This patch adds new function libc_feholdsetround_noex_aarch64_ctx, enabling

further optimization. libc_feholdsetround_aarch64_ctx now only needs to
read the FPCR in the typical case, avoiding a redundant FPSR read.
Performance results show a good improvement (5-10% on sin()) on cores with
expensive FPCR/FPSR instructions.
This commit is contained in:
Wilco Dijkstra 2014-08-07 16:29:55 +00:00
parent 538e9e454d
commit 656b84c2ef
2 changed files with 32 additions and 3 deletions

View File

@ -1,3 +1,8 @@
2014-08-07 Wilco Dijkstra <wdijkstr@arm.com>
* sysdeps/aarch64/fpu/math_private.h
(libc_feholdsetround_noex_aarch64_ctx): New function.
2014-08-07 Wilco Dijkstra <wdijkstr@arm.com>
* sysdeps/arm/armv6/strcpy.S (strcpy):

View File

@ -228,12 +228,9 @@ static __always_inline void
libc_feholdsetround_aarch64_ctx (struct rm_ctx *ctx, int r)
{
fpu_control_t fpcr;
fpu_fpsr_t fpsr;
int round;
_FPU_GETCW (fpcr);
_FPU_GETFPSR (fpsr);
ctx->env.__fpsr = fpsr;
/* Check whether rounding modes are different. */
round = (fpcr ^ r) & _FPU_FPCR_RM_MASK;
@ -263,6 +260,33 @@ libc_feresetround_aarch64_ctx (struct rm_ctx *ctx)
#define libc_feresetroundf_ctx libc_feresetround_aarch64_ctx
#define libc_feresetroundl_ctx libc_feresetround_aarch64_ctx
static __always_inline void
libc_feholdsetround_noex_aarch64_ctx (struct rm_ctx *ctx, int r)
{
fpu_control_t fpcr;
fpu_fpsr_t fpsr;
int round;
_FPU_GETCW (fpcr);
_FPU_GETFPSR (fpsr);
ctx->env.__fpsr = fpsr;
/* Check whether rounding modes are different. */
round = (fpcr ^ r) & _FPU_FPCR_RM_MASK;
ctx->updated_status = round != 0;
/* Set the rounding mode if changed. */
if (__glibc_unlikely (round != 0))
{
ctx->env.__fpcr = fpcr;
_FPU_SETCW (fpcr ^ round);
}
}
#define libc_feholdsetround_noex_ctx libc_feholdsetround_noex_aarch64_ctx
#define libc_feholdsetround_noexf_ctx libc_feholdsetround_noex_aarch64_ctx
#define libc_feholdsetround_noexl_ctx libc_feholdsetround_noex_aarch64_ctx
static __always_inline void
libc_feresetround_noex_aarch64_ctx (struct rm_ctx *ctx)
{