// test case, compile with "-O1 -fstrict-aliasing -ftree-vectorize" : void foo (float a[32], float b[2][32]) { int i; for (i = 0; i < 32; i++) a[i] = (b[0][i] > b[1][i]) ? b[0][i] : b[1][i]; } // compiler output: psy.i: In function 'foo': psy.i:6: error: unrecognizable insn: (insn 98 97 99 11 (set (reg:V4SF 110) (unspec:V4SF [ (reg:V4SF 89 [ vect_var_.40 ]) (reg:V4SF 85 [ vect_var_.47 ]) ] 52)) -1 (nil) (nil)) psy.i:6: internal compiler error: in extract_insn, at recog.c:2084
This is a target bug as the tree optimizers get it right: MEM[base: vect_p.55, index: D.1801] = VEC_COND_EXPR < vect_var_.40 > vect_var_.47 , vect_var_.40 , vect_var_.47 > ;
Confirmed, a 4.1 regression.
Patch.
Subject: Bug 23816 CVSROOT: /cvs/gcc Module name: gcc Changes by: rth@gcc.gnu.org 2005-09-13 19:02:05 Modified files: gcc : ChangeLog gcc/config/i386: sse.md Added files: gcc/testsuite/gcc.dg/vect: pr23816-1.c pr23816-2.c Log message: PR target/23816 * config/i386/sse.md (*ieee_sminv4sf3, *ieee_smaxv4sf3) (*ieee_sminv2df3, *ieee_smaxv2df3): New insn patterns. Patches: http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.9946&r2=2.9947 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/config/i386/sse.md.diff?cvsroot=gcc&r1=1.24&r2=1.25 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/gcc.dg/vect/pr23816-1.c.diff?cvsroot=gcc&r1=NONE&r2=1.1 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/gcc.dg/vect/pr23816-2.c.diff?cvsroot=gcc&r1=NONE&r2=1.1
Fixed.