Fix SIMD temp register interference #58914
Conversation
On Windows/x86, if a call returns a SIMD8 in two registers (eax, edx), and needs to be assembled and stored into a single SIMD register, AND we don't have the SSE41 ABI available (because the hardware doesn't support it, or we specifically disable it), then we need a temporary register. In the failure case, the target register and temporary register were the same, and then interfered with each other. The fix is to force the temporary register to be different from the target register. Fixes dotnet#58899
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT Issue DetailsOn Windows/x86, if a call returns a SIMD8 in two registers Fixes #58899
|
|
@tannergooding @kunalspathak @dotnet/jit-contrib PTAL |
|
Testing is successful; just need a review to merge. |
kunalspathak
left a comment
kunalspathak
left a comment
There was a problem hiding this comment.
Looks like it got introduced when we added support genReg1/genReg2->SIMD8 in
#52581
LGTM.
|
/backport to release/6.0 |
|
Started backporting to release/6.0: https://github.com/dotnet/runtime/actions/runs/1224803781 |
2 participants
On Windows/x86, if a call returns a SIMD8 in two registers
(eax, edx), and needs to be assembled and stored into a single
SIMD register, AND we don't have the SSE41 ABI available (because
the hardware doesn't support it, or we specifically disable it),
then we need a temporary register. In the failure case, the target
register and temporary register were the same, and then interfered
with each other. The fix is to force the temporary register to
be different from the target register.
No SPMI asm diffs
Fixes #58899