[libc-commits] [libc] [libc][libm][GPU] Add support for fast vendor math (PR #66439)

[Matt Arsenault via libc-commits]

================
@@ -117,6 +117,17 @@ add_entrypoint_object(
     -O2
 )
 
+add_entrypoint_object(
+  fast_cosf
+  SRCS
+    cosf.cpp
+  HDRS
+    ../../cosf.h
+  COMPILE_OPTIONS
+    ${bitcode_link_flags}
+    -D__CLANG_GPU_APPROX_TRANSCENDENTALS__ -O3 -ffast-math
----------------
arsenm wrote:

I think for something like this you should only enable approximate functions, or whatever minimum set of flags you need. You don't need to break inf/nan handling for most of these (plus the cases that require explicit handling are optimizable out once we propagate nofpclass by default)

https://github.com/llvm/llvm-project/pull/66439