https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102285
--- Comment #18 from qinzhao at gcc dot gnu.org --- (In reply to Richard Biener from comment #13) > We could try sth like the following which should cover this testcase and > also hopefully doesn't break anything else. > > diff --git a/gcc/expr.c b/gcc/expr.c > index e0bcbccd905..ab3b5b26c20 100644 > --- a/gcc/expr.c > +++ b/gcc/expr.c > @@ -5322,7 +5322,7 @@ non_mem_decl_p (tree base) > /* Returns true if REF refers to an object that does not > reside in memory and has non-BLKmode. */ > > -static inline bool > +bool > mem_ref_refers_to_non_mem_p (tree ref) > { > tree base; > diff --git a/gcc/expr.h b/gcc/expr.h > index 94a85b40dca..85650e96834 100644 > --- a/gcc/expr.h > +++ b/gcc/expr.h > @@ -346,4 +346,6 @@ extern void expand_operands (tree, tree, rtx, rtx*, rtx*, > /* Return an rtx for the size in bytes of the value of an expr. */ > extern rtx expr_size (tree); > > +extern bool mem_ref_refers_to_non_mem_p (tree ref); > + > #endif /* GCC_EXPR_H */ > diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c > index 8312d08aab2..c1d23dfd21e 100644 > --- a/gcc/internal-fn.c > +++ b/gcc/internal-fn.c > @@ -3015,8 +3015,10 @@ expand_DEFERRED_INIT (internal_fn, gcall *stmt) > reg_lhs = true; > else > { > - rtx tem = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE); > - reg_lhs = !MEM_P (tem); > + tree lhs_base = lhs; > + while (handled_component_p (lhs_base)) > + lhs_base = TREE_OPERAND (lhs_base, 0); > + reg_lhs = mem_refers_to_non_mem_p (lhs_base); > } > > if (!reg_lhs) I tried the above patch, it even doesn't work for this testing case. After the patch, this call to .DEFERRED_INIT will be expanded through REG path, however, the existing problem in the LHS IR broke the "expand_assignment" too. so, I still think that the root cause is the IR for LHS of the following call to .DEFERRED_INIT is not correct: MEM[(int[0:D.1993] *)&fb.3] = .DEFERRED_INIT (16, 2, 1); the LHS is: MEM[(int[0:D.1993] *)&fb.3] this is the MEM_REF, whose base address is a ADDR_EXPR of a VAR_DECL, however, this VAR_DECL is not addressable, will be in REG. this is conflicting: how can a variable that in a REG be address taken?