~ /TOC\[tc0\], 0\n/) - { ($p, $r) = split(/TOC\[tc0\], 0\n/, $chk[$i]); $printDS = 1;} - else { $r = $chk[$i]; $printDS = 0; }; - $chk[$i] = &mangle_powerpc_tailjump($r); - }; - &print_doctored($chk[$i], 0); - if ($TargetPlatform =~ /^powerpc-|^rs6000-/ && $printDS) { -#ok if ($chksymb[$i] !~ /\_regMain/) { - print OUTASM "\.csect ${chksymb[$i]}[DS]\n"; - print OUTASM "${p}TOC[tc0], 0\n"; -#ok } - } + &print_doctored($chk[$i], 0); } } elsif ( $chkcat[$i] eq 'toss' ) { @@ -910,28 +791,9 @@ sub mangle_asm { $consist =~ s/\//./g; $consist =~ s/-/_/g; $consist =~ s/[^A-Za-z0-9_.]/ZZ/g; # ToDo: properly? - # - # Using a cygnus-2.7-96q4 gcc build on hppas, the - # consistency chunk for ghc_cc_ID often (but not always!) - # gets lumped with a bunch of .IMPORT directives containing info on - # the code or data space nature of external symbols. We can't - # toss these, so once the consistency ID has been turned into - # a representable symbol, we substitute it for the symbol - # that the string was attached to in the first place (ghc_cc_ID.) - # (The original string is also substituted away.) - # - # This change may affect the code output on other platforms in - # adverse ways, hence we restrict this hack hppa targets only. - # - # -- 2/98 SOF - if ( $TargetPlatform =~ /^hppa/ ) { - $chk[$i] =~ s/^$TUS[@]?ghc.*c_ID$TPOSTLBL/$consist/o; - $chk[$i] =~ s/\t$T_hsc_cc_PAT/$T_HDR_misc/o; - $consist = $chk[$i]; #clumsily - } print OUTASM $T_HDR_consist, "${consist}${T_POST_LBL}\n"; - } elsif ( $TargetPlatform !~ /^(mips)-/ ) { # we just don't try in those case (ToDo) + } elsif ( $TargetPlatform !~ /^(mips|powerpc)-/ ) { # we just don't try in those case (ToDo) # on mips: consistency string is just a v # horrible bunch of .bytes, # which I am too lazy to sort out (WDP 95/05) @@ -946,7 +808,6 @@ sub mangle_asm { print OUTASM "___stg_split_marker",$chksymb[$i],"${T_POST_LBL}\n"; } elsif ( $chkcat[$i] eq 'closure' - || $chkcat[$i] eq 'srt' || $chkcat[$i] eq 'infotbl' || $chkcat[$i] eq 'slow' || $chkcat[$i] eq 'fast' ) { # do them in that order @@ -959,31 +820,19 @@ sub mangle_asm { $chkcat[$closurechk{$symb}] = 'DONE ALREADY'; } - # SRT - if ( defined($srtchk{$symb}) ) { - print OUTASM $T_HDR_srt; - print OUTASM $chk[$srtchk{$symb}]; - $chkcat[$srtchk{$symb}] = 'DONE ALREADY'; - } - # INFO TABLE if ( defined($infochk{$symb}) ) { print OUTASM $T_HDR_info; - if ($TargetPlatform =~ /^powerpc-|^rs6000-/) { - if ( !defined($slowchk{$symb}) && defined($fastchk{$symb}) ) { - $fastname = $chk[$fastchk{$symb}]; - $fastname =~ s/([_A-Za-z]*_fast\d+):.*(.*\n)*/\1/; - $chk[$infochk{$symb}] =~ s/\.long StdErrorCode/\.long $fastname/; - } - $chk[$infochk{$symb}] =~ s/\.long ([_A-Za-z]\S+_entry)/\.long \.\1/; - $chk[$infochk{$symb}] =~ s/\.long ([A-Za-z]\S+_upd)/\.long \.\1/; - print OUTASM $chk[$infochk{$symb}]; - } else { - print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1); - } + print OUTASM &rev_tbl($symb, $chk[$infochk{$symb}], 1); # entry code will be put here! + # paranoia + if ( $chk[$infochk{$symb}] =~ /$TDOTWORD[@]?\s+([A-Za-z0-9_]+_entry)$/o + && $1 ne "${T_US}${symb}_entry" ) { + print STDERR "!!! entry point???\n",$chk[$infochk{$symb}]; + } + $chkcat[$infochk{$symb}] = 'DONE ALREADY'; } @@ -993,42 +842,22 @@ sub mangle_asm { # teach it to drop through to the fast entry point: $c = $chk[$slowchk{$symb}]; - if ($TargetPlatform =~ /^powerpc-|^rs6000-/) { - ($p, $r) = split(/TOC\[tc0\], 0\n/, $c); - if ($symb =~ /^[_A-Z]/) - { - print OUTASM "\t\.globl \.${chksymb[$i]}_entry\n"; - print OUTASM "\.csect ${symb}_entry[DS]\n"; - print OUTASM "${p}TOC[tc0], 0\n"; - }; - $r =~ s/\.csect \.text\[PR\]\n//; # todo: properly - andre - $c = &mangle_powerpc_tailjump($r); - }; - if ( defined($fastchk{$symb}) ) { if ( $TargetPlatform =~ /^alpha-/ ) { $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/; } elsif ( $TargetPlatform =~ /^hppa/ ) { $c =~ s/^\s+ldil.*\n\s+ldo.*\n\s+bv.*\n(.*\n)?\s+\.EXIT/$1\t.EXIT/; } elsif ( $TargetPlatform =~ /^i386-/ ) { - # Reg alloc depending, gcc generated code may jump to the fast entry point via - # a number of registers. - $c =~ s/^\tmovl \$${T_US}${symb}_fast\d*,\%edx\n\tjmp \*\%edx\n//; - $c =~ s/^\tmovl \$${T_US}${symb}_fast\d*,\%ecx\n\tjmp \*\%ecx\n//; - $c =~ s/^\tmovl \$${T_US}${symb}_fast\d*,\%eax\n\tjmp \*\%eax\n//; - # The next two only apply if we're not stealing %esi or %edi. - $c =~ s/^\tmovl \$${T_US}${symb}_fast\d*,\%esi\n\tjmp \*\%esi\n// if ($StolenX86Regs < 3); - $c =~ s/^\tmovl \$${T_US}${symb}_fast\d*,\%edi\n\tjmp \*\%edi\n// if ($StolenX86Regs < 4); + $c =~ s/^\tmovl \$${T_US}${symb}_fast\d+,\%edx\n\tjmp \*\%edx\n//; + $c =~ s/^\tmovl \$${T_US}${symb}_fast\d+,\%eax\n\tjmp \*\%eax\n//; } elsif ( $TargetPlatform =~ /^mips-/ ) { $c =~ s/^\tjmp \$31,\(\$27\),0\n\t\.align 4\n\t\.end/\t.align 4\n\t.end/; } elsif ( $TargetPlatform =~ /^m68k-/ ) { $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n\tnop\n//; $c =~ s/^\tjmp ${T_US}${symb}_fast\d+.*\n//; - } elsif ( $TargetPlatform =~ /^powerpc-|^rs6000-/ ) { - $c =~ s/^\tb \.${T_US}${symb}_fast\d+\n//; } elsif ( $TargetPlatform =~ /^sparc-/ ) { - $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n\t\s*nop\n//; - $c =~ s/^\tcall\s+${T_US}${symb}_fast\d+,.*\n(\t\s*[a-z].*\n)/$1/; + $c =~ s/^\tcall ${T_US}${symb}_fast\d+,.*\n\tnop\n//; + $c =~ s/^\tcall ${T_US}${symb}_fast\d+,.*\n(\t[a-z].*\n)/$1/; } else { print STDERR "$Pgm: mystery slow-fast dropthrough: $TargetPlatform\n"; } @@ -1051,7 +880,6 @@ sub mangle_asm { # FAST ENTRY POINT if ( defined($fastchk{$symb}) ) { - $c = $chk[$fastchk{$symb}]; if ( ! defined($slowchk{$symb}) # ToDo: the || clause can go once we're no longer # concerned about producing exactly the same output as before @@ -1059,19 +887,7 @@ sub mangle_asm { ) { print OUTASM $T_HDR_fast; } - if ($TargetPlatform =~ /^powerpc-|^rs6000-/) { - local(@lbls) = split(/:/, $c); - $fullname = $lbls[0]; - $fullname =~ s/$T_MOVE_DIRVS//g; - if ( $fullname =~ /^[A-Z]/) - { print OUTASM "\t\.globl \.${fullname}\n"; - } else { -# print OUTASM "\t\.lglobl \.${fullname}\n"; #todo: rm - andre - }; - $c =~ s/((.*\n)*)\t.long \S+, TOC\[tc0\], 0\n\.csect \.text\[PR\]\n((.*\n)*)/\1\3/; - $c = &mangle_powerpc_tailjump($c); - }; - &print_doctored($c, 0); + &print_doctored($chk[$fastchk{$symb}], 0); $chkcat[$fastchk{$symb}] = 'DONE ALREADY'; } @@ -1082,19 +898,16 @@ sub mangle_asm { # VECTOR TABLE if ( defined($vectorchk{$symb}) ) { print OUTASM $T_HDR_vector; - if ($TargetPlatform =~ /^powerpc-|^rs6000-/) { - if ( $symb =~ /^[A-Z]/) { - print OUTASM "\t\.globl \.${symb}_vtbl\n"; - print OUTASM "\t\.globl ${symb}_vtbl\n"; - }; - $chk[$vectorchk{$symb}] =~ s/\.long (\S+)/\.long \.\1/g; - print OUTASM ".${symb}_vtbl:\n"; - print OUTASM $chk[$vectorchk{$symb}]; - } else { - print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0); - } + print OUTASM &rev_tbl($symb, $chk[$vectorchk{$symb}], 0); # direct return code will be put here! $chkcat[$vectorchk{$symb}] = 'DONE ALREADY'; + } + + # DIRECT RETURN + if ( defined($directchk{$symb}) ) { + print OUTASM $T_HDR_direct; + &print_doctored($chk[$directchk{$symb}], 0); + $chkcat[$directchk{$symb}] = 'DONE ALREADY'; } elsif ( $TargetPlatform =~ /^alpha-/ ) { # Alphas: the commented nop is for the splitter, to ensure @@ -1106,20 +919,6 @@ sub mangle_asm { print OUTASM "\t# nop\n"; } - } elsif ( $chkcat[$i] eq 'toc' ) { - # silly optimisation to print tocs, since they come in groups... - print OUTASM $T_HDR_toc; - local($j) = $i; - while ($chkcat[$j] eq 'toc') - { if ( $chk[$j] !~ /\.tc UpdatePAP\[TC\]/ # not needed: always turned into a jump. - ) - { - print OUTASM $chk[$j]; - } - $chkcat[$j] = 'DONE ALREADY'; - $j++; - } - } else { &tidy_up_and_die(1,"$Pgm: unknown chkcat (ghc-asm: $TargetPlatform)\n$chkcat[$i]\n$chk[$i]\n"); } @@ -1127,9 +926,8 @@ sub mangle_asm { print OUTASM $EXTERN_DECLS if $TargetPlatform =~ /^mips-/; - if ($TargetPlatform =~ /^powerpc-|^rs6000-/) { - print OUTASM ".csect .text[PR]\n_section_.text:\n.csect .data[RW]\n\t.long _section_.text\n" - }; + print OUTASM ".csect .text[PR]\n_section_.text:\n.csect .data[RW]\n\t.long _section_.text\n" + if $TargetPlatform =~ /^powerpc-/; # finished close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n"); @@ -1167,8 +965,7 @@ sub print_doctored { local($_, $need_fallthru_patch) = @_; if ( $TargetPlatform !~ /^i386-/ - || ! /^\t[a-z]/ # no instructions in here, apparently - || /^${T_US}_reg[A-Za-z0-9_]+${T_POST_LBL}/) { + || ! /^\t[a-z]/ ) { # no instructions in here, apparently print OUTASM $_; return; } @@ -1176,6 +973,34 @@ sub print_doctored { local($entry_patch) = ''; local($exit_patch) = ''; + local($call_entry_patch)= ''; + local($call_exit_patch) = ''; + +#OLD: # first, convert calls to *very magic form*: (ToDo: document + # for real!) from + # + # pushl $768 + # call _?PerformGC_wrapper + # addl $4,%esp + # to + # movl $768, %eax + # call _?PerformGC_wrapper + # + # The reason we do this now is to remove the apparent use of + # %esp, which would throw off the "what patch code do we need" + # decision. + # + # Special macros in ghc/includes/COptWraps.lh, used in + # ghc/runtime/CallWrap_C.lc, are required for this to work! + # + + s/^\tpushl \$(\d+)\n\tcall ${T_DO_GC}\n\taddl \$4,\%esp\n/\tmovl \$$1,\%eax\n\tcall ${T_DO_GC}\n/go; + s/^\tpushl \%eax\n\tcall ${T_DO_GC}\n\taddl \$4,\%esp\n/\tcall ${T_DO_GC}\n/go; + s/^\tpushl \%edx\n\tcall ${T_DO_GC}\n\taddl \$4,\%esp\n/\tmovl \%edx,\%eax\n\tcall ${T_DO_GC}\n/go; + +#= if ( $StolenX86Regs <= 4 ) { # %ecx is ordinary reg +#= s/^\tpushl \%ecx\n\tcall ${T_DO_GC}\n\taddl \$4,\%esp\n/\tmovl \%ecx,\%eax\n\tcall ${T_DO_GC}\n/go; +#= } # gotta watch out for weird instructions that # invisibly smash various regs: @@ -1210,13 +1035,10 @@ sub print_doctored { # movl $_blah, # jmp * # - -# the short form may tickle perl bug: -# s/^\tmovl \$${T_US}(.*),(\%e[abcd]x)\n\tjmp \*$2/\tjmp $T_US$1/g; - s/^\tmovl \$${T_US}(.*),\%eax\n\tjmp \*\%eax/\tjmp $T_US$1/g; - s/^\tmovl \$${T_US}(.*),\%ebx\n\tjmp \*\%ebx/\tjmp $T_US$1/g; - s/^\tmovl \$${T_US}(.*),\%ecx\n\tjmp \*\%ecx/\tjmp $T_US$1/g; - s/^\tmovl \$${T_US}(.*),\%edx\n\tjmp \*\%edx/\tjmp $T_US$1/g; + # which is easily fixed as: + # + # sigh! try to hack around it... + # if ($StolenX86Regs <= 2 ) { # YURGH! spurious uses of esi? s/^\tmovl (.*),\%esi\n\tjmp \*%esi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g; @@ -1227,43 +1049,49 @@ sub print_doctored { } if ($StolenX86Regs <= 3 ) { # spurious uses of edi? s/^\tmovl (.*),\%edi\n\tjmp \*%edi\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g; - s/^\tjmp \*(-?\d*\(.*\%edi.*\))\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g; + s/^\tjmp \*(-?\d*)\((.*\%edi.*)\)\n/\tmovl $2,\%eax\n\tjmp \*$1\(\%eax\)\n/g; s/^\tjmp \*\%edi\n/\tmovl \%edi,\%eax\n\tjmp \*\%eax\n/g; die "$Pgm: (mangler) still have jump involving \%edi!\n$_" if /(jmp|call) .*\%edi/; } +#= if ($StolenX86Regs <= 4 ) { # spurious uses of ecx? +#= s/^\tmovl (.*),\%ecx\n\tjmp \*%ecx\n/\tmovl $1,\%eax\n\tjmp \*\%eax\n/g; +#= s/^\tjmp \*(-?\d*)\((.*\%ecx.*)\)\n/\tmovl $2,\%eax\n\tjmp \*$1\(\%eax\)\n/g; +#= s/^\tjmp \*\%ecx\n/\tmovl \%ecx,\%eax\n\tjmp \*\%eax\n/g; +#= die "$Pgm: (mangler) still have jump involving \%ecx!\n$_" +#= if /(jmp|call) .*\%ecx/; +#= } # OK, now we can decide what our patch-up code is going to # be: - - # Offsets into register table - you'd better update these magic - # numbers should you change its contents! - # local($OFFSET_R1)=0; No offset for R1 in new RTS. - local($OFFSET_Hp)=92; - - # Note funky ".=" stuff; we're *adding* to these _patch guys if ( $StolenX86Regs <= 2 - && ( /[^0-9]\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi) - $entry_patch .= "\tmovl \%esi,(\%ebx)\n"; - $exit_patch .= "\tmovl (\%ebx),\%esi\n"; - + && ( /32\(\%ebx\)/ || /\%esi/ || /^\tcmps/ ) ) { # R1 (esi) + $entry_patch .= "\tmovl \%esi,32(\%ebx)\n"; + $exit_patch .= "\tmovl 32(\%ebx),\%esi\n"; # nothing for call_{entry,exit} because %esi is callee-save } if ( $StolenX86Regs <= 3 - && ( /${OFFSET_Hp}\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # Hp (edi) - $entry_patch .= "\tmovl \%edi,${OFFSET_Hp}(\%ebx)\n"; - $exit_patch .= "\tmovl ${OFFSET_Hp}(\%ebx),\%edi\n"; - + && ( /64\(\%ebx\)/ || /\%edi/ || /^\t(scas|cmps)/ ) ) { # SpA (edi) + $entry_patch .= "\tmovl \%edi,64(\%ebx)\n"; + $exit_patch .= "\tmovl 64(\%ebx),\%edi\n"; # nothing for call_{entry,exit} because %edi is callee-save } - +#= if ( $StolenX86Regs <= 4 +#= && ( /80\(\%ebx\)/ || /\%ecx/ || /^\t(rep|loop)/ ) ) { # Hp (ecx) +#= $entry_patch .= "\tmovl \%ecx,80(\%ebx)\n"; +#= $exit_patch .= "\tmovl 80(\%ebx),\%ecx\n"; +#= +#= $call_exit_patch .= "\tmovl \%ecx,108(\%ebx)\n"; +#= $call_entry_patch .= "\tmovl 108(\%ebx),\%ecx\n"; +#= # I have a really bad feeling about this if we ever +#= # have a nested call... +#= # NB: should just hide it somewhere in the C stack. +#= } # -------------------------------------------------------- # next, here we go with non-%esp patching! # s/^(\t[a-z])/$entry_patch$1/; # before first instruction - -# Before calling GC we must set up the exit condition before the call -# and entry condition when we come back + s/^(\tcall .*\n(\taddl \$\d+,\%esp\n)?)/$call_exit_patch$1$call_entry_patch/g; # _all_ calls # fix _all_ non-local jumps: @@ -1275,14 +1103,47 @@ sub print_doctored { s/^\tJMP___SL/\tjmp \*${T_X86_PRE_LLBL}/go; s/^\tJMP___L/\tjmp ${T_X86_PRE_LLBL}/go; + # fix post-PerformGC wrapper (re-)entries ??? + if ($StolenX86Regs == 2 ) { die "ARGH! Jump uses \%esi or \%edi with -monly-2-regs:\n$_" if /^\t(jmp|call) .*\%e(si|di)/; +#= die "ARGH! Jump uses \%esi, \%edi, or \%ecx with -monly-2-regs:\n$_" +#= if /^\t(jmp|call) .*\%e(si|di|cx)/; } elsif ($StolenX86Regs == 3 ) { die "ARGH! Jump uses \%edi with -monly-3-regs:\n$_" if /^\t(jmp|call) .*\%edi/; +#= die "ARGH! Jump uses \%edi or \%ecx with -monly-3-regs:\n$_" +#= if /^\t(jmp|call) .*\%e(di|cx)/; +#= } elsif ($StolenX86Regs == 4 ) { +#= die "ARGH! Jump uses \%ecx with -monly-4-regs:\n$_" +#= if /^\t(jmp|call) .*\%ecx/; } + # final peephole fixes + + s/^\tmovl \%eax,36\(\%ebx\)\n\tjmp \*36\(\%ebx\)\n/\tmovl \%eax,36\(\%ebx\)\n\tjmp \*\%eax\n/; +# the short form may tickle perl bug: +# s/^\tmovl \$${T_US}(.*),(\%e[abcd]x)\n\tjmp \*$2/\tjmp $T_US$1/g; + s/^\tmovl \$${T_US}(.*),\%eax\n\tjmp \*\%eax/\tjmp $T_US$1/g; + s/^\tmovl \$${T_US}(.*),\%ebx\n\tjmp \*\%ebx/\tjmp $T_US$1/g; + s/^\tmovl \$${T_US}(.*),\%ecx\n\tjmp \*\%ecx/\tjmp $T_US$1/g; + s/^\tmovl \$${T_US}(.*),\%edx\n\tjmp \*\%edx/\tjmp $T_US$1/g; + + # Hacks to eliminate some reloads of Hp. Worth about 5% code size. + # We could do much better than this, but at least it catches about + # half of the unnecessary reloads. + # Note that these will stop working if either: + # (i) the offset of Hp from BaseReg changes from 80, or + # (ii) the register assignment of BaseReg changes from %ebx + + s/^\tmovl 80\(\%ebx\),\%e.x\n\tmovl \$(.*),(-?[0-9]*)\(\%e.x\)\n\tmovl 80\(\%ebx\),\%e(.)x/\tmovl 80\(\%ebx\),\%e$3x\n\tmovl \$$1,$2\(\%e$3x\)/g; + + s/^\tmovl 80\(\%ebx\),\%e(.)x\n\tmovl (.*),\%e(.)x\n\tmovl \%e$3x,(-?[0-9]*\(\%e$1x\))\n\tmovl 80\(\%ebx\),\%e$1x/\tmovl 80\(\%ebx\),\%e$1x\n\tmovl $2,\%e$3x\n\tmovl \%e$3x,$4/g; + + s/^\tmovl 80\(\%ebx\),\%edx((\n\t(movl|addl) .*,((-?[0-9]*\(.*)|(\%e[abc]x)))+)\n\tmovl 80\(\%ebx\),\%edx/\tmovl 80\(\%ebx\),\%edx$1/g; + s/^\tmovl 80\(\%ebx\),\%eax((\n\t(movl|addl) .*,((-?[0-9]*\(.*)|(\%e[bcd]x)))+)\n\tmovl 80\(\%ebx\),\%eax/\tmovl 80\(\%ebx\),\%eax$1/g; + # -------------------------------------------------------- # that's it -- print it # @@ -1300,8 +1161,54 @@ sub print_doctored { \begin{code} sub init_FUNNY_THINGS { %KNOWN_FUNNY_THING = ( - # example - # "${T_US}stg_.*{T_POST_LBL}", 1, + "${T_US}CheckHeapCode${T_POST_LBL}", 1, + "${T_US}CommonUnderflow${T_POST_LBL}", 1, + "${T_US}Continue${T_POST_LBL}", 1, + "${T_US}EnterNodeCode${T_POST_LBL}", 1, + "${T_US}ErrorIO_call_count${T_POST_LBL}", 1, + "${T_US}ErrorIO_innards${T_POST_LBL}", 1, + "${T_US}IndUpdRetDir${T_POST_LBL}", 1, + "${T_US}IndUpdRetV0${T_POST_LBL}", 1, + "${T_US}IndUpdRetV1${T_POST_LBL}", 1, + "${T_US}IndUpdRetV2${T_POST_LBL}", 1, + "${T_US}IndUpdRetV3${T_POST_LBL}", 1, + "${T_US}IndUpdRetV4${T_POST_LBL}", 1, + "${T_US}IndUpdRetV5${T_POST_LBL}", 1, + "${T_US}IndUpdRetV6${T_POST_LBL}", 1, + "${T_US}IndUpdRetV7${T_POST_LBL}", 1, + "${T_US}PrimUnderflow${T_POST_LBL}", 1, + "${T_US}StackUnderflowEnterNode${T_POST_LBL}", 1, + "${T_US}StdErrorCode${T_POST_LBL}", 1, + "${T_US}UnderflowVect0${T_POST_LBL}", 1, + "${T_US}UnderflowVect1${T_POST_LBL}", 1, + "${T_US}UnderflowVect2${T_POST_LBL}", 1, + "${T_US}UnderflowVect3${T_POST_LBL}", 1, + "${T_US}UnderflowVect4${T_POST_LBL}", 1, + "${T_US}UnderflowVect5${T_POST_LBL}", 1, + "${T_US}UnderflowVect6${T_POST_LBL}", 1, + "${T_US}UnderflowVect7${T_POST_LBL}", 1, + "${T_US}UpdErr${T_POST_LBL}", 1, + "${T_US}UpdatePAP${T_POST_LBL}", 1, + "${T_US}WorldStateToken${T_POST_LBL}", 1, + "${T_US}_Enter_Internal${T_POST_LBL}", 1, + "${T_US}_PRMarking_MarkNextAStack${T_POST_LBL}", 1, + "${T_US}_PRMarking_MarkNextBStack${T_POST_LBL}", 1, + "${T_US}_PRMarking_MarkNextCAF${T_POST_LBL}", 1, + "${T_US}_PRMarking_MarkNextGA${T_POST_LBL}", 1, + "${T_US}_PRMarking_MarkNextRoot${T_POST_LBL}", 1, + "${T_US}_PRMarking_MarkNextSpark${T_POST_LBL}", 1, + "${T_US}_PRMarking_MarkNextEvent${T_POST_LBL}", 1, + "${T_US}_PRMarking_MarkNextClosureInFetchBuffer${T_POST_LBL}", 1, + "${T_US}_Scavenge_Forward_Ref${T_POST_LBL}", 1, + "${T_US}__std_entry_error__${T_POST_LBL}", 1, + "${T_US}_startMarkWorld${T_POST_LBL}", 1, + "${T_US}resumeThread${T_POST_LBL}", 1, + "${T_US}startCcRegisteringWorld${T_POST_LBL}", 1, + "${T_US}startEnterFloat${T_POST_LBL}", 1, + "${T_US}startEnterInt${T_POST_LBL}", 1, + "${T_US}startPerformIO${T_POST_LBL}", 1, + "${T_US}startStgWorld${T_POST_LBL}", 1, + "${T_US}stopPerformIO${T_POST_LBL}", 1 ); } \end{code} @@ -1322,31 +1229,29 @@ sub rev_tbl { local(@words) = (); local($after) = ''; local(@lines) = split(/\n/, $tbl); - local($i, $j); #local ($i, $extra, $words_to_pad, $j); - + local($i, $extra, $words_to_pad, $j); + # see comment in mangleAsm as to why this silliness is needed. local($TDOTWORD) = ${T_DOT_WORD}; local($TDOTGLOBAL) = ${T_DOT_GLOBAL}; local($TUS) = ${T_US}; local($TPOSTLBL) = ${T_POST_LBL}; - # Deal with the header... - for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t?$TDOTWORD\s+/o; $i++) { + for ($i = 0; $i <= $#lines && $lines[$i] !~ /^\t$TDOTWORD\s+/o; $i++) { $label .= $lines[$i] . "\n", next if $lines[$i] =~ /^[A-Za-z0-9_]+_info$TPOSTLBL[@]?$/o || $lines[$i] =~ /$TDOTGLOBAL/o - || $lines[$i] =~ /^$TUS[@]?\S+_vtbl$TPOSTLBL[@]?$/o; + || $lines[$i] =~ /^$TUS[@]?vtbl_\S+$TPOSTLBL[@]?$/o; $before .= $lines[$i] . "\n"; # otherwise... } - # Grab the table data... if ( $TargetPlatform !~ /^hppa/ ) { - for ( ; $i <= $#lines && $lines[$i] =~ /^\t?$TDOTWORD\s+/o; $i++) { + for ( ; $i <= $#lines && $lines[$i] =~ /^\t$TDOTWORD\s+/o; $i++) { push(@words, $lines[$i]); } } else { # hppa weirdness - for ( ; $i <= $#lines && $lines[$i] =~ /^\s+($TDOTWORD|\.IMPORT)/; $i++) { + for ( ; $i <= $#lines && $lines[$i] =~ /^\s+\.(word|IMPORT)/; $i++) { if ($lines[$i] =~ /^\s+\.IMPORT/) { push(@imports, $lines[$i]); } else { @@ -1358,13 +1263,8 @@ sub rev_tbl { } } - # now throw away the first word (SRT) iff it is empty. - # The .zero business is for Linux/ELF. - # The .skip business is for Sparc/Solaris/ELF. - # The .blockz business is for HPPA. - if ($discard1 && $words[0] =~ /^\t?($TDOTWORD\s+0|\.zero\s+4|\.skip\s+4|\.blockz\s+4)/) { - shift(@words) - } + # now throw away the first word (entry code): + shift(@words) if $discard1; # Padding removed to reduce code size and improve performance on Pentiums. # Simon M. 13/4/96 @@ -1387,7 +1287,7 @@ sub rev_tbl { $tbl = $before . (($TargetPlatform !~ /^hppa/) ? '' : join("\n", @imports) . "\n") - . join("\n", @words) . "\n" + . join("\n", (reverse @words)) . "\n" . $label . $after; # print STDERR "before=$before\n"; @@ -1399,36 +1299,44 @@ sub rev_tbl { } \end{code} -The HP is a major nuisance. The threaded code mangler moved info -tables from data space to code space, but unthreaded code in the RTS -still has references to info tables in data space. Since the HP -linker is very precise about where symbols live, we need to patch the -references in the unthreaded RTS as well. - \begin{code} -sub mini_mangle_asm_hppa { +sub mini_mangle_asm_i386 { local($in_asmf, $out_asmf) = @_; + &init_TARGET_STUFF(); + + # see mangleAsm comment + local($TUS) = ${T_US}; + local($TPOSTLBL)=${T_POST_LBL}; + open(INASM, "< $in_asmf") || &tidy_up_and_die(1,"$Pgm: failed to open `$in_asmf' (to read)\n"); open(OUTASM,"> $out_asmf") || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n"); while () { - s/_info,DATA/_info,CODE/; # Move _info references to code space - s/P%_PR/_PR/; print OUTASM; + + next unless + /^$TUS[@]?(PerformGC|StackOverflow|Yield|PerformReschedule)_wrapper$TPOSTLBL\n/o; + print OUTASM "\tmovl \%esp, ${T_US}__temp_esp\n"; + print OUTASM "\tmovl \%eax, ${T_US}__temp_eax\n"; } # finished: close(OUTASM) || &tidy_up_and_die(1,"Failed writing to $out_asmf\n"); close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n"); } - \end{code} - + +The HP is a major nuisance. The threaded code mangler moved info +tables from data space to code space, but unthreaded code in the RTS +still has references to info tables in data space. Since the HP +linker is very precise about where symbols live, we need to patch the +references in the unthreaded RTS as well. + \begin{code} -sub mini_mangle_asm_powerpc { +sub mini_mangle_asm_hppa { local($in_asmf, $out_asmf) = @_; open(INASM, "< $in_asmf") @@ -1437,15 +1345,8 @@ sub mini_mangle_asm_powerpc { || &tidy_up_and_die(1,"$Pgm: failed to open `$out_asmf' (to write)\n"); while () { - s/long _PRStart/long ._PRStart/; - s/long _PRIn_/long ._PRIn_/; - s/long _Dummy_(\S+)_entry/long ._Dummy_\1_entry/; - s/long _PRMarking_MarkNextRoot\[DS\]/long ._PRMarking_MarkNextRoot/; - s/long _PRMarking_MarkNextCAF\[DS\]/long ._PRMarking_MarkNextCAF/; - s/long _PRMarking_MarkNextAStack\[DS\]/long ._PRMarking_MarkNextAStack/; - s/long _PRMarking_MarkNextBStack\[DS\]/long ._PRMarking_MarkNextBStack/; - s/\.tc EnterNodeCode\[TC]\,EnterNodeCode\[DS\]/\.tc EnterNodeCode\[TC]\,.EnterNodeCode/; # CONC - s/\.tc CheckHeapCode\[TC]\,CheckHeapCode\[DS\]/\.tc CheckHeapCode\[TC]\,.CheckHeapCode/; # CONC + s/_info,DATA/_info,CODE/; # Move _info references to code space + s/P%_PR/_PR/; print OUTASM; } @@ -1454,26 +1355,6 @@ sub mini_mangle_asm_powerpc { close(INASM) || &tidy_up_and_die(1,"Failed reading from $in_asmf\n"); } -sub mangle_powerpc_tailjump { - local($c) = @_; - local($maybe_more) = 1; - while (($c =~ /\tlw?z? \d+,LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n/) && $maybe_more) - { $maybe_more = 0; - $lcsymb = $c; - $lcsymb =~ s/(.*\n)*\tlw?z? \d+,LC\.\.(\d+)\(2\)\n\tmtctr \d+\n\tbctr\n(.*\n)*/\2/; -# the checks for r1 and r2 are mostly paranoia... - $r1 = $c; - $r1 =~ s/(.*\n)*\tlw?z? (\d+),LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n(.*\n)*/\2/; - $r2 = $c; - $r2 =~ s/(.*\n)*\tlw?z? \d+,LC\.\.(\d+)\(2\)\n\tmtctr (\d+)\n\tbctr\n(.*\n)*/\3/; - if (r1 == r2) - { $maybe_more = 1; - $c =~ s/((.*\n)*)\tlw?z? \d+,LC\.\.\d+\(2\)\n\tmtctr \d+\n\tbctr\n((.*\n)*)/\1\tb $tocequiv{$lcsymb}\n\3/; - } - }; - $c; -} - # make "require"r happy... 1; \end{code}