bit_value_unop and bit_value_binop require constant values to be INTEGER_CSTs:
gcc_assert ((rval.lattice_val == CONSTANT && TREE_CODE (rval.value) == INTEGER_CST) || wi::sext (rval.mask, TYPE_PRECISION (TREE_TYPE (rhs))) == -1); However, when deciding whether to record a constant value, the for_bits_p handling in get_value_for_expr used a negative test for ADDR_EXPR: else if (is_gimple_min_invariant (expr) && (!for_bits_p || TREE_CODE (expr) != ADDR_EXPR)) This patch uses a positive test for INTEGER_CST instead. Existing tests showed the need for this once polynomial constants are added. Tested on aarch64-linux-gnu, x86_64-linux-gnu and powerpc64le-linus-gnu. OK to install? Richard 2017-09-20 Richard Sandiford <richard.sandif...@linaro.org> Alan Hayward <alan.hayw...@arm.com> David Sherwood <david.sherw...@arm.com> gcc/ * tree-ssa-ccp.c (get_value_for_expr): Use a positive test for INTEGER_CST rather than a negative test for ADDR_EXPR. Index: gcc/tree-ssa-ccp.c =================================================================== --- gcc/tree-ssa-ccp.c 2017-08-10 14:36:07.842479033 +0100 +++ gcc/tree-ssa-ccp.c 2017-09-20 13:10:31.466004578 +0100 @@ -617,7 +617,7 @@ get_value_for_expr (tree expr, bool for_ } } else if (is_gimple_min_invariant (expr) - && (!for_bits_p || TREE_CODE (expr) != ADDR_EXPR)) + && (!for_bits_p || TREE_CODE (expr) == INTEGER_CST)) { val.lattice_val = CONSTANT; val.value = expr;