Complete vld1 instructions with some corrections #1216
Conversation
|
r? @Amanieu (rust-highfive has picked a reviewer for you, use r? to override) |
| name = vmull | ||
| n-suffix | ||
| name = vmull_n | ||
| no-q |
There was a problem hiding this comment.
Maybe this change should be documented in the PR or in an extra commit/extra PR as it changes the names of public intrinsics (to their correct names).
There was a problem hiding this comment.
Thanks for pointing it out! I added relevant documents to the description of the PR.
| @@ -3568,7 +3733,7 @@ generate int16x4_t:i16:int16x4_t, int32x2_t:i32:int32x2_t | |||
|
|
|||
| /// Vector saturating doubling multiply high with scalar | |||
| name = vqdmulhq_n | |||
| out-suffix | |||
| no-q | |||
SparrowLii
changed the title
| #[target_feature(enable = "neon,aes")] | ||
| #[cfg_attr(test, assert_instr(ldr))] | ||
| pub unsafe fn vld1q_p64(ptr: *const p64) -> poly64x2_t { | ||
| transmute(u64x2::new(*ptr, *ptr.offset(1))) |
There was a problem hiding this comment.
There was a concern raised in #1148 that this doesn't get optimized down to a single instruction.
There was a problem hiding this comment.
I agree, all others have been already changed to use read_unaligned() instead which cannot be seen in the diff as it this PR is based on an older commit.
There was a problem hiding this comment.
That makes sense. I will rebase the PR and make changes accordingly.
Update stdarch submodule This is mainly to fix the critical issue of aarch64 store intrinsics overwriting additional memory, see rust-lang/stdarch#1220 Changes: * aarch64/armv7: additional vld1/vst1 intrinsics + perf fixes for existing ones * rust-lang/stdarch#1205 * rust-lang/stdarch#1207 * rust-lang/stdarch#1216 * armv7: Make FMA work with vfpv4 and optimize * rust-lang/stdarch#1219 * Non-visible changes to the testing framework * rust-lang/stdarch#1208 * rust-lang/stdarch#1211 * rust-lang/stdarch#1213 * rust-lang/stdarch#1215 * rust-lang/stdarch#1218
This PR does the following:
(1) completes
vld1neon instructions, mainly for*_p64s(2) replaces
cryptofeature withaesfor arm(3) adds
"core_arch/src/arm_shared/neon"dir instdarch-verify. Based on this, the function names ofvmull_n_s16,vmull_n_s32,vmull_n_u16,vmull_n_u32,vqdmulhq_n_s16,vqdmulhq_n_s32have been corrected(4) optimizes
stdarch-genso thatvld2*(and other) instructions can be generated more clearly(For the convenience of reviewing, they are not generated in this PR)