Assertion failed 'block->lastNode()->OperGet() == GT_RETFILT' during 'Generate code' (IL size 1480; hash 0x46e9aa75; FullOpts)

[kunalspathak]

Repros on windows x86

Environment:

set COMPlus_JitStress=1

Antigen repro: https://gist.github.com/kunalspathak/13eab94a2e925edd3a39e244d5463a8f

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch, @kunalspathak
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Repros on windows x86

Environment:

set COMPlus_JitStress=1

Antigen repro: https://gist.github.com/kunalspathak/13eab94a2e925edd3a39e244d5463a8f

Author:	kunalspathak
Assignees:	-
Labels:	arch-x86, os-windows, area-CodeGen-coreclr
Milestone:	-

[kunalspathak]

@BruceForstall - Can this be related to your recent refactoring work?

[BruceForstall]

@BruceForstall - Can this be related to your recent refactoring work?

Unknown, but I'll take a look

[BruceForstall]

This doesn't fail in the current main, but does fail with a tree from several days ago (e.g., 3713168). I'm assuming that some recent JIT change perturbed the IR but didn't actually fix the issue.

[BruceForstall]

@jakobbotsch This is due to physical promotion. I didn't know physical promotion was part of our usual stress mix at this point; is that appropriate?

On x86, GT_RETFILT must be the last operation of a filter/finally return. But it appears physical promotion inserts a "read back" after the RETFILT node:

*** NYI: Conservatively marked as read-back
STMT00096 ( 0x344[E-] ... 0x344 )
               [000568] -----------                         *  RETFILT   void
Reading back replacement V11.[000..004) -> V98 at the end of BB35

Note; the IR dump after the phase is:

***** BB35
STMT00096 ( 0x344[E-] ... 0x344 )
               [000568] -----------                         *  RETFILT   void

***** BB35
STMT00191 ( ??? ... ??? )
               [001190] -A---------                         *  ASG       int
               [001188] D------N---                         +--*  LCL_VAR   int    V98 tmp78
               [001189] -----------                         \--*  LCL_FLD   int    V11 loc10        [+0]

but the STMT00191 and [001190] and related nodes do not appear in the dump before this location. It would be helpful to dump them when they are created so it's more obvious to see who is creating them and why.

As mentioned above, this doesn't fail in main. Has it been intentionally fixed, or did IR perturbations make it "go away"?

[jakobbotsch]

@jakobbotsch This is due to physical promotion. I didn't know physical promotion was part of our usual stress mix at this point; is that appropriate?

It is meant to be fully functional (but maybe not profitable), so I'd say so.

On x86, GT_RETFILT must be the last operation of a filter/finally return. But it appears physical promotion inserts a "read back" after the RETFILT node:

*** NYI: Conservatively marked as read-back
STMT00096 ( 0x344[E-] ... 0x344 )
               [000568] -----------                         *  RETFILT   void
Reading back replacement V11.[000..004) -> V98 at the end of BB35

Note; the IR dump after the phase is:

***** BB35
STMT00096 ( 0x344[E-] ... 0x344 )
               [000568] -----------                         *  RETFILT   void

***** BB35
STMT00191 ( ??? ... ??? )
               [001190] -A---------                         *  ASG       int
               [001188] D------N---                         +--*  LCL_VAR   int    V98 tmp78
               [001189] -----------                         \--*  LCL_FLD   int    V11 loc10        [+0]

but the STMT00191 and [001190] and related nodes do not appear in the dump before this location. It would be helpful to dump them when they are created so it's more obvious to see who is creating them and why.

Makes sense, I will make sure to dump these when they are created. FWIW, physical promotion just uses fgInsertStmtNearEnd, so it seems that helper is lacking handling for these blocks. I'll fix it.

As mentioned above, this doesn't fail in main. Has it been intentionally fixed, or did IR perturbations make it "go away"?

#85105 may have altered the IR as a side effect.