[PATCH] Fix LTO decimal ICE

[Jakub Jelinek <jakub@redhat.com>]

On Wed, May 21, 2014 at 08:52:03PM +0200, Jakub Jelinek wrote:
> FAIL: c-c++-common/ubsan/float-cast-overflow-10.c  -O2 -flto -fno-use-linker-plugin -flto-partition=none  (internal compiler error)
> FAIL: c-c++-common/ubsan/float-cast-overflow-10.c  -O2 -flto -fno-use-linker-plugin -flto-partition=none  (test for excess errors)
> FAIL: c-c++-common/ubsan/float-cast-overflow-10.c  -O2 -flto -fuse-linker-plugin -fno-fat-lto-objects  (internal compiler error)
> FAIL: c-c++-common/ubsan/float-cast-overflow-10.c  -O2 -flto -fuse-linker-plugin -fno-fat-lto-objects  (test for excess errors)
> FAIL: c-c++-common/ubsan/float-cast-overflow-7.c  -O2 -flto -fno-use-linker-plugin -flto-partition=none  (internal compiler error)
> FAIL: c-c++-common/ubsan/float-cast-overflow-7.c  -O2 -flto -fno-use-linker-plugin -flto-partition=none  (test for excess errors)
> FAIL: c-c++-common/ubsan/float-cast-overflow-7.c  -O2 -flto -fuse-linker-plugin -fno-fat-lto-objects  (internal compiler error)
> FAIL: c-c++-common/ubsan/float-cast-overflow-7.c  -O2 -flto -fuse-linker-plugin -fno-fat-lto-objects  (test for excess errors)
> 
> The LTO ICEs on float-cast-overflow-{7,10}.c seems to be related to decimal support:
> /usr/src/gcc/gcc/testsuite/c-c++-common/ubsan/float-cast-overflow-7.c:147:1: internal compiler error: in decimal_to_decnumber, at dfp.c:138
> 0x116b4cb decimal_to_decnumber
>         ../../gcc/dfp.c:138
...

This bug is because we leave garbage in REAL_VALUE_TYPE padding bits, but
then use memcmp on REAL_VALUE_TYPE objects.  All other places use memset
first to clear the padding bits, so I've committed this fix as obvious
to trunk and 4.9 (without testcase, because I couldn't reproduce it on
anything smaller than float-cast-overflow-{7,10}.c and those require
further gcc patches.

> The execution test FAILs for -m32 are:
> ==4494==Sanitizer CHECK failed: ../../../../../libsanitizer/ubsan/ubsan_value.cc:98 ((0 && "unexpected floating point bit width")) != (0) (0, 0)

This one can be fixed by handling 96 the same as 80 and 128, apparently even 
clang/llvm uses 96 on i?86 and crashes in libubsan the same way.  Marek, can
you please handle the LLVM bureaucracy?

2014-05-22  Jakub Jelinek  <jakub@redhat.com>

	* tree-streamer-in.c (unpack_ts_real_cst_value_fields): Make sure
	all padding bits in REAL_VALUE_TYPE are cleared.

--- gcc/tree-streamer-in.c.jj	2014-05-20 16:37:05.000000000 +0200
+++ gcc/tree-streamer-in.c	2014-05-22 09:40:01.300112220 +0200
@@ -168,6 +168,9 @@ unpack_ts_real_cst_value_fields (struct
   REAL_VALUE_TYPE r;
   REAL_VALUE_TYPE *rp;
 
+  /* Clear all bits of the real value type so that we can later do
+     bitwise comparisons to see if two values are the same.  */
+  memset (&r, 0, sizeof r);
   r.cl = (unsigned) bp_unpack_value (bp, 2);
   r.decimal = (unsigned) bp_unpack_value (bp, 1);
   r.sign = (unsigned) bp_unpack_value (bp, 1);


	Jakub