On Fri, Sep 30, 2016 at 11:28:28AM -0600, Jeff Law wrote: > On 09/30/2016 11:01 AM, James Greenhalgh wrote: > > > >Hi, > > > >This patch ports the logic from m68k's TARGET_FLT_EVAL_METHOD to the new > >target hook TARGET_C_EXCESS_PRECISION. > > > >Patch tested by building an m68k-none-elf toolchain and running > >m68k.exp (without the ability to execute) with no regressions, and manually > >inspecting the output assembly code when compiling > >testsuite/gcc.target/i386/excess-precision* to show no difference in > >code-generation. > > > >OK? > > > >Thanks, > >James > > > >--- > >gcc/ > > > >2016-09-30 James Greenhalgh <james.greenha...@arm.com> > > > > * config/m68k/m68k.c (m68k_excess_precision): New. > > (TARGET_C_EXCESS_PRECISION): Define. > OK when prereqs are approved. Similarly for other targets where you > needed to add this hook.
Thanks Jeff, Andreas, I spotted a very silly bug when I was retesting this patch set - when I swapped the namespace for the new traget macro it changed from TARGET_EXCESS_PRECISION to TARGET_C_EXCESS_PRECISION but I failed to update the m68k patch to reflect that. This second revision fixes that (obvious) oversight. OK? Thanks, James --- gcc/ 2016-10-14 James Greenhalgh <james.greenha...@arm.com> * config/m68k/m68k.c (m68k_excess_precision): New. (TARGET_C_EXCESS_PRECISION): Define.
diff --git a/gcc/config/m68k/m68k.c b/gcc/config/m68k/m68k.c index b152ca8..3edeb71 100644 --- a/gcc/config/m68k/m68k.c +++ b/gcc/config/m68k/m68k.c @@ -183,6 +183,8 @@ static rtx m68k_function_arg (cumulative_args_t, machine_mode, static bool m68k_cannot_force_const_mem (machine_mode mode, rtx x); static bool m68k_output_addr_const_extra (FILE *, rtx); static void m68k_init_sync_libfuncs (void) ATTRIBUTE_UNUSED; +static enum flt_eval_method +m68k_excess_precision (enum excess_precision_type); /* Initialize the GCC target structure. */ @@ -323,6 +325,9 @@ static void m68k_init_sync_libfuncs (void) ATTRIBUTE_UNUSED; #undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA #define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA m68k_output_addr_const_extra +#undef TARGET_C_EXCESS_PRECISION +#define TARGET_C_EXCESS_PRECISION m68k_excess_precision + /* The value stored by TAS. */ #undef TARGET_ATOMIC_TEST_AND_SET_TRUEVAL #define TARGET_ATOMIC_TEST_AND_SET_TRUEVAL 128 @@ -6531,4 +6536,36 @@ m68k_epilogue_uses (int regno ATTRIBUTE_UNUSED) == m68k_fk_interrupt_handler)); } + +/* Implement TARGET_C_EXCESS_PRECISION. + + Set the value of FLT_EVAL_METHOD in float.h. When using 68040 fp + instructions, we get proper intermediate rounding, otherwise we + get extended precision results. */ + +static enum flt_eval_method +m68k_excess_precision (enum excess_precision_type type) +{ + switch (type) + { + case EXCESS_PRECISION_TYPE_FAST: + /* The fastest type to promote to will always be the native type, + whether that occurs with implicit excess precision or + otherwise. */ + return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT; + case EXCESS_PRECISION_TYPE_STANDARD: + case EXCESS_PRECISION_TYPE_IMPLICIT: + /* Otherwise, the excess precision we want when we are + in a standards compliant mode, and the implicit precision we + provide can be identical. */ + if (TARGET_68040 || ! TARGET_68881) + return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT; + + return FLT_EVAL_METHOD_PROMOTE_TO_LONG_DOUBLE; + default: + gcc_unreachable (); + } + return FLT_EVAL_METHOD_UNPREDICTABLE; +} + #include "gt-m68k.h"