JIT: Emulate missing x86 shift instructions for xplat intrinsics #111108
Conversation
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
dotnet-policy-service
bot
added
the
community-contribution
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
| emitAttr attr3 = attr; | ||
| if (hasTupleTypeInfo(ins) && ((insTupleTypeInfo(ins) & INS_TT_MEM128) != 0)) | ||
| { | ||
| // Shift instructions take xmm for the 3rd operand regardless of instruction size. | ||
| attr3 = EA_16BYTE; | ||
| } |
There was a problem hiding this comment.
Not quite sure this is the "best" fix.
This feels like a case where we could utilize the existing tags for memory operand size for the register as well rather than hardcoding a one-off special case.
Not going to block on it, especially since this is only used for disassembly; but its something that would be nice to have improved for this and other cases where the last operand is a different register size than the others.
There was a problem hiding this comment.
Can you elaborate on your preferred fix here? This is already using the tuple type rather than hard-coded instruction list -- I just mention shift in the comment explicitly because it's the only thing I'm aware of that uses full xmm/m128 for what is ultimately an 8-bit scalar, and this should be the only encoding form where that xmm fix is required.
|
CC. @dotnet/jit-contrib for secondary review |
This extends the byte shift emulation to handle arithmetic shift and adds emulation of qword arithmetic shift for platforms without AVX-512/AVX10.
SPMI diffs show the updated codegen.