JIT ARM32: Force tailcall nullchecks with LR #68107
Conversation
For interface tail calls we need to allocate registers for both the target and a null-check. The register that can be used for the target is highly constrained: it essentially only go into r12 as the arg registers may be busy and that is the only volatile register left that will not be overridden by the epilog. This leaves just `LR` left for the null check, and LSRA does not seem to handle the low amount of freedom too well (these are all allocated at the same location, which might have something to do with it). To help LSRA just force the null-check to happen with LR for fast tailcalls, which is going to be killed going into the epilog anyway. Fix dotnet#66563
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT Issue DetailsFor interface tail calls we need to allocate registers for both the To help LSRA just force the null-check to happen with LR for fast Fix #66563 cc @dotnet/jit-contrib @kunalspathak, this is the alternative fix
|
For interface tail calls we need to allocate registers for both the target and a null-check. The register that can be used for the target is highly constrained: it essentially only go into r12 as the arg registers may be busy and that is the only volatile register left that will not be overridden by the epilog. This leaves just `LR` left for the null check, and LSRA does not seem to handle the low amount of freedom too well (these are all allocated at the same location, which might have something to do with it). To help LSRA just force the null-check to happen with LR for fast tailcalls, which is going to be killed going into the epilog anyway. Fix dotnet#66563
ghost
locked as resolved and limited conversation to collaborators
2 participants
For interface tail calls we need to allocate registers for both the
target and a null-check. The register that can be used for the target is
highly constrained: it essentially only go into r12 as the arg registers
may be busy and that is the only volatile register left that will not be
overridden by the epilog. This leaves
just
LRleft for the null check, and LSRA does not seem to handle thelow amount of freedom too well (these are all allocated at the same
location, which might have something to do with it).
To help LSRA just force the null-check to happen with LR for fast
tailcalls, which is going to be killed going into the epilog anyway.
Fix #66563
cc @dotnet/jit-contrib @kunalspathak, this is the alternative fix