(The same phenomenon happens in 4.2 subversion alpha.)
Incorrect code is generated when using -ftree-vectorize on the SSE2 Pentium 4
target. (The same code works on the AMD64 64-bit target.)
See attached source code.
$ gcc -O2 -Wall -march=pentium4 -mfpmath=sse -ftree-vectorize
-ftree-vectorizer-verbose=1 gcc_plantage2.c -o ./gcc_plantage2
$ ./gcc_plantage2
will segfault (the same program works perfectly without -ftree-vectorize).
The segfault appears at the second movapd instruction:
movapd .LC0, %xmm0
.L17:
movapd %xmm0, (%eax)
addl $1, %edx
addl $16, %eax
cmpl %edx, %ebx
ja .L17
%eax is at this point equal to 4 modulo 16, which results in a segmentation
fault, since movapd assumes 16-byte alignment.
--
Summary: wrong alignment or incorrect address computation in
vectorized code on Pentium 4 SSE
Product: gcc
Version: 4.0.3
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: David dot Monniaux at ens dot fr
GCC build triplet: i486-linux-gnu
GCC host triplet: i486-linux-gnu
GCC target triplet: i486-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25413
