JIT: support dead block elimination for EH handlers

[BruceForstall]

fgRemoveDeadBlocks() removes unreachable code by starting from a set of initial blocks and walking control-flow paths until all reachable blocks are marked, then removing any unmarked blocks.

However, this will not remove any EH handlers (finally/fault/catch/filter/filter-handler) even if the associated try is not reachable.

Change the logic to only add the EH handler entries corresponding to a try to the reachable set once a reachable try body is found.

This will also require ensuring the related EH table entry is removed if necessary.

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

Issue Details

---

fgRemoveDeadBlocks() removes unreachable code by starting from a set of initial blocks and walking control-flow paths until all reachable blocks are marked, then removing any unmarked blocks.

However, this will not remove any EH handlers (finally/fault/catch/filter/filter-handler) even if the associated try is not reachable.

Change the logic to only add the EH handler entries corresponding to a try to the reachable set once a reachable try body is found.

This will also require ensuring the related EH table entry is removed if necessary.

Author:	BruceForstall
Assignees:	BruceForstall
Labels:	area-CodeGen-coreclr
Milestone:	8.0.0

[EgorBo]

Example:

int Test() => 1;

void Foo()
{
    if (Test() != 1)
    {
        // unreachable
        try
        {
            Console.WriteLine();
        }
        catch
        {
        }
    }
}

it's expected for Foo to be lowered to no-op, but instead, it emits:

; Assembly listing for method Program:Foo():this (FullOpts)
G_M52879_IG01:
       push     rbp
       sub      rsp, 16
       lea      rbp, [rsp+10H]
       mov      qword ptr [rbp-10H], rsp
						;; size=14 bbWeight=1 PerfScore 2.75
G_M52879_IG02:
       add      rsp, 16
       pop      rbp
       ret      
						;; size=6 bbWeight=1 PerfScore 1.75
G_M52879_IG03:
       int3     
						;; size=1 bbWeight=0 PerfScore 0.00
G_M52879_IG04:
       push     rbp
       sub      rsp, 16
       mov      rbp, qword ptr [rcx]
       mov      qword ptr [rsp], rbp
       lea      rbp, [rbp+10H]
						;; size=16 bbWeight=0 PerfScore 0.00
G_M52879_IG05:
       lea      rax, G_M52879_IG02
						;; size=7 bbWeight=0 PerfScore 0.00
G_M52879_IG06:
       add      rsp, 16
       pop      rbp
       ret      
; Total bytes of code 50

[saucecontrol]

I frequently use a pattern where I use exception filters with a JIT-time constant value. These are unmanaged callbacks, which SEH unwinds properly on Windows without the catch, so I want to catch only on Linux. The filter makes the catch unreachable on Windows, but the EH block stays (actually I get 2 -- one for the filtered case and one without)

static readonly bool capture = !OperatingSystem.IsWindows();

static int foo(int x)
{
    int res = 0;
    try {
        res = checked(x * x);
    }
    catch when (capture) {
    }
    
    return res;
}

Which compiles to:

       push     rbp
       sub      rsp, 16
       lea      rbp, [rsp+0x10]
       mov      qword ptr [rbp-0x10], rsp
						;; size=14 bbWeight=1 PerfScore 2.75
       xor      eax, eax
       mov      dword ptr [rbp-0x04], eax
						;; size=5 bbWeight=1 PerfScore 1.25
       imul     edi, edi
       jo       SHORT G_M23486_IG04
       mov      dword ptr [rbp-0x04], edi
       jmp      SHORT G_M23486_IG05
						;; size=10 bbWeight=1 PerfScore 6.00
G_M23486_IG04:
       call     CORINFO_HELP_OVERFLOW
       int3     
						;; size=6 bbWeight=0 PerfScore 0.00
G_M23486_IG05:
       mov      eax, dword ptr [rbp-0x04]
						;; size=3 bbWeight=1 PerfScore 1.00
       add      rsp, 16
       pop      rbp
       ret      
						;; size=6 bbWeight=1 PerfScore 1.75
       push     rbp
       sub      rsp, 16
       mov      rbp, qword ptr [rdi]
       mov      qword ptr [rsp], rbp
       lea      rbp, [rbp+0x10]
						;; size=16 bbWeight=0 PerfScore 0.00
       xor      eax, eax
						;; size=2 bbWeight=0 PerfScore 0.00
       add      rsp, 16
       pop      rbp
       ret      
						;; size=6 bbWeight=0 PerfScore 0.00
       push     rbp
       sub      rsp, 16
       mov      rbp, qword ptr [rdi]
       mov      qword ptr [rsp], rbp
       lea      rbp, [rbp+0x10]
						;; size=16 bbWeight=0 PerfScore 0.00
       lea      rax, G_M23486_IG05
						;; size=7 bbWeight=0 PerfScore 0.00
       add      rsp, 16
       pop      rbp
       ret      
						;; size=6 bbWeight=0 PerfScore 0.00
; Total bytes of code 97

[BruceForstall]

This is a slightly different case of "dead EH clause": the filter-handler (aka "catch") is dead because the filter always returns zero. Since the filter always returns zero, it's unnecessary.

[AndyAyersMS]

Egor's example is now cleaned up:

; Method Program:<<Main>$>g__Foo|0_1() (FullOpts)
G_M38873_IG01:  ;; offset=0x0000
						;; size=0 bbWeight=1 PerfScore 0.00

G_M38873_IG02:  ;; offset=0x0000
       ret      
						;; size=1 bbWeight=1 PerfScore 1.00
; Total bytes of code: 1

but the "known filter value" is not yet a thing... guessing it is pretty rare (since filters are rare).

If an EH region is discovered unreachable late in the phase list (say after the main opt phases) we might not be able to remove it, especially so after funclets have been created.