[llvm-commits] CVS: llvm/utils/codegen-diff
Brian Gaeke
gaeke at cs.uiuc.edu
Thu Jul 3 16:16:01 PDT 2003
Changes in directory llvm/utils:
codegen-diff added (r1.1)
---
Log message:
here little scriptie, nice scriptie...
---
Diffs of the changes:
Index: llvm/utils/codegen-diff
diff -c /dev/null llvm/utils/codegen-diff:1.1
*** /dev/null Thu Jul 3 16:15:12 2003
--- llvm/utils/codegen-diff Thu Jul 3 16:15:02 2003
***************
*** 0 ****
--- 1,112 ----
+ #!/usr/bin/perl
+
+ sub parse_objdump_file {
+ my ($filename) = @_;
+ my @result;
+ open (INPUT, $filename) or die "$filename: $!\n";
+ while (<INPUT>) {
+ if (/\s*([0-9a-f]*):\t(([0-9a-f]{2} )+) *\t(.*)$/) {
+ my ($addr, $bytes, $instr) = ($1, $2, $4);
+ $addr = "0x" . $addr;
+ $bytes =~ s/\s*(.*\S)\s*/$1/; # trim any remaining whitespace
+ $instr =~ s/\s*(.*\S)\s*/$1/;
+ push (@result, {'addr' => $addr, 'bytes' => $bytes, 'instr' => $instr});
+ }
+ }
+ close INPUT;
+ return @result;
+ }
+
+ sub parse_gdb_file {
+ my ($filename) = @_;
+ my @result;
+ my $got_addr;
+ open (INPUT, $filename) or die "$filename: $!\n";
+ while (<INPUT>) {
+ if (/^(0x[0-9a-f]*):\t([^\t]*)\t[^:]*:\t((0x[0-9a-f]{2}\s*)+)\s*$/) {
+ my ($addr, $bytes, $instr) = ($1, $3, $2);
+ $bytes =~ s/0x//g;
+ $bytes =~ s/\s+/ /g; # regularize whitespace
+ $bytes =~ s/\s*(.*\S)\s*/$1/; # trim any remaining whitespace
+ $instr =~ s/\s*(.*\S)\s*/$1/;
+ push (@result, {'addr' => $addr, 'bytes' => $bytes, 'instr' => $instr});
+ } elsif (/^(0x[0-9a-f]*):\t$/) { # deal with gdb's line breaker
+ $got_addr = $1;
+ } elsif ($got_addr && /^ ([^\t]*)\t[^:]*:\t((0x[0-9a-f]{2}\s*)+)\s*$/) {
+ my ($addr, $bytes, $instr) = ($got_addr, $2, $1);
+ $bytes =~ s/0x//g;
+ $bytes =~ s/\s+/ /g; # regularize whitespace
+ $bytes =~ s/\s*(.*\S)\s*/$1/; # trim any remaining whitespace
+ $instr =~ s/\s*(.*\S)\s*/$1/;
+ push (@result, {'addr' => $addr, 'bytes' => $bytes, 'instr' => $instr});
+ undef $got_addr;
+ }
+ }
+ close INPUT;
+ return @result;
+ }
+
+ sub binary_diffs {
+ my ($objdump_file, $gdb_file) = @_;
+ my @file1 = parse_objdump_file ($objdump_file);
+ my @file2 = parse_gdb_file ($gdb_file);
+ my $lastrecord = ($#file1 >= $#file2) ? ($#file1) : ($#file2);
+ for (my $i = 0; $i <= $lastrecord; ++$i) {
+ my $d1 = $file1[$i];
+ my $d2 = $file2[$i];
+ if ($d1->{'bytes'} ne $d2->{'bytes'}) {
+ printf "0x%08x:\t%30s \t%s\n", 0+$d1->{'addr'}, $d1->{'bytes'}, $d1->{'instr'};
+ printf "0x%08x:\t%30s \t%s\n\n", 0+$d2->{'addr'}, $d2->{'bytes'}, $d2->{'instr'};
+ }
+ }
+ }
+
+ $objdump_file = $ARGV[0];
+ $gdb_file = $ARGV[1];
+ binary_diffs ($objdump_file, $gdb_file);
+ exit (0);
+ __END__
+ =pod
+
+ =head1 NAME
+
+ codegen-diff
+
+ =head1 SYNOPSIS
+
+ codegen-diff I<OBJDUMP-OUTPUT-FILE> I<GDB-DISASSEMBLY-FILE>
+
+ =head1 DESCRIPTION
+
+ B<codegen-diff> is a program that tries to show you the differences
+ between the code that B<llc> generated and the code that B<lli> generated.
+
+ The way you use it is as follows: first, you create I<OBJDUMP-OUTPUT-FILE>
+ by running B<objdump> on the B<llc> compiled and linked binary. You need to
+ trim down the result so it contains only the function of interest.
+
+ Second, you create I<GDB-DISASSEMBLY-FILE> by running B<gdb>, with my patch
+ to print out hex bytes in the B<disassemble> command output, on
+ B<lli>. Set a breakpoint in C<Emitter::finishFunction()> and wait until
+ the function you want is compiled. Then use the B<disassemble> command
+ to print out the assembly dump of the function B<lli> just compiled.
+ (Use C<lli -debug> to find out where the function starts and ends in memory.)
+ It's easiest to save this output by using B<script>.
+
+ Finally, you run B<codegen-diff>, as indicated in the Synopsis section of
+ this manpage. It will print out a two-line stanza for each mismatched
+ instruction, with the B<llc> version first, and the B<lli> version second.
+
+ =head1 AUTHOR
+
+ B<codegen-diff> was written by Brian Gaeke.
+
+ =head1 SEE ALSO
+
+ L<gdb(1)>, L<objdump(1)>, L<script(1)>.
+
+ You will need my B<gdb> patch:
+
+ http://llvm.cs.uiuc.edu/~gaeke/gdb-disassembly-print-bytes.patch
+
+ =cut
More information about the llvm-commits
mailing list