[project @ 2005-11-18 14:01:33 by simonmar]

Discard various ways in which gcc zeroes stack slots in the prologue.
So far in my investigations these have always been unnecessary, they
appear to be the result of missed optimisations by gcc, so cross
fingers and discard them.  New variants have just shown up because I
started compiling the RTS with -optc-O2.

diff --git a/ghc/driver/mangler/ghc-asm.lprl b/ghc/driver/mangler/ghc-asm.lprl
index 3406419..1c75f00 100644 (file)
--- a/ghc/driver/mangler/ghc-asm.lprl
+++ b/ghc/driver/mangler/ghc-asm.lprl
@@ -822,23 +822,27 @@ sub mangle_asm {
                        }
                    }
 
-                   $p =~ s/^\tpushl\s+\%edi\n//;
-                   $p =~ s/^\tpushl\s+\%esi\n//;
-                   $p =~ s/^\tpushl\s+\%ebx\n//;
-                   $p =~ s/^\tmovl\s+\%esi,\s*\d*\(\%esp\)\n//;
-                   $p =~ s/^\tmovl\s+\%edi,\s*\d*\(\%esp\)\n//;
+               # gcc 3.4.3 puts this kind of stuff in the prologue, eg.
+               # when compiling PrimOps.cmm with -optc-O2:
+               #        xorl    %ecx, %ecx
+               #        xorl    %edx, %edx
+               #        movl    %ecx, 16(%esp)
+               #        movl    %edx, 20(%esp)
+               # but then the code of the function doesn't assume
+               # anything about the contnets of these stack locations.
+               # I think it's to do with the use of inline functions for
+               # PK_Word64() and friends, where gcc is initialising the
+               # contents of the struct to zero, and failing to optimise
+               # away the initialisation.  Let's live dangerously and
+               # discard these initalisations.
+
+                   $p =~ s/^\tpushl\s+\%e(di|si|bx)\n//g;
+                   $p =~ s/^\txorl\s+\%e(ax|cx|dx),\s*\%e(ax|cx|dx)\n//g;
+                   $p =~ s/^\tmovl\s+\%e(ax|cx|dx|si|di),\s*\d*\(\%esp\)\n//g;
+                   $p =~ s/^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n//g;
                    $p =~ s/^\tsubl\s+\$\d+,\s*\%esp\n//;
                     $p =~ s/^\tmovl\s+\$\d+,\s*\%eax\n\tcall\s+__alloca\n// if ($TargetPlatform =~ /^.*-(cygwin32|mingw32)/);
 
-                   # GCC 3.1 is in the habit of adding spurious writes to the
-                   # stack in the prologue.  Just to be on the safe side,
-                   # chuck these over the fence into the main code.
-                   while ($p =~ /^\tmovl\s+\$\d+,\s*\d*\(\%esp\)\n/) {
-                         # print "Spurious instruction: $&";
-                         $p = $` . $';
-                         $r = $& . $r;
-                   }
-
                     if ($TargetPlatform =~ /^i386-apple-darwin/) {
                         $pcrel_label = $p;
                         $pcrel_label =~ s/(.|\n)*^(\"?L\d+\$pb\"?):\n(.|\n)*/$2/ or $pcrel_label = "";