cmd/compile: optimize away select receive cases with known-zero chans

[bradfitz]

Go version

go version go1.26-devel_6dceff8bad Thu Sep 25 09:24:46 2025 -0700 darwin/arm64

What did you see happen?

I'd hope Go would notice this select is known to always be nil and optimize it away in the select:

package main

type T struct {
	SomeSymbol int
}

var v any

func main() {
	var c <-chan T
	select {
	case v = <-c:
	}
}

Instead, go tool compile -S x.go says:

main.main STEXT size=128 args=0x0 locals=0x28 funcid=0x0 align=0x0
        0x0000 00000 (/Users/bradfitz/hack/go/x.go:9)   TEXT    main.main(SB), ABIInternal, $48-0
        0x0000 00000 (/Users/bradfitz/hack/go/x.go:9)   MOVD    16(g), R16
        0x0004 00004 (/Users/bradfitz/hack/go/x.go:9)   PCDATA  $0, $-2
        0x0004 00004 (/Users/bradfitz/hack/go/x.go:9)   CMP     R16, RSP
        0x0008 00008 (/Users/bradfitz/hack/go/x.go:9)   BLS     108
        0x000c 00012 (/Users/bradfitz/hack/go/x.go:9)   PCDATA  $0, $-1
        0x000c 00012 (/Users/bradfitz/hack/go/x.go:9)   MOVD.W  R30, -48(RSP)
        0x0010 00016 (/Users/bradfitz/hack/go/x.go:9)   MOVD    R29, -8(RSP)
        0x0014 00020 (/Users/bradfitz/hack/go/x.go:9)   SUB     $8, RSP, R29
        0x0018 00024 (/Users/bradfitz/hack/go/x.go:9)   FUNCDATA        $0, gclocals·g5+hNtRBP6YXNjfog7aZjQ==(SB)
        0x0018 00024 (/Users/bradfitz/hack/go/x.go:9)   FUNCDATA        $1, gclocals·g5+hNtRBP6YXNjfog7aZjQ==(SB)
        0x0018 00024 (/Users/bradfitz/hack/go/x.go:12)  MOVD    ZR, R0
        0x001c 00028 (/Users/bradfitz/hack/go/x.go:12)  MOVD    $main..autotmp_3-8(SP), R1
        0x0020 00032 (/Users/bradfitz/hack/go/x.go:12)  PCDATA  $1, $0
        0x0020 00032 (/Users/bradfitz/hack/go/x.go:12)  CALL    runtime.chanrecv2(SB)
        0x0024 00036 (/Users/bradfitz/hack/go/x.go:12)  MOVD    main..autotmp_3-8(SP), R0
        0x0028 00040 (/Users/bradfitz/hack/go/x.go:12)  CALL    runtime.convT64(SB)
        0x002c 00044 (/Users/bradfitz/hack/go/x.go:12)  MOVD    $type:<unlinkable>.T(SB), R2
        0x0034 00052 (/Users/bradfitz/hack/go/x.go:12)  PCDATA  $0, $-3
        0x0034 00052 (/Users/bradfitz/hack/go/x.go:12)  MOVD    R2, main.v(SB)
        0x003c 00060 (/Users/bradfitz/hack/go/x.go:12)  PCDATA  $0, $-4
        0x003c 00060 (/Users/bradfitz/hack/go/x.go:12)  MOVWU   runtime.writeBarrier(SB), R2
        0x0044 00068 (/Users/bradfitz/hack/go/x.go:12)  PCDATA  $0, $-1
        0x0044 00068 (/Users/bradfitz/hack/go/x.go:12)  PCDATA  $0, $-2
        0x0044 00068 (/Users/bradfitz/hack/go/x.go:12)  CBZW    R2, 88
        0x0048 00072 (/Users/bradfitz/hack/go/x.go:12)  MOVD    main.v+8(SB), R1
        0x0050 00080 (/Users/bradfitz/hack/go/x.go:12)  CALL    runtime.gcWriteBarrier2(SB)
        0x0054 00084 (/Users/bradfitz/hack/go/x.go:12)  STP     (R0, R1), (R25)
        0x0058 00088 (/Users/bradfitz/hack/go/x.go:12)  MOVD    R0, main.v+8(SB)
        0x0060 00096 (/Users/bradfitz/hack/go/x.go:14)  PCDATA  $0, $-1
        0x0060 00096 (/Users/bradfitz/hack/go/x.go:14)  MOVD    -8(RSP), R29
        0x0064 00100 (/Users/bradfitz/hack/go/x.go:14)  MOVD.P  48(RSP), R30
        0x0068 00104 (/Users/bradfitz/hack/go/x.go:14)  RET     (R30)
        0x006c 00108 (/Users/bradfitz/hack/go/x.go:14)  NOP
        0x006c 00108 (/Users/bradfitz/hack/go/x.go:9)   PCDATA  $1, $-1
        0x006c 00108 (/Users/bradfitz/hack/go/x.go:9)   PCDATA  $0, $-2
        0x006c 00108 (/Users/bradfitz/hack/go/x.go:9)   MOVD    R30, R3
        0x0070 00112 (/Users/bradfitz/hack/go/x.go:9)   CALL    runtime.morestack_noctxt(SB)
        0x0074 00116 (/Users/bradfitz/hack/go/x.go:9)   PCDATA  $0, $-1
        0x0074 00116 (/Users/bradfitz/hack/go/x.go:9)   JMP     0

And the T type needlessly shows up in my binary:

% ./bin/go build -ldflags=-dumpdep x.go 2>&1 | grep SomeSymbol
type:main.T <UsedInIface> -> type:.namedata.SomeSymbol.

While this may look contrived, this ends up happening in real programs that use build tags to set boolean constants that guard conditional initialization of chans to receive on in selects.

/cc @randall77 @dr2chase @adonovan @mdempsky

What did you expect to see?

0x0018 00024 (/Users/bradfitz/hack/go/x.go:10)  CALL    runtime.block(SB)

... instead of runtime.chanrecv2, as if you commented out at select receive case and it were just select{}.

And I shouldn't see type info for T or SomeSymbol in my binary.

[gabyhelp]

Related Issues

• cmd/compile/internal/walk: unreachable branch generated for select cases #50823 (closed)
• cmd/compile: elimination of unused runtime calls #17216
• cmd/compile: eliminate more bounds checks #17370
• Go compiler should be more intelligent #20309 (closed)
• cmd/compile: should delete loops without side-effects #69889
• cmd/compile: eliminate some bounds checks #5364 (closed)
• cmd/compile: double zeroing and unnecessary copying/stack use #67957 (closed)
• cmd/compile: incorrect optimizations. #30339 (closed)
• cmd/compile: unneccesary 0 check in division #45928 (closed)
• cmd/compile: pointless static copy code #17113 (closed)

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[mdempsky]

Select desugaring into runtime calls happens before SSA, so might be difficult to optimize this in SSA. The desugaring could be extended to handle the case you presented, but I’m guessing in your real use case the code is somewhat more complex? Could you share a more realistic snippet of code that you’d like to get optimized this way?

[bradfitz]

The code that motivated this discussion was https://github.com/tailscale/tailscale/pull/17272/files#diff-314ba0d799f70c8998940903efb541e511f352b39a9eeeae8d475c921d66c2acR617 and I alluded to the more complex code above, but it's basically just:

package main

type T struct {
	SomeSymbol int
}

const someFalseConstant = false

var v any

func main() {
	var c <-chan T
	if someFalseConstant {
		c = getPubSubChan[T]()
	}
	select {
	case v = <-c:
	}
}

// getPubSubChan is something in another package that doesn't know about T, but uses generics
// and reflect and reflect.TypeFor[T] etc.
func getPubSubChan[T any]() chan T { return make(chan T) }

[bradfitz]

Select desugaring into runtime calls happens before SSA

I was afraid that would be the state of things.

[JunyangShao]

@golang/compiler

[prattmic]

To be clear, I assume your motivation here is about binary size of including the type in the binary and extra instructions in the select. As opposed to the runtime cost of chanrecv vs block?

[bradfitz]

Correct.

[mdempsky]

@bradfitz Thanks for the link to real code.

I think that would be feasible to optimize actually. We already have code to recognize variables that are never reassigned like this.