Add hexagon implementations for compiler_rt builtins

[androm3da]

No description provided.

[alexrp]

I see these also referenced in LLVM:

• __hexagon_floattidf
• __hexagon_floattisf
• __hexagon_fixunssfti
• __hexagon_fixunsdfti
• __hexagon_fixsfti
• __hexagon_fixdfti
• __hexagon_fast2_sqrtf
• __hexagon_fast2_sqrtdf2
• __hexagon_adddf3 and __hexagon_fast_adddf3
• __hexagon_subdf3 and __hexagon_fast_subdf3
• __hexagon_muldf3 and __hexagon_fast_muldf3
• __hexagon_divdf3 and __hexagon_fast_divdf3
• __hexagon_divsf3 and __hexagon_fast_divsf3

Do we need these?

[androm3da]

I see these also referenced in LLVM:

• __hexagon_floattidf
• __hexagon_floattisf
• __hexagon_fixunssfti
• __hexagon_fixunsdfti
• __hexagon_fixsfti
• __hexagon_fixdfti

These are not defined in compiler-rt/lib/builtins/hexagon/ so I think they're safe to omit. But I'll follow up with the team to understand why these references aren't satisfied.

• __hexagon_fast2_sqrtf
• __hexagon_fast2_sqrtdf2
• __hexagon_adddf3 and __hexagon_fast_adddf3
• __hexagon_subdf3 and __hexagon_fast_subdf3
• __hexagon_muldf3 and __hexagon_fast_muldf3
• __hexagon_divdf3 and __hexagon_fast_divdf3
• __hexagon_divsf3 and __hexagon_fast_divsf3

Do we need these?

These were unintentionally omitted. I'll see about revising this PR to include these.

I assume it's safe/permitted/expected for me to use .set directives to create aliases from within an inline assembly block? Otherwise I will need to find another mechanism to publish the aliases.

[alexrp]

I assume it's safe/permitted/expected for me to use .set directives to create aliases from within an inline assembly block? Otherwise I will need to find another mechanism to publish the aliases.

Can you clarify what aliases you're referring to?

[androm3da]

Can you clarify what aliases you're referring to?

The __hexagon_fast_* / __qdsp_* / __hexagon_fast2_* aliases, for example:

https://github.com/llvm/llvm-project/blob/e14827f0828d14ef17ab76316e8449d1b76e2617/compiler-rt/lib/builtins/hexagon/dfaddsub.S#L58-L71

https://github.com/rust-lang/compiler-builtins/blob/a09218f1c4d23ffbd97d68f0fefb5feed2469dc5/src/hexagon/dfaddsub.s#L7-L15

[alexrp]

Hmm, I'm not sure if doing that in inline assembly is a good idea. 🤔 But if these are just aliases of the same symbol, perhaps you can just do multiple @export()s with different names?

[androm3da]

Hmm, I'm not sure if doing that in inline assembly is a good idea. 🤔 But if these are just aliases of the same symbol, perhaps you can just do multiple @export()s with different names?

I thought such a thing might be possible, so yeah, that'll do. I will work on it.

[androm3da]

Addressed review comments, fixed __hexagon_subdf3 issue (thanks @SidManning)

[alexrp]

Small nitpick for consistency.

[alexrp]

These are not defined in compiler-rt/lib/builtins/hexagon/ so I think they're safe to omit. But I'll follow up with the team to understand why these references aren't satisfied.

This looks like an LLVM bug to me. They seem to be used for i128 <-> f32/f64 conversions. Clang disables __int128 for Hexagon, so it's not an issue there, but it is an issue for Zig:

❯ cat test.zig
export fn cvt(d: i128) f64 {
    return @floatFromInt(d);
}
❯ zig4 build-lib test.zig -fno-compiler-rt -target hexagon-freestanding
❯ llvm-objdump -r --disassemble-symbols=cvt libtest.a

libtest.a(./libtest.a.o):       file format elf32-hexagon

Disassembly of section .text:

00000000 <cvt>:
       0:       02 c0 9d a0     a09dc002 {      allocframe(#0x10) }
       4:       04 c2 03 f5     f503c204 {      r5:4 = combine(r3,r2) }
       8:       04 c0 01 f5     f501c004 {      r5:4 = combine(r1,r0) }
       c:       ff e2 de a7     a7dee2ff {      memd(r30+#-0x8) = r3:2 }
      10:       fe e0 de a7     a7dee0fe {      memd(r30+#-0x10) = r1:0 }
      14:       00 c0 00 5a     5a00c000 {      call 0x14 }                     00000014:  R_HEX_B22_PCREL      __hexagon_floattidf
      18:       1e c0 1e 96     961ec01e {      dealloc_return }

[alexrp]

So, I think these calls should just all be deleted. That way, the code generator will emit calls to the usual __floattidf, etc. (Unless of course these Hexagon-optimized routines actually exist somewhere and just haven't been upstreamed to LLVM yet.)

[androm3da]

(Unless of course these Hexagon-optimized routines actually exist somewhere and just haven't been upstreamed to LLVM yet.)

I think this may be the case. I'll check and update the PR soon.

[alexrp]

Ok, given llvm/llvm-project#117423, I think this PR is good to go after the remaining align(32) suggestions are accepted.

[androm3da]

Ok, given llvm/llvm-project#117423, I think this PR is good to go after the remaining align(32) suggestions are accepted.

Ok - I've taken those, thanks.

re: llvm/llvm-project#117423 -- if it turns out that it's not preferred to remove those, then I'll probably add the specializations to compiler-rt instead (and then corresponding ones here).