https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118067
--- Comment #6 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to Uroš Bizjak from comment #5) > The problematic insn is still: > > (insn 9 5 10 2 (parallel [ > (set (reg:HI 99 [ _2 ]) > (lshiftrt:HI (subreg:HI (reg/v:V32HI 165 [ u ]) 0) > (const_int 1 [0x1]))) > (clobber (reg:CC 17 flags)) > ]) "pr118067.c":10:8 1215 {*lshrhi3_1} > (expr_list:REG_UNUSED (reg:CC 17 flags) > (nil))) > > But now reload tries to reload input operand with: > > Inserting insn reload before: > 191: r337:HI=r165:V32HI#0 > > Creating newreg=338, assigning class NO_REGS to secondary r338 > 191: r337:HI=r165:V32HI#0 > Inserting the sec. move after: > 192: r337:HI#0=r338:SI Here, reload could spill HImode value from r165 to memory and reload it from memory to r337. *movhi_internal pattern supports both instructions: (define_insn "*movhi_internal" [(set (match_operand:HI 0 "nonimmediate_operand" "=r,r,r,m ,*k,*k ,r ,m ,*k ,?r,?*v,*Yv,*v,*v,jm,m") (match_operand:HI 1 "general_operand" "r ,n,m,rn,r ,*km,*k,*k,CBC,*v,r ,C ,*v,m ,*x,*v"))] Please note that tuning for generic target disables direct moves from XMM to GPR, so alternative (?r,*v) is effectively disabled. Adding -mtune=intel to compile flags avoids the ICE.