Don't update side-effects in a statement that doesn't exist

[EgorBo]

Fixes #113334

There were two issues:

1. fgMorphBlock could fold the entire upper bound check into no-op meaning there are no statements to update side-effects for
2. (unrelated to the bug) - we shouldn't mix different bounds checks nodes with different throw kind

IR for lower-bound block and then upper-bound block:

------------ BB04 [0003] [012..012) -> BB06(0),BB05(1) (cond), preds={BB01} succs={BB05,BB06}

***** BB04 [0003]
STMT00006 ( 0x012[E-] ... ??? )
N004 (  5,  5) [000136] -----------                         *  JTRUE     void
N003 (  3,  3) [000135] J----------                         \--*  GT        int
N001 (  1,  1) [000133] -----------                            +--*  CNS_INT   int    0
N002 (  1,  1) [000134] -----------                            \--*  CNS_INT   int    0 $40

------------ BB05 [0004] [012..012) -> BB02(1),BB06(0) (cond), preds={BB04} succs={BB06,BB02}

***** BB05 [0004]
STMT00007 ( 0x012[E-] ... ??? )
N004 (  5,  5) [000139] -----------                         *  JTRUE     void
N003 (  3,  3) [000138] J----------                         \--*  LT        int
N001 (  1,  1) [000137] -----------                            +--*  CNS_INT   int    0
N002 (  1,  1) [000131] -----------                            \--*  CNS_INT   int    1 $43

[jakobbotsch]

I wonder -- do we gain a lot from calling morph over removing the bounds checks manually?

[EgorBo]

I wonder -- do we gain a lot from calling morph over removing the bounds checks manually?

@jakobbotsch Yes, it's needed to handle constant indices - there are no phases to fold JTRUE(1 > 0) after it (for upper bound check). I can fold it by hands if we want to avoid morph.. that will just be more verbose.

I just didn't expect the lower bound check can be also foldable - seems to happen only around bounds checks with SIMD.

[Copilot]

Pull Request Overview

This pull request addresses a bug where side-effects were not updated due to an entirely folded upper-bound check and also separates bounds check nodes by throw kind.

• Fix the issue in fgMorphBlock where no statements are updated when the upper-bound check is a no-op.
• Prevent mixing of different bounds check nodes with different throw kinds.
• Introduce a regression test to verify the fix.