[arm64] Avoid sign-extending TYP_INT register moves#129864
Conversation
Revise ins_Move_Extend to emit mov instead of a sxtw for TYP_INT moves. Add an optional dstType arg to assert the move is valid and non-widening. Fixes dotnet#129052. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
github-actions
Bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
|
Impacts about 150K instructions across 62K methods in SPMI. @tannergooding PTAL |
There was a problem hiding this comment.
Pull request overview
This PR adjusts ARM64 JIT move/extend selection so that TYP_INT register-to-register moves no longer emit sxtw (sign-extend) and instead use mov, avoiding unnecessary work when the value remains int-typed. It also adds an optional dstType parameter to help assert move validity (non-widening) and introduces a JIT disasm-based regression test.
Changes:
- Update
ins_Move_Extend(ARM64) to returnINS_movforTYP_INTreg-reg moves instead ofINS_sxtw. - Thread an optional
dstTypethroughinst_Mov_Extend/ins_Move_Extendand add a debug assert to validate non-widening moves. - Add a new ARM64 disasm-check regression test under
InstructionCombining.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| src/coreclr/jit/instr.cpp | ARM64 ins_Move_Extend now prefers mov for TYP_INT reg-reg moves; adds optional dstType assertion plumbing. |
| src/coreclr/jit/codegen.h | Updates declarations to include the new optional dstType parameter. |
| src/tests/JIT/opt/InstructionCombining/IntMoveNoSignExtend.cs | New disasm-check test validating no sxtw is emitted for int moves and that real int -> long widening still uses sxtw. |
| src/tests/JIT/opt/InstructionCombining/IntMoveNoSignExtend.csproj | New test project wiring with disasm checks and environment variable settings. |
Broaden the int/long mov comment to mention the EA_4BYTE zero extension, and simplify the test's sxtw disasm check to a bare ARM64-NOT. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
|
@tannergooding ping |
| * dstType - type the value will be consumed as; defaults to srcType. Must share srcType's | ||
| * actual type, as genuine widening (e.g. int -> long) must use a cast instead. |
There was a problem hiding this comment.
I'm not sure this comment is entirely accurate "as is" because GT_CAST itself calls ins_Move_Extend to get the register it needs for some code, such as float->double (genFloatToFloatCast)
I'm also not quite sure this is the "right fix"... You're adding in a new parameter that nothing actually uses and which just causes all architectures to do "extra work" here, when only Arm64 has changed and is really just no longer ever selecting sxtw.
I'd think this is rather something that no architecture should ever be producing an int->long widening here for, and so we're effectively only handling reg->reg and small->int widening loads (i.e. using the same general if (!varTypeIsSmall(srcType)) { } else if (varTypeIsUnsigned(srcType)) { } else { } pattern every other arch is doing)
-or- that widening up to the full register is needed for Arm64 for some platforms and so there needs to be an extra argument passed down for such architectures, which is actually used to select mov vs sxtw, and that one of the half dozen callsites (I think genReturn is the one you're effectively fixing here) needs to be updated to handle that arg on those platforms.
Revise ins_Move_Extend to emit mov instead of a sxtw for TYP_INT moves. Add an optional dstType arg to assert the move is valid and non-widening.
Fixes #129052.