| Summary: | [4.6/4.7 Regression] ICE in convert_move | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Jakub Jelinek <jakub> |
| Component: | middle-end | Assignee: | Jakub Jelinek <jakub> |
| Status: | RESOLVED FIXED | ||
| Severity: | normal | CC: | yufeng |
| Priority: | P3 | Keywords: | ice-on-valid-code |
| Version: | 4.6.0 | ||
| Target Milestone: | 4.6.1 | ||
| See Also: | https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54429 | ||
| Host: | Target: | ||
| Build: | Known to work: | 4.5.2 | |
| Known to fail: | 4.6.0, 4.7.0 | Last reconfirmed: | 2011-03-29 15:50:00 |
| Bug Depends on: | 48358 | ||
| Bug Blocks: | |||
| Attachments: |
gcc46-pr48335.patch
gcc46-pr48335.patch gcc46-pr48335.patch |
||
And:
/* { dg-do compile } */
/* { dg-options "-O2 -fno-tree-sra" } */
typedef __float128 T __attribute__((__may_alias__));
struct S
{
_Complex double d __attribute__((aligned (16)));
};
void bar (struct S);
void
foo (T x)
{
struct S s;
*(T *) &s.d = x;
__real__ s.d *= 7.0;
bar (s);
}
seems to be quietly miscompiled (instead of storing the 128 bit __float128 over both real and imaginary parts (it is __may_alias__, so it should be fine aliasing-wise) it converts the __float128 to double and stores just over real part. In 4.5 &s.d was present and s was addressable, but ADDR_EXPR in MEM_EXPR is ignored and thus in 4.6 we happily put s into (concat:DC (reg:DF ...) (reg:DF ...)).
Yet another testcase:
/* { dg-do compile } */
/* { dg-options "-O2 -fno-tree-sra" } */
typedef long long T __attribute__((__may_alias__));
struct S
{
_Complex float d __attribute__((aligned (8)));
};
void bar (struct S);
void
foo (T x)
{
struct S s;
*(T *) &s.d = x;
__real__ s.d *= 7.0;
bar (s);
}
which ICEs too. BTW, in the original testcase from
https://bugzilla.redhat.com/show_bug.cgi?id=691133 -fno-tree-sra wasn't used, just delta did a poor job on it even after many hours (got it to 32KB), so I've decided to try to write a testcase myself and for that I needed -fno-tree-sra.
Created attachment 23814 [details] gcc46-pr48335.patch WIP patch I've been playing with. Unfortunately, it shows that even without CONCAT we ICE, e.g. on: struct S { float d; }; void bar (struct S); void foo (int x) { struct S s = { .d = 0.0f }; *(char *) &s.d = x; s.d *= 7.0; bar (s); } at -O2 -fno-tree-sra. Created attachment 23815 [details] gcc46-pr48335.patch Updated patch, with more testcases. gcc.dg/pr48335-2.c still ICEs with -m32 and gcc.dg/pr48335-3.c for both -m32/-m64 (gcc.dg/pr48335-3.c is not complex related BTW). A patch that might fix most of the issues could also be
Index: gcc/tree-ssa.c
===================================================================
--- gcc/tree-ssa.c (revision 171716)
+++ gcc/tree-ssa.c (working copy)
@@ -1976,7 +1976,8 @@ maybe_optimize_var (tree var, bitmap add
/* Do not change TREE_ADDRESSABLE if we need to preserve var as
a non-register. Otherwise we are confused and forget to
add virtual operands for it. */
- && (!is_gimple_reg_type (TREE_TYPE (var))
+ && ((!is_gimple_reg_type (TREE_TYPE (var))
+ && TYPE_MODE (TREE_TYPE (var)) == BLKmode)
|| !bitmap_bit_p (not_reg_needs, DECL_UID (var))))
{
TREE_ADDRESSABLE (var) = 0;
pessimizing tree alias analysis (but only to be as bad as 4.5, maybe
a little worse for register types with components (vectors and complex vars)).
Created attachment 23818 [details] gcc46-pr48335.patch Updated patch, all the new tests now compile, on i386/x86_64/ppc/ppc64/s390/s390x. Author: jakub Date: Fri Apr 1 21:13:29 2011 New Revision: 171855 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=171855 Log: PR middle-end/48335 * expr.c (expand_assignment): Handle all possibilities if TO_RTX is CONCAT. * expmed.c (store_bit_field_1): Avoid trying to create invalid SUBREGs. (store_split_bit_field): If SUBREG_REG (op0) or op0 itself has smaller mode than word, return it for offset 0 and const0_rtx for out-of-bounds stores. If word is const0_rtx, skip it. * gcc.c-torture/compile/pr48335-1.c: New test. * gcc.dg/pr48335-1.c: New test. * gcc.dg/pr48335-2.c: New test. * gcc.dg/pr48335-3.c: New test. * gcc.dg/pr48335-4.c: New test. * gcc.dg/pr48335-5.c: New test. * gcc.dg/pr48335-6.c: New test. * gcc.dg/pr48335-7.c: New test. * gcc.dg/pr48335-8.c: New test. * gcc.target/i386/pr48335-1.c: New test. Added: trunk/gcc/testsuite/gcc.c-torture/compile/pr48335-1.c trunk/gcc/testsuite/gcc.dg/pr48335-1.c trunk/gcc/testsuite/gcc.dg/pr48335-2.c trunk/gcc/testsuite/gcc.dg/pr48335-3.c trunk/gcc/testsuite/gcc.dg/pr48335-4.c trunk/gcc/testsuite/gcc.dg/pr48335-5.c trunk/gcc/testsuite/gcc.dg/pr48335-6.c trunk/gcc/testsuite/gcc.dg/pr48335-7.c trunk/gcc/testsuite/gcc.dg/pr48335-8.c trunk/gcc/testsuite/gcc.target/i386/pr48335-1.c Modified: trunk/gcc/ChangeLog trunk/gcc/expmed.c trunk/gcc/expr.c trunk/gcc/testsuite/ChangeLog Author: jakub Date: Wed Apr 6 20:39:20 2011 New Revision: 172063 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=172063 Log: Backported from mainline 2011-04-01 Jakub Jelinek <jakub@redhat.com> PR middle-end/48335 * expr.c (expand_assignment): Handle all possibilities if TO_RTX is CONCAT. * expmed.c (store_bit_field_1): Avoid trying to create invalid SUBREGs. (store_split_bit_field): If SUBREG_REG (op0) or op0 itself has smaller mode than word, return it for offset 0 and const0_rtx for out-of-bounds stores. If word is const0_rtx, skip it. * gcc.c-torture/compile/pr48335-1.c: New test. * gcc.dg/pr48335-1.c: New test. * gcc.dg/pr48335-2.c: New test. * gcc.dg/pr48335-3.c: New test. * gcc.dg/pr48335-4.c: New test. * gcc.dg/pr48335-5.c: New test. * gcc.dg/pr48335-6.c: New test. * gcc.dg/pr48335-7.c: New test. * gcc.dg/pr48335-8.c: New test. * gcc.target/i386/pr48335-1.c: New test. Added: branches/gcc-4_6-branch/gcc/testsuite/gcc.c-torture/compile/pr48335-1.c branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/pr48335-1.c branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/pr48335-2.c branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/pr48335-3.c branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/pr48335-4.c branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/pr48335-5.c branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/pr48335-6.c branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/pr48335-7.c branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/pr48335-8.c branches/gcc-4_6-branch/gcc/testsuite/gcc.target/i386/pr48335-1.c Modified: branches/gcc-4_6-branch/gcc/ChangeLog branches/gcc-4_6-branch/gcc/expmed.c branches/gcc-4_6-branch/gcc/expr.c branches/gcc-4_6-branch/gcc/testsuite/ChangeLog Fixed. (In reply to comment #6) > Created attachment 23818 [details] > gcc46-pr48335.patch > > Updated patch, all the new tests now compile, on > i386/x86_64/ppc/ppc64/s390/s390x. Hi Jakub, The test gcc.dg/pr48335-2.c unfortunately fails in the arm-eabi testing with an ICE: gcc.dg/pr48335-2.c: In function 'f1': gcc.dg/pr48335-2.c:19:30: internal compiler error: in expand_expr_addr_expr_1, at expr.c:6922 I've raised the bug 48493 for this. I'm not sure how it is related with this bugzilla as the compiler before your commit fails on the test as well. Cheers, Yufeng |
/* { dg-do compile } */ /* { dg-options "-O2 -fno-tree-sra -msse2" } */ #include <emmintrin.h> struct S { _Complex double d __attribute__((aligned (16))); }; void bar (struct S); void foo (__m128d x, struct S y) { struct S s; _mm_store_pd ((double *) &s.d, x); __real__ s.d *= 7.0; bar (s); } ICEs in convert_move, starting with http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=161655