[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #14 from hjl dot tools at gmail dot com 2009-02-18 23:42 --- Fixed. -- hjl dot tools at gmail dot com changed: What|Removed |Added Status|NEW |RESOLVED Resolution||FIXED http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #13 from hjl at gcc dot gnu dot org 2009-02-18 21:40 --- Subject: Bug 39224 Author: hjl Date: Wed Feb 18 21:40:08 2009 New Revision: 144272 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=144272 Log: 2009-02-18 H.J. Lu PR target/39224 * config/i386/i386.c (ix86_return_in_memory): Properly check ABI. Modified: trunk/gcc/ChangeLog trunk/gcc/config/i386/i386.c -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #12 from hjl dot tools at gmail dot com 2009-02-18 19:50 --- The updated patch is at http://gcc.gnu.org/ml/gcc-patches/2009-02/msg00871.html -- hjl dot tools at gmail dot com changed: What|Removed |Added URL|http://gcc.gnu.org/ml/gcc- |http://gcc.gnu.org/ml/gcc- |patches/2009- |patches/2009- |02/msg00870.html|02/msg00871.html http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #11 from hjl dot tools at gmail dot com 2009-02-18 19:22 --- A patch is posted at http://gcc.gnu.org/ml/gcc-patches/2009-02/msg00870.html -- hjl dot tools at gmail dot com changed: What|Removed |Added URL||http://gcc.gnu.org/ml/gcc- ||patches/2009- ||02/msg00870.html http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #10 from hjl dot tools at gmail dot com 2009-02-18 18:53 --- The problem is callee returns long double via a pointer to a structure. But caller thinks callee returns long double in rax/edx. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #9 from hjl dot tools at gmail dot com 2009-02-18 15:50 --- For sysv/x86-64, XFmode is 16byte with 16byte alignment. It is passed in memory and returned in in %st0/%st1. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #8 from ktietz at gcc dot gnu dot org 2009-02-18 14:23 --- (In reply to comment #7) > (In reply to comment #6) > > (In reply to comment #5) > > > (In reply to comment #4) > > > > ok, I found the issue, which causes here the problem. > > > > The x86_64 abi returns TFmode in rax,edx which is stored in aligned > > > > stack > > > > > > XFmode > > > > > > > right, sorry I meant XFmode > > > > w64 ABI doesn't cover XFmode. Why do you use integer registers for > XFmode? It doesn't make any senses to me. > As I said above there are two reasons. First 32-bit code supports XFmode, too. Also the vendor specific library supports 96-bit floating point, too. But not as native type, they use a struct for this (what is the current implementation in gcc to pass them as memory reference). So there is no good reason for not supporting XFmode types for w64 IMHO. This issue here seems to be more the store size in x86_64 abi. Hasn't x86_64 also the need to have those types store 16-bytes aligned? -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #7 from hjl dot tools at gmail dot com 2009-02-18 14:13 --- (In reply to comment #6) > (In reply to comment #5) > > (In reply to comment #4) > > > ok, I found the issue, which causes here the problem. > > > The x86_64 abi returns TFmode in rax,edx which is stored in aligned stack > > > > XFmode > > > > right, sorry I meant XFmode > w64 ABI doesn't cover XFmode. Why do you use integer registers for XFmode? It doesn't make any senses to me. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #6 from ktietz at gcc dot gnu dot org 2009-02-18 12:11 --- (In reply to comment #5) > (In reply to comment #4) > > ok, I found the issue, which causes here the problem. > > The x86_64 abi returns TFmode in rax,edx which is stored in aligned stack > > XFmode > right, sorry I meant XFmode -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #5 from ubizjak at gmail dot com 2009-02-18 12:06 --- (In reply to comment #4) > ok, I found the issue, which causes here the problem. > The x86_64 abi returns TFmode in rax,edx which is stored in aligned stack XFmode -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #4 from ktietz at gcc dot gnu dot org 2009-02-18 10:45 --- ok, I found the issue, which causes here the problem. The x86_64 abi returns TFmode in rax,edx which is stored in aligned stack variable as 96 bits, but the upper 32-bits (which have to be zero initialized) aren't set. For w64 there are stored 128-bit in this case as desired. So the compare of those two results is failing, because on x86_64 abi the upper 32-bits have random values. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #3 from ktietz at gcc dot gnu dot org 2009-02-18 08:47 --- (In reply to comment #2) > I am verifying it at the moment for w64 target, if we have here same issues. > Yes, on w64 targets we have the same issue. By adding print methods, it seems that the return value of the foreign abi is passed/treated wrong. This testcases were run by me last time about beginning of January this year and they seem to work fine at this date. -- ktietz at gcc dot gnu dot org changed: What|Removed |Added Status|UNCONFIRMED |NEW Ever Confirmed|0 |1 Last reconfirmed|-00-00 00:00:00 |2009-02-18 08:47:04 date|| http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #2 from ktietz at gcc dot gnu dot org 2009-02-18 08:18 --- (In reply to comment #1) > I think long double on w64 is the same as double. I am not sure if > gcc.dg/callabi/func-1.c is a valid test. > the long double is supported as 96-bit floating point for gcc. This isn't as ms_abi declares it, but I follow here the 32-bit port. So we allow higher precision, and by vendor specific headers it has not much affects, beside that the vendor specific scanf/printf functions can't print it proper. Therefore the mingw runtime provides C99 compliant (and gnu-style compiant) alternatives to print/scan them. I am verifying it at the moment for w64 target, if we have here same issues. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
--- Comment #1 from hjl dot tools at gmail dot com 2009-02-18 06:45 --- I think long double on w64 is the same as double. I am not sure if gcc.dg/callabi/func-1.c is a valid test. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
[Bug target/39224] ABI attribute doesn't work with long double
-- hjl dot tools at gmail dot com changed: What|Removed |Added Target Milestone|--- |4.4.0 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39224
