diff options
| author | Vincent Chen <vincentc@andestech.com> | 2018-11-22 11:14:36 +0800 |
|---|---|---|
| committer | Greentime Hu <greentime@andestech.com> | 2018-11-22 18:13:27 +0800 |
| commit | 44e92e0364adfd7b6759084e02a550d06336d896 (patch) | |
| tree | 4fb8af79a5ce25f9467bdc22f44e67f6ae42b6d6 /arch/nds32/include/asm | |
| parent | 1ac832509f2ea1b566f0c06f98f308f58b03d098 (diff) | |
nds32: support denormalized result through FP emulator
Currently, the nds32 FPU dose not support the arithmetic of denormalized
number. When the nds32 FPU finds the result of the instruction is a
denormlized number, the nds32 FPU considers it to be an underflow condition
and rounds the result to an appropriate number. It may causes some loss
of precision. This commit proposes a solution to re-execute the
instruction by the FPU emulator to enhance the precision. To transfer
calculations from user space to kernel space, this feature will enable
the underflow exception trap by default. Enabling this feature may cause
some side effects:
1. Performance loss due to extra FPU exception
2. Need another scheme to control real underflow trap
A new parameter, UDF_trap, which is belong to FPU context is used
to control underflow trap.
User can configure this feature via CONFIG_SUPPORT_DENORMAL_ARITHMETIC
Signed-off-by: Vincent Chen <vincentc@andestech.com>
Acked-by: Greentime Hu <greentime@andestech.com>
Signed-off-by: Greentime Hu <greentime@andestech.com>
Diffstat (limited to 'arch/nds32/include/asm')
| -rw-r--r-- | arch/nds32/include/asm/elf.h | 11 | ||||
| -rw-r--r-- | arch/nds32/include/asm/fpu.h | 11 | ||||
| -rw-r--r-- | arch/nds32/include/asm/syscalls.h | 1 |
3 files changed, 23 insertions, 0 deletions
diff --git a/arch/nds32/include/asm/elf.h b/arch/nds32/include/asm/elf.h index f5f9cf7e0544..95f3ea253e4c 100644 --- a/arch/nds32/include/asm/elf.h +++ b/arch/nds32/include/asm/elf.h @@ -9,6 +9,7 @@ */ #include <asm/ptrace.h> +#include <asm/fpu.h> typedef unsigned long elf_greg_t; typedef unsigned long elf_freg_t[3]; @@ -159,8 +160,18 @@ struct elf32_hdr; #endif + +#if IS_ENABLED(CONFIG_FPU) +#define FPU_AUX_ENT NEW_AUX_ENT(AT_FPUCW, FPCSR_INIT) +#else +#define FPU_AUX_ENT NEW_AUX_ENT(AT_IGNORE, 0) +#endif + #define ARCH_DLINFO \ do { \ + /* Optional FPU initialization */ \ + FPU_AUX_ENT; \ + \ NEW_AUX_ENT(AT_SYSINFO_EHDR, \ (elf_addr_t)current->mm->context.vdso); \ } while (0) diff --git a/arch/nds32/include/asm/fpu.h b/arch/nds32/include/asm/fpu.h index 9b1107b58e23..019f1bcfc5ee 100644 --- a/arch/nds32/include/asm/fpu.h +++ b/arch/nds32/include/asm/fpu.h @@ -28,7 +28,18 @@ extern int do_fpuemu(struct pt_regs *regs, struct fpu_struct *fpu); #define sNAN64 0xFFFFFFFFFFFFFFFFULL #define sNAN32 0xFFFFFFFFUL +#if IS_ENABLED(CONFIG_SUPPORT_DENORMAL_ARITHMETIC) +/* + * Denormalized number is unsupported by nds32 FPU. Hence the operation + * is treated as underflow cases when the final result is a denormalized + * number. To enhance precision, underflow exception trap should be + * enabled by default and kerenl will re-execute it by fpu emulator + * when getting underflow exception. + */ +#define FPCSR_INIT FPCSR_mskUDFE +#else #define FPCSR_INIT 0x0UL +#endif extern const struct fpu_struct init_fpuregs; diff --git a/arch/nds32/include/asm/syscalls.h b/arch/nds32/include/asm/syscalls.h index 78778ecff60c..da32101b455d 100644 --- a/arch/nds32/include/asm/syscalls.h +++ b/arch/nds32/include/asm/syscalls.h @@ -7,6 +7,7 @@ asmlinkage long sys_cacheflush(unsigned long addr, unsigned long len, unsigned int op); asmlinkage long sys_fadvise64_64_wrapper(int fd, int advice, loff_t offset, loff_t len); asmlinkage long sys_rt_sigreturn_wrapper(void); +asmlinkage long sys_udftrap(int option); #include <asm-generic/syscalls.h> |