diff options
| author | Joe Ramsay <Joe.Ramsay@arm.com> | 2024-05-16 09:16:36 +0100 |
|---|---|---|
| committer | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2024-05-21 14:38:22 +0100 |
| commit | c39cf53702e570b64a70d8f7fdfe277017acfb5f (patch) | |
| tree | b703107f8e69e8ee3896e6608c16fbf0985bf4bf | |
| parent | c35cad016b2a5bc71fa2330f26e35afdef327ccc (diff) | |
| download | glibc-c39cf53702e570b64a70d8f7fdfe277017acfb5f.tar.gz glibc-c39cf53702e570b64a70d8f7fdfe277017acfb5f.tar.xz glibc-c39cf53702e570b64a70d8f7fdfe277017acfb5f.zip | |
Compile libmvec with -fno-math-errno
Rounding intrinsics may not be inlined without -fno-math-errno. libmvec is free to do what it likes with errno, so disable it for better performance. Tested with no regression on aarch64 and x86_64. Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
| -rw-r--r-- | Makeconfig | 7 |
1 files changed, 4 insertions, 3 deletions
diff --git a/Makeconfig b/Makeconfig index 59e21d4c82..9fe664ae3a 100644 --- a/Makeconfig +++ b/Makeconfig @@ -899,9 +899,10 @@ endif +math-flags = -frounding-math # Logically only "libnldbl", "nonlib" and "testsuite" should be using -# -fno-math-errno. However due to GCC bug #88576, only "libm" can use -# -fno-math-errno. -+extra-math-flags = $(if $(filter libm,$(in-module)),-fno-math-errno,-fmath-errno) +# -fno-math-errno. However due to GCC bug #88576, only "libm" and +# "libmvec" can use -fno-math-errno. ++extra-math-flags = $(if $(filter libmvec libm,$(in-module)),\ + -fno-math-errno,-fmath-errno) # Use 64 bit time_t support for installed programs installed-modules = nonlib nscd ldconfig locale_programs \ |