JIT does not eliminate null check for "o ??= new()"

[tfenise]

Description

https://sharplab.io/#v2:EYLgxg9gTgpgtADwGwBYA0AXEBDAzgWwB8ABAJgEYBYAKGIGYACMhgYQYG8aHunHjykTFAwCy5ABQRgAKxhgMDCAEoOXHuoCWAMwbjJDAPwGAvAwB2MAO6KZcjOKUqNucwFcANu5UYAFlAjWFtYAoghgMAAOGBoQZg4A3GrcAL5JDGn0TAJCoqSStvKKKpzU6po6ehAMpkE2svIOTi5mHl4Mvv6BVgyh4VExcUqJpTypI9wZfNnEwiJ04hpmChrFaeW6GgwAPAwADE7Ve8NlPNob23vefgHm3b2R0bEJaWPqk1mCM6Io+fUKyqpxgwxskgA=

static void M1(object o) {
    if ((o ??= new object()) is null) throw new Exception();
}
    
static void M2(object o) {
    if ((o = new object()) is null) throw new Exception();
}

static void M3(int i) {
    if (i < 0) i = 0;
    if (i < 0) throw new Exception();
}
    
static void M4(object o) {
}

While the JIT can optimize M2(almost, but still curiously worse than M4) and M3 to a nop, it cannot optimize M1 similarly.

Found while discussing #75936.

Configuration

sharplab says "Core CLR 6.0.822.36306 on amd64".

I also tried 7.0.100-rc.1.22431.12 on amd64. The codegen for M1 was not really better.

Regression?

Probably not.

[ghost]

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

Issue Details

---

Description

https://sharplab.io/#v2:EYLgxg9gTgpgtADwGwBYA0AXEBDAzgWwB8ABAJgEYBYAKGIGYACMhgYQYG8aHunHjykTFAwCy5ABQRgAKxhgMDCAEoOXHuoCWAMwbjJDAPwGAvAwB2MAO6KZcjOKUqNucwFcANu5UYAFlAjWFtYAoghgMAAOGBoQZg4A3GrcAL5JDGn0TAJCoqSStvKKKpzU6po6ehAMpkE2svIOTi5mHl4Mvv6BVgyh4VExcUqJpTypI9wZfNnEwiJ04hpmChrFaeW6GgwAPAwADE7Ve8NlPNob23vefgHm3b2R0bEJaWPqk1mCM6Io+fUKyqpxgwxskgA=

static void M1(object o) {
    if ((o ??= new object()) is null) throw new Exception();
}
    
static void M2(object o) {
    if ((o = new object()) is null) throw new Exception();
}

static void M3(int i) {
    if (i < 0) i = 0;
    if (i < 0) throw new Exception();
}
    
static void M4(object o) {
}

While the JIT can optimize M2(almost, but still curiously worse than M4) and M3 to a nop, it cannot optimize M1 similarly.

Found while discussing #75936.

Configuration

sharplab says "Core CLR 6.0.822.36306 on amd64".

I also tried 7.0.100-rc.1.22431.12 on amd64. The codegen for M1 was not really better.

Regression?

Probably not.

Author:	tfenise
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr
Milestone:	-

[EgorBo]

Looks like Jump-Threading + PHI issue?

BB03 is never true because V02 is either come form V02 != 0 == true or it comes from BB02 where it's set to non-null
cc @AndyAyersMS to confirm

[AndyAyersMS]

Likely BB03 has some side effecting code; this currently blocks jump threading.

[AndyAyersMS]

Likely BB03 has some side effecting code; this currently blocks jump threading.

No, you were right, this is the PHI case: #48115 (comment)

optRedundantRelop in BB03; jump tree is
N004 (  7,  6) [000012] -----------                         *  JTRUE     void   $VN.Void
N003 (  5,  4) [000011] J------N---                         \--*  NE        int    $101
N001 (  3,  2) [000009] -----------                            +--*  LCL_VAR   ref    V02 tmp1         u:2 (last use) $1c0
N002 (  1,  1) [000010] -----------                            \--*  CNS_INT   ref    null $VN.Null
 ... checking previous tree
N005 (  0,  0) [000038] -A------R--                         *  ASG       ref    $VN.Void
N004 (  0,  0) [000036] D------N---                         +--*  LCL_VAR   ref    V02 tmp1         d:2 $VN.Void
N003 (  0,  0) [000037] -----------                         \--*  PHI       ref    $1c0
N001 (  0,  0) [000040] ----------- pred BB02                  +--*  PHI_ARG   ref    V02 tmp1         u:3 $180
N002 (  0,  0) [000039] ----------- pred BB01                  \--*  PHI_ARG   ref    V02 tmp1         u:1 $80
 -- prev tree is a phi
Relop [000011] BB03 value unknown, trying inference