Implement the fneg instruction. #1462
Conversation
| ; %3: double = fneg %1 | ||
| {{_}} {{_}}: mov r.64.x, 0x8000000000000000 | ||
| {{_}} {{_}}: movq fp.128.y, r.64.x | ||
| {{_}} {{_}}: xorpd fp.128.a, fp.128.y |
There was a problem hiding this comment.
I was curious that %3 was allocated a fresh register (fp.128.a), instead of reusing the register previously assigned to the now-dead %2 (fp.128.z). Could indicate an error in my register allocator changes?
There was a problem hiding this comment.
We currently prefer to allocate "higher" registers, which is pointless in FP-land, but in GP-land means we're less likely to get twonked by x64's fondness for RAX and the sysv abi's preference for "lower" registers in CALLs. Does that guarantee what you're seeing is intended? Nope. But it might give you a clue to what look for.
|
LGTM. |
ltratt
requested review from
jacob-hughes
and removed request for
jacob-hughes
|
LLVM synced. OK to squash? |
|
Please squash. |
Required for LuaJIT tests.
|
squashed. |
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
|
Once #1463 is in, this should merge. It's the ambiguous methods warning again. |
|
Not yet :) |
|
I deliberately merged both at once, because CI can run the merge together. There is no harm in doing so. |
Required for LuaJIT tests.
Please carefully check the register allocator changes. I had to implement
RegConstraint::Temporaryinassign_fp_regs().Requires: ykjit/ykllvm#214