Hi! I forgot to handle the case where lowpart_subreg returns a VOIDmode CONST_INT, in that case convert_mode_scalar obviously doesn't work.
The following patch fixes that. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2022-10-19 Jakub Jelinek <ja...@redhat.com> PR middle-end/107262 * expr.cc (convert_mode_scalar): For BFmode -> SFmode conversions of constants, use simplify_unary_operation if fromi has VOIDmode instead of recursive convert_mode_scalar. * gcc.dg/pr107262.c: New test. --- gcc/expr.cc.jj 2022-10-14 09:35:56.134991155 +0200 +++ gcc/expr.cc 2022-10-18 14:07:24.330512745 +0200 @@ -416,8 +416,15 @@ convert_mode_scalar (rtx to, rtx from, i rtx tof = NULL_RTX; if (fromi) { - rtx toi = gen_reg_rtx (toi_mode); - convert_mode_scalar (toi, fromi, 1); + rtx toi; + if (GET_MODE (fromi) == VOIDmode) + toi = simplify_unary_operation (ZERO_EXTEND, toi_mode, + fromi, fromi_mode); + else + { + toi = gen_reg_rtx (toi_mode); + convert_mode_scalar (toi, fromi, 1); + } toi = maybe_expand_shift (LSHIFT_EXPR, toi_mode, toi, GET_MODE_PRECISION (to_mode) --- gcc/testsuite/gcc.dg/pr107262.c.jj 2022-10-18 14:09:44.114607505 +0200 +++ gcc/testsuite/gcc.dg/pr107262.c 2022-10-18 14:09:33.708749313 +0200 @@ -0,0 +1,13 @@ +/* PR middle-end/107262 */ +/* { dg-do compile } */ +/* { dg-options "-O2 -ffast-math" } */ +/* { dg-add-options bfloat16 } */ +/* { dg-require-effective-target bfloat16_runtime } */ + +__bf16 +foo (__bf16 a) +{ + __bf16 b = 0; + b /= a; + return b; +} Jakub