* [patch] fix sysv ppc small struct ffi_call
@ 2013-02-07 20:58 Andreas Tobler
2013-02-08 20:33 ` Anthony Green
0 siblings, 1 reply; 2+ messages in thread
From: Andreas Tobler @ 2013-02-07 20:58 UTC (permalink / raw)
To: libffi-discuss, Anthony Green
[-- Attachment #1: Type: text/plain, Size: 590 bytes --]
Hi Anthony,
Unfortunately almost a year too late I noticed a severe breakage in the
area of small structs with the sysv ABI on powerpc. Not the gcc_sysv.
At the time when the soft float patch for ppc went in this breakage
happened.
This is the git commit:
https://github.com/atgreen/libffi/commit/52891f8a93f9b8de801cca4cf05639422dc9773e
Attached the patch which fixes that. I tested on ppc-unknown-freebsd and
ppc-unknown-linux, no regressions.
Would you mind applying this patch for the 3.0.12 spin?
And thank you for applying the first fix where LONGDOUBLE==DOUBLE.
TIA!
Andreas
[-- Attachment #2: libffi_sysv_ppc.diff --]
[-- Type: text/plain, Size: 3269 bytes --]
--- src/powerpc/ffi.c.orig 2013-02-07 19:51:28.000000000 +0100
+++ src/powerpc/ffi.c 2013-02-07 19:56:08.000000000 +0100
@@ -48,6 +48,11 @@
FLAG_RETURNS_128BITS = 1 << (31-27), /* cr6 */
+ FLAG_SYSV_SMST_R4 = 1 << (31-26), /* use r4 for FFI_SYSV 8 byte
+ structs. */
+ FLAG_SYSV_SMST_R3 = 1 << (31-25), /* use r3 for FFI_SYSV 4 byte
+ structs. */
+
FLAG_ARG_NEEDS_COPY = 1 << (31- 7),
#ifndef __NO_FPRS__
FLAG_FP_ARGUMENTS = 1 << (31- 6), /* cr1.eq; specified by ABI */
@@ -686,18 +691,35 @@
break;
case FFI_TYPE_STRUCT:
- /*
- * The final SYSV ABI says that structures smaller or equal 8 bytes
- * are returned in r3/r4. The FFI_GCC_SYSV ABI instead returns them
- * in memory.
- *
- * NOTE: The assembly code can safely assume that it just needs to
- * store both r3 and r4 into a 8-byte word-aligned buffer, as
- * we allocate a temporary buffer in ffi_call() if this flag is
- * set.
- */
- if (cif->abi == FFI_SYSV && size <= 8)
- flags |= FLAG_RETURNS_SMST;
+ if (cif->abi == FFI_SYSV)
+ {
+ /* The final SYSV ABI says that structures smaller or equal 8 bytes
+ are returned in r3/r4. The FFI_GCC_SYSV ABI instead returns them
+ in memory. */
+
+ /* Treat structs with size <= 8 bytes. */
+ if (size <= 8)
+ {
+ flags |= FLAG_RETURNS_SMST;
+ /* These structs are returned in r3. We pack the type and the
+ precalculated shift value (needed in the sysv.S) into flags.
+ The same applies for the structs returned in r3/r4. */
+ if (size <= 4)
+ {
+ flags |= FLAG_SYSV_SMST_R3;
+ flags |= 8 * (4 - size) << 8;
+ break;
+ }
+ /* These structs are returned in r3 and r4. See above. */
+ if (size <= 8)
+ {
+ flags |= FLAG_SYSV_SMST_R3 | FLAG_SYSV_SMST_R4;
+ flags |= 8 * (8 - size) << 8;
+ break;
+ }
+ }
+ }
+
intarg_count++;
flags |= FLAG_RETVAL_REFERENCE;
/* Fall through. */
--- src/powerpc/sysv.S.orig 2013-02-07 19:50:25.000000000 +0100
+++ src/powerpc/sysv.S 2013-02-07 19:51:06.000000000 +0100
@@ -142,14 +142,19 @@
#endif
L(small_struct_return_value):
- /*
- * The C code always allocates a properly-aligned 8-byte bounce
- * buffer to make this assembly code very simple. Just write out
- * r3 and r4 to the buffer to allow the C code to handle the rest.
- */
- stw %r3, 0(%r30)
- stw %r4, 4(%r30)
- b L(done_return_value)
+ extrwi %r6,%r31,2,19 /* number of bytes padding = shift/8 */
+ mtcrf 0x02,%r31 /* copy flags to cr[24:27] (cr6) */
+ extrwi %r5,%r31,5,19 /* r5 <- number of bits of padding */
+ subfic %r6,%r6,4 /* r6 <- number of useful bytes in r3 */
+ bf- 25,L(done_return_value) /* struct in r3 ? if not, done. */
+/* smst_one_register: */
+ slw %r3,%r3,%r5 /* Left-justify value in r3 */
+ mtxer %r6 /* move byte count to XER ... */
+ stswx %r3,0,%r30 /* ... and store that many bytes */
+ bf+ 26,L(done_return_value) /* struct in r3:r4 ? */
+ add %r6,%r6,%r30 /* adjust pointer */
+ stswi %r4,%r6,4 /* store last four bytes */
+ b L(done_return_value)
.LFE1:
END(ffi_call_SYSV)
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [patch] fix sysv ppc small struct ffi_call
2013-02-07 20:58 [patch] fix sysv ppc small struct ffi_call Andreas Tobler
@ 2013-02-08 20:33 ` Anthony Green
0 siblings, 0 replies; 2+ messages in thread
From: Anthony Green @ 2013-02-08 20:33 UTC (permalink / raw)
To: Andreas Tobler; +Cc: libffi-discuss
On Thu, Feb 7, 2013 at 3:58 PM, Andreas Tobler <andreast-list@fgznet.ch> wrote:
> Attached the patch which fixes that. I tested on ppc-unknown-freebsd and
> ppc-unknown-linux, no regressions.
>
> Would you mind applying this patch for the 3.0.12 spin?
Done -- Thank you Andreas!
AG
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2013-02-08 20:33 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2013-02-07 20:58 [patch] fix sysv ppc small struct ffi_call Andreas Tobler
2013-02-08 20:33 ` Anthony Green
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).