COMMON: Use compiler intrinsics in SWAP_BYTES_16 when available #3962
Conversation
common/endian.h
Outdated
| @@ -106,6 +106,20 @@ | |||
| return result; | |||
| } | |||
| } | |||
|
|
|||
| // Test for GCC-compatible | |||
| #elif GCC_ATLEAST(4, 8) || defined(__clang__) | |||
There was a problem hiding this comment.
I wonder if this one shouldn't come first (before PSP one) because it's cleaner than writing a piece of assembly code.
There was a problem hiding this comment.
Forget about it, I just saw how it's done for 32-bits.
There was a problem hiding this comment.
It might still be worth assessing how useful the MIPS routines are with modern compilers - according to the Compiler Explorer, GCC 5.4 for MIPS at least already generates wsbh/ror instructions when using the GCC builtins, and for MIPS64r2 the assembly code looks like it'll be slower than the builtins because it doesn't use dsbh/dshd for SWAP_BYTES_64. I'm not sure about the special Allegrex instructions, though.
There was a problem hiding this comment.
For Allegrex, I will do some test but it should be OK according to the code:
pspdev/gcc@1530c47#diff-dad60b351908e704b73fd0b46c72674f8fc9fff6fea1900b902d8b04fae3fb1fR5871
There was a problem hiding this comment.
I can confirm that with latest PSP toolchain (GCC 11.2.0 allegrex), __builtin_bswap32 generates wsbw and __builtin_bswap16 generates wsbh.
There was a problem hiding this comment.
OK, I've added another commit that removes the MIPS versions.
No description provided.