Hi! In TYPE_MODE we have a hack for vector types where we dynamically adjust it based on whether it is used from a function where the vector mode is or isn't supported (depending on target attribute, function multiversioning, etc.), but we don't have anything like that in DECL_MODE.
The following is just a small hack that should be backportable to adjust similarly DECL_MODE when looking at fields - for static VAR_DECLs we already have code to cope with that, furthermore we need to cope there with the case where DECL_RTL is created in one context and used in another one. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2017-12-01 Jakub Jelinek <ja...@redhat.com> PR target/78643 PR target/80583 * expr.c (get_inner_reference): If DECL_MODE of a non-bitfield is BLKmode for vector field with vector raw mode, use TYPE_MODE instead of DECL_MODE. * gcc.target/i386/pr80583.c: New test. --- gcc/expr.c.jj 2017-11-27 09:27:41.000000000 +0100 +++ gcc/expr.c 2017-12-01 11:36:19.011863308 +0100 @@ -7032,7 +7032,16 @@ get_inner_reference (tree exp, HOST_WIDE size. */ mode = TYPE_MODE (DECL_BIT_FIELD_TYPE (field)); else if (!DECL_BIT_FIELD (field)) - mode = DECL_MODE (field); + { + mode = DECL_MODE (field); + /* For vector fields re-check the target flags, as DECL_MODE + could have been set with different target flags than + the current function has. */ + if (mode == BLKmode + && VECTOR_TYPE_P (TREE_TYPE (field)) + && VECTOR_MODE_P (TYPE_MODE_RAW (TREE_TYPE (field)))) + mode = TYPE_MODE (TREE_TYPE (field)); + } else if (DECL_MODE (field) == BLKmode) blkmode_bitfield = true; --- gcc/testsuite/gcc.target/i386/pr80583.c.jj 2017-12-01 11:44:28.106603314 +0100 +++ gcc/testsuite/gcc.target/i386/pr80583.c 2017-12-01 11:44:12.000000000 +0100 @@ -0,0 +1,13 @@ +/* PR target/80583 */ +/* { dg-do compile } */ +/* { dg-options "-O0 -mno-avx" } */ + +typedef int V __attribute__((__vector_size__(32))); +struct S { V a; }; + +V __attribute__((target ("avx"))) +foo (struct S *b) +{ + V x = b->a; + return x; +} Jakub