On Thu, Feb 16, 2017 at 05:46:30PM +0000, Segher Boessenkool wrote:
> The testcases pr59833.c and pr61441.c check whether signaling NaNs as
> input to some operation result in quiet NaNs. Without -fsignaling-nans
> this is not guaranteed to happen. So, this patch add this option to
> these testcases.
>
> Tested on powerpc64-linux {-m32,-m64}, on aarch64-linux, and on
> x86_64-linux (where there is no issignaling, huh -- I tested on gcc20).
pr61441.c now fails since your change on i?86-linux (and x86_64-linux -m32).
Before combine we have:
(insn 6 5 7 2 (set (reg:DF 89 [ x ])
(float_extend:DF (mem/u/c:SF (symbol_ref/u:SI ("*.LC0") [flags 0x2]) [1
S4 A32]))) "pr61441.c":12 152 {*extendsfdf2}
(expr_list:REG_EQUAL (const_double:DF +SNaN [+SNaN])
(nil)))
(insn 7 6 8 2 (set (mem:DF (pre_dec:SI (reg/f:SI 7 sp)) [2 S8 A64])
(reg:DF 89 [ x ])) "pr61441.c":13 119 {*pushdf}
(expr_list:REG_DEAD (reg:DF 89 [ x ])
(expr_list:REG_ARGS_SIZE (const_int 16 [0x10])
(nil))))
(not sure if that REG_EQUAL is not already invalid for -fsignaling-nans),
and combine turns that into:
(insn 7 6 8 2 (set (mem:DF (pre_dec:SI (reg/f:SI 7 sp)) [2 S8 A64])
(const_double:DF +SNaN [+SNaN])) "pr61441.c":13 119 {*pushdf}
(expr_list:REG_ARGS_SIZE (const_int 16 [0x10])
(nil)))
which doesn't drop the signal bit.
Jakub
