This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[doc patch] document powerpc abi changes
- From: Alan Modra <amodra at bigpond dot net dot au>
- To: gcc-patches at gcc dot gnu dot org
- Cc: David Edelsohn <dje at watson dot ibm dot com>, gerald at pfeifer dot com
- Date: Mon, 22 Nov 2004 13:55:09 +1030
- Subject: [doc patch] document powerpc abi changes
I would like to apply the following, which documents PowerPC ABI
changes (that I know about!) for gcc-3.4. Tested by
http://validator.w3.org/
--- htdocs/gcc-3.4/changes.html 2004-11-15 07:55:42.000000000 +1030
+++ changes.html 2004-11-22 13:28:25.150853629 +1030
@@ -955,7 +955,16 @@
aimed at reducing the reliance on assembler macros.</li>
</ul>
-<h3>PowerPC Darwin</h3>
+<h3>PowerPC</h3>
+ <ul>
+ <li>GCC 3.4 releases have a number of fixes for PowerPC and PowerPC64
+ <a href="powerpc-abi.html"> ABI incompatibilities</a>
+ regarding the way parameters are passed during functions calls.
+ These changes may result in incompatibility between code compiled
+ with GCC 3.3 and GCC 3.4.</li>
+ </ul>
+
+<h4>PowerPC Darwin</h4>
<ul>
<li>Support for shared/dylib gcc libraries has been added. It is
enabled by default on <code>powerpc-apple-darwin7.0.0</code>
@@ -965,6 +974,23 @@
powerpc-apple-darwin7.0.0</code> you need to install <a
href="http://www.opendarwin.org/projects/dlcompat/";>dlcompat</a>.
</li>
+
+ <li>128 bit IBM extended precision format support added for long
+ doubles.</li>
+ </ul>
+
+<h4>PowerPC64 Linux</h4>
+ <ul>
+ <li> By default, PowerPC64 Linux now uses natural alignment of
+ structure elements. The old four byte alignment for doubles, with
+ special rules for structs starting with a double, can be chosen with
+ -malign-power. This change may result in incompatibility between
+ code compiled with GCC 3.3 and GCC 3.4.</li>
+
+ <li>-mabi=altivec is now the default rather than -mabi=no-altivec.</li>
+
+ <li>128 bit IBM extended precision format support added for long
+ doubles.</li>
</ul>
<h3>S/390 and zSeries</h3>
--- /dev/null 2003-09-22 23:30:38.000000000 +0930
+++ powerpc-abi.html 2004-11-22 13:31:16.534935949 +1030
@@ -0,0 +1,132 @@
+<html>
+
+<head>
+<title>PowerPC ABI Changes in GCC 3.4 - GNU Project - Free Software Foundation (FSF)</title>
+</head>
+
+<body>
+<h1>PowerPC ABI Changes in GCC 3.4</h1>
+
+<p>The GCC 3.4 release series fixed several cases in which earlier
+ releases did not follow the proper calling conventions. These
+ fixes unfortunately did not all happen in one release, so in certain
+ rare cases code will be incompatible when using different version
+ compilers within the release series.</p>
+
+
+<h2>Fixed in 3.4.0</h2>
+
+ <ul>
+ <li>_Complex integer types smaller than _Complex long, and _Complex
+ floating point types smaller than _Complex double were passed and
+ returned incorrectly in one register instead of the real and
+ imaginary components being in separate registers. This has been
+ fixed for PowerPC64 Linux and AIX, with PowerPC Linux and Darwin
+ opting to stay compatible with GCC 3.3.</li>
+ </ul>
+
+ <ul>
+ <li>Under -mabi=altivec, PowerPC64 Linux used to pass vector args to
+ non-prototyped functions only in vector registers. Without a
+ prototype the arg is supposed to be passed in both vector registers
+ and in gprs or on the stack.</li>
+ </ul>
+
+ <ul>
+ <li>DWARF2 debug information failed to use the correct numbers for
+ registers other than gprs or fprs. This affected both PowerPC and
+ PowerPC64 Linux.</li>
+ </ul>
+
+ <ul>
+ <li>Many bugs fixed related to mixed-mode code, that is, use of
+ 64-bit instructions and data on a 32-bit ABI. Previously, 64-bit
+ values such as long long were passed in one register instead of
+ being split into separate 32-bit registers as required by the ABI.</li>
+ </ul>
+
+ <ul>
+ <li>Zero size arrays were incorrectly passed by reference in GCC
+ 3.3.1 and GCC 3.3.0, a regression from GCC 3.2. This was also fixed
+ in GCC 3.3.2.</li>
+ </ul>
+
+ <ul>
+ <li>Small structures passed by value were passed at the wrong end of
+ registers, or padded at the wrong end in memory, or passed on the
+ stack when they should be passed in registers. AIX and PowerPC64
+ Linux support has been changed to comply with their respective ABIs,
+ which necessarily introduces an incompatibility with GCC 3.3.
+ "Small" here means less than eight bytes for the 64 bit
+ ABIs, and less than four bytes for AIX-32.</li>
+ </ul>
+
+
+<h2>Fixed in 3.4.1</h2>
+
+ <ul>
+ <li>Under -mabi=no-altivec, vector parameters were not aligned, and
+ wrong code was used to access them if the target function used
+ vector instructions. Furthermore, -maltivec affected the way vector
+ parameters were passed and returned. This affected all ABIs, both
+ 32 bit and 64 bit. The fix necessarily introduces an incompatibility
+ between GCC 3.3 and GCC 3.4, but not for the more common
+ -mabi=altivec case.</li>
+ </ul>
+
+ <ul>
+ <li>PowerPC Linux had a number of problems with functions accepting
+ variable arguments, with disagreement between caller and callee over
+ where arguments were to be found. Affected were: all _Complex types,
+ 128 bit long double, and vectors under -mabi=altivec when all altivec
+ registers had been used.</li>
+ </ul>
+
+ <ul>
+ <li>PowerPC64 Linux and AIX generated incorrect code for certain
+ _Complex types in functions accepting variable arguments. _Complex
+ integer types smaller than _Complex long, and _Complex floating
+ point types smaller than _Complex double, were not properly
+ repackaged from registers used to pass the values.</li>
+ </ul>
+
+ <ul>
+ <li>A 128 bit long double partially passed in fp13 (because
+ fp2..fp12 were used for other fp args) also incorrectly used fp14
+ instead of passing the remainder of the arg on the stack. Affected
+ PowerPC64 Linux, AIX, and Darwin.</li>
+ </ul>
+
+
+<h2>Fixed in 3.4.2</h2>
+
+ <ul>
+ <li>Arguments requiring alignment, and also passed partly in
+ registers and partly on the stack, did not have all the stack words
+ written. This bug affected AIX-32 and Darwin in the following cases:
+ <ul><li>Under -mabi=altivec, vector args passed in the variable argument
+ part of a function with an ellipsis in the prototype.</li>
+ <li>Under -mabi=no-altivec, all vector args.</li></ul>
+ To be affected by this bug, the vector arg needs to be passed in r9,
+ r10 and two stack words, and r8 not be used for passing other args.</li>
+ </ul>
+
+ <ul>
+ <li>AIX-32 didn't pass 128 bit long doubles correctly when they
+ needed to be passed both in gprs and fprs (variable arg functions,
+ or not prototyped). For example, if the long double should have
+ been passed partially in r8,r9,r10 then only r8 was set.</li>
+ </ul>
+
+ <ul>
+ <li>A 128 bit long double passed partially in fprs and partially on
+ the stack caused a compiler abort on PowerPC64 Linux and AIX-64.</li>
+ </ul>
+
+ <ul>
+ <li>Mixed-mode parameter passing bugs fixed for PowerPC Linux.
+ Somehow, PowerPC Linux missed the fixes applied for the other 32 bit
+ ABIs for GCC 3.4.0.</li>
+ </ul>
+</body>
+</html>
--
Alan Modra
IBM OzLabs - Linux Technology Centre