http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55742
--- Comment #38 from Sriraman Tallam <tmsriram at google dot com> 2013-01-18 19:53:16 UTC --- (In reply to comment #32) > Created attachment 29207 [details] > gcc48-pr55742.patch > > This bug is open for way too long given its severity, so let's start talking > over patches. > > This patch attempts to implement what I understand from Jason's comments, just > with "default" instead of "any", because it is indeed the default target > attribute (whatever you specify on the command line). > > Say on: > void foo (); > void foo () __attribute__((target ("avx"))); > void foo () __attribute__((target ("default"))); > __attribute__((target ("default"))) void foo () > { > } > __attribute__((target ("avx"))) void foo () > { > } > void (*fn) () = foo; > > first we merge the first two decls, because only if target attribute is > present > on both, we consider it for multi-versioning, for compatibility with 4.7 and > older. On e.g. > void foo (); > void foo () __attribute__((target ("sse4"))); > void foo () __attribute__((target ("default"))); > void foo () > { > } > we reject the last fn definition, because at that point foo is already known > to > be multi-versioned, thus it is required that target attribute is specified for > foo (either "default", or some other). Unfortunately, for this case the error > is reported twice for some reason. > > The #c0 testcase now compiles. > > Now, the issues I discovered with multiversioning, still unfixed by the patch: > 1) the mv*.C testcases should be moved, probably to g++.dg/ext/mv*.C > 2) can you please explain the mess in handle_target_attribute? > /* Do not strip invalid target attributes for targets which support function > multiversioning as the target string is used to determine versioned > functions. */ > else if (! targetm.target_option.valid_attribute_p (*node, name, args, > flags) > && ! targetm.target_option.supports_function_versions ()) > *no_add_attrs = true; > Why do you need that? Consider complete garbage in target attribute > arguments, > which is errored about, but the above for i386/x86_64 keeps the target > attribute around anyway, leading to lots of ICEs everywhere: Without bringing in your patch, I removed this line with patch: --- gcc/c-family/c-common.c (revision 195302) +++ gcc/c-family/c-common.c (working copy) @@ -8763,8 +8763,7 @@ multiversioning as the target string is used to determine versioned functions. */ else if (! targetm.target_option.valid_attribute_p (*node, name, args, - flags) - && ! targetm.target_option.supports_function_versions ()) + flags)) *no_add_attrs = true; return NULL_TREE; and then tried the new compiler on the following example: int foo (); int foo () __attribute__ ((target("mmx"))); int main () { return foo (); } int foo () { return 0; } int __attribute__ ((target("mmx"))) foo () { return 0; } and with -mno-mmx added to the compile options, everything is fine. However, with -mmmx in the compile options, I get: fe_example.cc: In function ‘int foo()’: fe_example.cc:16:1: error: redefinition of ‘int foo()’ foo () ^ fe_example.cc:10:1: error: ‘int foo()’ previously defined here foo () Reason is the stripping of target attributes that do not make sense. But, for MV that creates duplicate functions. I can change this to only keep the attribute tagged it is recognized by the target. That way I will strip out erroneous values for target attribute. > Consider e.g.: > __attribute__((target ("default"))) void foo (void) > { > } > __attribute__((target (128))) void foo (void) > { > } > 3) the multiversioning code assumes that target has a single argument, but it > can have more than one. Say for: > __attribute__((target ("avx,popcnt"))) void foo (void) > { > } > __attribute__((target ("popcnt","avx"))) void bar (void) > { > } > the compiler handles those two as equivalent, but with -Dbar=foo > multi-versioning only considers the first string out of that.