[forceinline] add Kunpeng forceinline optimization

[lxq015]

[port] forceinline: add Kunpeng forceinline optimization for runtime hot paths

Port of https://gitcode.com/openeuler/golang/pull/66 to the ZTE RISC-V toolchain.

Summary

Adds a Kunpeng-targeted forceinline pass that forcibly inlines selected hot functions in runtime.mallocgc and related allocation/GC paths, bypassing the normal inline cost budget. The feature is fully gated by a new compiler debug flag and is off by default, so existing builds are unaffected.

Two new -d flags are introduced:

• -d=forceinline=1 — enable the default force-inline function list
• -d=forceinlinelog — print force-inline decision logs (for diagnostics)

Usage

./make.bash
GOMAXPROCS=1 go test -gcflags='all=-d=forceinline=1' -bench=BenchmarkMalloc -run=^$ -count=3 runtime/

Verification

1. Functional / correctness

Ran the bundled script test to verify the gating semantics:

Case Result
-d=forceinline=1 default list takes effect ✅
-d=forceinline=0 disables the feature ✅
-d=forceinline=2 rejected with -forceinline does not support setting to 2 ✅
-l has higher priority than forceinline ✅
Binary equivalence (two identical builds produce identical binaries) ✅

2. Benchmark

goos: linux
goarch: riscv64
pkg: runtime
cpu: Spacemit(R) X60 
                 │before_noforceinline │  forceinline_withmallocgc  │ forceinline_nomallocgc │
                  │       sec/op        │   sec/op    vs base         │   sec/op     vs base          │
Malloc8            101.1n ± 7%   104.3n ± 9%       ~ (p=0.485)         101.8n ± 3%       ~ (p=0.900)
Malloc16           178.8n ± 3%   171.4n ± 1%  -4.17% (p=0.002)         178.1n ± 1%  -0.42% (p=0.006)
Malloc32           214.9n ± 2%   211.3n ± 0%  -1.70% (p=0.002)         215.6n ± 1%       ~ (p=0.818)
MallocTypeInfo8    152.4n ± 2%   143.6n ± 1%  -5.81% (p=0.002)         152.7n ± 7%       ~ (p=0.818)
MallocTypeInfo16   198.2n ± 9%   211.0n ± 3%       ~ (p=0.065)         209.3n ± 5%       ~ (p=0.132)
MallocTypeInfo32   242.3n ± 2%   231.1n ± 3%  -4.62% (p=0.002)         236.5n ± 2%  -2.37% (p=0.009)
MallocLargeStruct  1.433µ ± 1%   1.380µ ± 1%  -3.70% (p=0.002)         1.423µ ± 2%       ~ (p=0.329)
geomean            236.0n        232.3n      -1.56%                    237.1n       +0.46%

The default forceinline set (which includes mallocgc) shows a -1.56% improvement relative to the baseline, with 5 statistically significant items (p=0.002). Excluding mallocgc from the forceinline set results in a +0.46% regression relative to the baseline.

Notes

The default force-inline function list was adapted for the Go 1.26.5 runtime layout (runtime.scanobject → runtime.scanObject; removed writeHeapBits.*, mallocgc1/2, deductAssistCredit2, which no longer exist in this version). See the in-code comment: "You may need update this list when you update the go version."

[wangpc-pp]

+2.

[BoyaoWang430]

Add concurrent:"ok" here. Without this tag, the compiler will set ConcurrentOk=false and force the backend worker count to 1, which will significantly increase build time.

[lxq015]

Thank you, fixed.

[BoyaoWang430]

+1

[ctk-1998]

+1

[ctk-1998]

Given that Go versions are subject to upgrades, the function list here may require updates accordingly; a means of recording this pending action is needed.

[lxq015]

Thank you for your suggestion. I've added assertion-based tests (TestScript/forceinline) that verify every function in the force-inline list is actually force-inlined. This acts as a safety net: if we upgrade the Go version but forget to adapt the list, the test will fail and all.bash will catch it.

[ctk-1998]

Do we have any performance data?

[ctk-1998]

IIUC, runtime.mallocgc has a high cost. If we force it to be inlined by default, it would significantly increase pressure of i-cache, resulting in performance degradation rather than improvement. Has this function been tested independently?

[lxq015]

Yes, the default forceinline set (which includes mallocgc) improves performance by 1.56% over the baseline, with 5 statistically significant results (p=0.002). Excluding mallocgc from the forceinline set leads to a 0.46% regression. Benchmark results have also been added to the PR description.

[newborn22]

+2