ARC:fpu: add extra capability check before use of sqrt and fma builtins
Checks
| Context |
Check |
Description |
| dj/TryBot-apply_patch |
success
|
Patch applied to master at the time it was sent
|
| dj/TryBot-32bit |
success
|
Build for i686
|
Commit Message
From: Pavel Kozlov <pavel.kozlov@synopsys.com>
Add extra check for compiler definitions to ensure that compiler provides
sqrt and fma hw fpu instructions else use software implementation.
As divide/sqrt and FMA hw support from CPU side is optional,
the compiler can be configured by options to generate hw FPU instructions,
but without use of FDDIV, FDSQRT, FSDIV, FSSQRT, FDMADD and FSMADD
instructions. In this case __builtin_sqrt and __builtin_sqrtf provided by
compiler can't be used inside the glibc code, as these builtins are used
in implementations of sqrt() and sqrtf() functions but at the same time
these builtins unfold to sqrt() and sqrtf(). So it is possible to receive
code like that:
0001c4b4 <__ieee754_sqrtf>:
1c4b4: 0001 0000 b 0 ;1c4b4 <__ieee754_sqrtf>
The same is also true for __builtin_fma and __builtin_fmaf.
---
sysdeps/arc/fpu/math-use-builtins-fma.h | 14 ++++++++++++--
sysdeps/arc/fpu/math-use-builtins-sqrt.h | 14 ++++++++++++--
2 files changed, 24 insertions(+), 4 deletions(-)
Comments
On 21/12/22 13:28, Pavel.Kozlov--- via Libc-alpha wrote:
> From: Pavel Kozlov <pavel.kozlov@synopsys.com>
>
> Add extra check for compiler definitions to ensure that compiler provides
> sqrt and fma hw fpu instructions else use software implementation.
>
> As divide/sqrt and FMA hw support from CPU side is optional,
> the compiler can be configured by options to generate hw FPU instructions,
> but without use of FDDIV, FDSQRT, FSDIV, FSSQRT, FDMADD and FSMADD
> instructions. In this case __builtin_sqrt and __builtin_sqrtf provided by
> compiler can't be used inside the glibc code, as these builtins are used
> in implementations of sqrt() and sqrtf() functions but at the same time
> these builtins unfold to sqrt() and sqrtf(). So it is possible to receive
> code like that:
>
> 0001c4b4 <__ieee754_sqrtf>:
> 1c4b4: 0001 0000 b 0 ;1c4b4 <__ieee754_sqrtf>
>
> The same is also true for __builtin_fma and __builtin_fmaf.
> ---
> sysdeps/arc/fpu/math-use-builtins-fma.h | 14 ++++++++++++--
> sysdeps/arc/fpu/math-use-builtins-sqrt.h | 14 ++++++++++++--
> 2 files changed, 24 insertions(+), 4 deletions(-)
>
> diff --git a/sysdeps/arc/fpu/math-use-builtins-fma.h b/sysdeps/arc/fpu/math-use-builtins-fma.h
> index eede75aa41be..082badf48201 100644
> --- a/sysdeps/arc/fpu/math-use-builtins-fma.h
> +++ b/sysdeps/arc/fpu/math-use-builtins-fma.h
> @@ -1,4 +1,14 @@
> -#define USE_FMA_BUILTIN 1
> -#define USE_FMAF_BUILTIN 1
> +#if defined __ARC_FPU_DP_DIV__
> +# define USE_SQRT_BUILTIN 1
> +#else
> +# define USE_SQRT_BUILTIN 0
> +#endif
> +
> +#if defined __ARC_FPU_SP_DIV__
> +# define USE_SQRTF_BUILTIN 1
> +#else
> +# define USE_SQRTF_BUILTIN 0
> +#endif
> +
> #define USE_FMAL_BUILTIN 0
> #define USE_FMAF128_BUILTIN 0
This is wrong, sqrt use macro do not belong for the fma switch file.
> diff --git a/sysdeps/arc/fpu/math-use-builtins-sqrt.h b/sysdeps/arc/fpu/math-use-builtins-sqrt.h
> index e94c915ba66a..a449bc609295 100644
> --- a/sysdeps/arc/fpu/math-use-builtins-sqrt.h
> +++ b/sysdeps/arc/fpu/math-use-builtins-sqrt.h
> @@ -1,4 +1,14 @@
> -#define USE_SQRT_BUILTIN 1
> -#define USE_SQRTF_BUILTIN 1
> +#if defined __ARC_FPU_DP_DIV__
> +# define USE_SQRT_BUILTIN 1
> +#else
> +# define USE_SQRT_BUILTIN 0
> +#endif
> +
> +#if defined __ARC_FPU_SP_DIV__
> +# define USE_SQRTF_BUILTIN 1
> +#else
> +# define USE_SQRTF_BUILTIN 0
> +#endif
> +
> #define USE_SQRTL_BUILTIN 0
> #define USE_SQRTF128_BUILTIN 0
This is ok.
> This is wrong, sqrt use macro do not belong for the fma switch file.
Thank you for the review and your notice. I was inattentive when
moving changes from the branch I had. This has been fixed in
v2 of the patch [1].
I've manually checked (by objdump -d output review) that now
expected code for libm is generated in different cases (when compiler
provides support for extra instructions and sets macroses
__ARC_FPU_DP_DIV__, __ARC_FPU_SP_DIV__, __ARC_FPU_DP_FMA__,
__ARC_FPU_SP_FMA__ and when not).
[1]
http://lists.infradead.org/pipermail/linux-snps-arc/2023-January/006771.html
--
Pavel
@@ -1,4 +1,14 @@
-#define USE_FMA_BUILTIN 1
-#define USE_FMAF_BUILTIN 1
+#if defined __ARC_FPU_DP_DIV__
+# define USE_SQRT_BUILTIN 1
+#else
+# define USE_SQRT_BUILTIN 0
+#endif
+
+#if defined __ARC_FPU_SP_DIV__
+# define USE_SQRTF_BUILTIN 1
+#else
+# define USE_SQRTF_BUILTIN 0
+#endif
+
#define USE_FMAL_BUILTIN 0
#define USE_FMAF128_BUILTIN 0
@@ -1,4 +1,14 @@
-#define USE_SQRT_BUILTIN 1
-#define USE_SQRTF_BUILTIN 1
+#if defined __ARC_FPU_DP_DIV__
+# define USE_SQRT_BUILTIN 1
+#else
+# define USE_SQRT_BUILTIN 0
+#endif
+
+#if defined __ARC_FPU_SP_DIV__
+# define USE_SQRTF_BUILTIN 1
+#else
+# define USE_SQRTF_BUILTIN 0
+#endif
+
#define USE_SQRTL_BUILTIN 0
#define USE_SQRTF128_BUILTIN 0