https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100041
--- Comment #15 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to Richard Biener from comment #12) > A possible solution might be to disallow the -m64 -m96bit-long-double > combination, the documentation suggests -m128bit-long-double was intended > as an "optimization" over the x86-32 ABI. Yes, I think this is the way to go. While it is possible to "fix" the ICE with: --cut here-- diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md index 2820f6d6188..e2fa01756fd 100644 --- a/gcc/config/i386/i386.md +++ b/gcc/config/i386/i386.md @@ -3050,7 +3050,7 @@ (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -16)))) (match_operand:XF 0 "nonmemory_no_elim_operand" "f,r,*r,C"))] - "TARGET_64BIT" + "TARGET_64BIT && !TARGET_128BIT_LONG_DOUBLE" "#" "&& 1" [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -16))) @@ -3077,7 +3077,7 @@ (define_insn "*pushxf" [(set (match_operand:XF 0 "push_operand" "=<,<,<,<,<") (match_operand:XF 1 "general_no_elim_operand" "f,r,*r,oF,oC"))] - "" + "!TARGET_64BIT || TARGET_128BIT_LONG_DOUBLE" { /* This insn should be already split before reg-stack. */ return "#"; --cut here-- where the above patch prevents the push from memory, it is quite obvious from ix86_split_long_move that TARGET_128BIT_LONG_DOUBLE is indeed intended as an optimization for 32bit targets only. So, let's kill -m96bit-long-double for TARGET_64BIT.