build: add link-time dead method elimination

[luoliwoshang]

Summary

• Add package metadata summaries for DCE facts, including ordinary edges, type children, interface method demands, interface conversions, reflection method use, and ABI method slots.
• Persist metadata in the package build cache and merge package-local summaries into a global metadata view for analysis.
• Add a metadata-driven deadcode analyzer that computes live ABI method slots using reachable interface demands, used-in-interface types, reflection rules, and MethodByName rules.
• Emit strong ABI type metadata overrides that clear dead method IFn/TFn references, allowing the linker to drop unused method bodies.
• Add metadata golden tests, metadump tooling, and focused analyzer/DCE override coverage.

Notes

• The temporary verbose cache-hit meta read timing output has been removed; -v cache hit logging now prints only the package path.

Test Plan

• go test ./internal/deadcode ./internal/dcepass ./internal/metadata ./internal/build -count=1

Demo Size Comparison

Built _demo/go demos with DCE from this PR into _demo/go/520-dce, and with local upstream/main (ec1b2f74) into _demo/go/520-nodce.

Build command shape:

• LLGO_BUILD_CACHE=off llgo build -a -o <output> .

Summary for demos that built successfully on both sides:

• Matched demos: 55
• No-DCE total size: 59,670,624 bytes
• DCE total size: 43,519,232 bytes
• Total saved: 16,151,392 bytes
• Total reduction: 27.07%
• Demos that grew: 0

Two demos failed on both sides and were excluded from the size total:

• defer/setjmp: C source files are present but this direct llgo build . path is not using cgo/SWIG.
• defer/test: existing source errors around c.Printf / c.Str.

All measured demos, sorted by saved bytes descending:

Demo	No DCE	DCE	Saved	Saved %
goimporter-1389	4267280	2875648	1391632	32.61%
embedunexport-1598	3056032	1762912	1293120	42.31%
gotypes	3072896	2458912	613984	19.98%
abimethod	1434896	912224	522672	36.43%
mimeheader	1527664	1005856	521808	34.16%
reflectpointerto	1719408	1236656	482752	28.08%
reflectmake	1905136	1443568	461568	24.23%
netip	1504976	1048512	456464	30.33%
gobuild-1389	1590144	1158624	431520	27.14%
randdemo	1415280	985376	429904	30.38%
gotoken	1433408	1003808	429600	29.97%
sysexec	1414528	985376	429152	30.34%
oslookpath	1414144	985008	429136	30.35%
sync	1419840	990800	429040	30.22%
createtemp-1654	1431072	1002992	428080	29.91%
gobuild	2199920	1776784	423136	19.23%
randcrypt	1557968	1143248	414720	26.62%
maphash	1576416	1162128	414288	26.28%
failed/stacktrace	1376288	963392	412896	30.00%
cgo	1376928	964128	412800	29.98%
readdir	1395248	982512	412736	29.58%
timedur	1375904	963440	412464	29.98%
mkdirdemo	1394336	981920	412416	29.58%
sysopen-1654	1394208	981840	412368	29.58%
reflectfunc	1396240	984240	412000	29.51%
reflectcopy	1413152	1002256	410896	29.08%
logdemo	1377216	980864	396352	28.78%
checkfile	1377344	981264	396080	28.76%
reflectname-1412	1376032	979968	396064	28.78%
oswritestring	1376704	981216	395488	28.73%
texttemplate	2172768	1929248	243520	11.21%
osfile	788480	554368	234112	29.69%
commandrun	1160880	938496	222384	19.16%
reflectindirect	687056	487888	199168	28.99%
reflectmethod	1716672	1543392	173280	10.09%
timer	508304	419856	88448	17.40%
async	319600	277024	42576	13.32%
gotime	354480	328496	25984	7.33%
syscall	173696	151856	21840	12.57%
ifaceconv	74736	72896	1840	2.46%
ifaceprom-1559	91120	89312	1808	1.98%
mapclosure	78416	77136	1280	1.63%
export	95072	93856	1216	1.28%
cabi	70592	69472	1120	1.59%
issue1538-floatcvtuint-over	70592	69472	1120	1.59%
defer	72384	71280	1104	1.53%
defer/stress	72320	71216	1104	1.53%
issue1538	70512	69408	1104	1.57%
complex	73440	72352	1088	1.48%
defer/setjmp/c_standard_demo/closure_test	72464	71376	1088	1.50%
math	73248	72176	1072	1.46%
defer/setjmp/c_standard_demo/llgo_test	70816	70816	0	0.00%
goroutine	71040	71040	0	0.00%
runtime	89760	89760	0	0.00%
statefn	71568	71568	0	0.00%

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 76.80328% with 283 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/metadata/decode.go	48.54%	53 Missing and 53 partials ⚠️
internal/dcepass/dcepass.go	59.71%	57 Missing and 28 partials ⚠️
internal/metadata/global_summary.go	77.38%	24 Missing and 21 partials ⚠️
internal/metadata/meta.go	67.56%	14 Missing and 10 partials ⚠️
internal/metadata/builder.go	84.84%	5 Missing and 5 partials ⚠️
internal/deadcode/analyze.go	95.91%	3 Missing and 3 partials ⚠️
internal/metadata/format.go	93.97%	3 Missing and 2 partials ⚠️
internal/metadata/encode.go	98.79%	0 Missing and 1 partial ⚠️
ssa/abitype.go	98.66%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[luoliwoshang]

补充一个 DCE 方法槽修复点：

对于被判定为 dead 的方法槽，IFn/TFn 不应该写成 null。这里需要和 Go linker 的做法一致，重定向到一个专门的 runtime.unreachableMethod。

原因是运行时构造 itab 时，fun[0] == 0 有特殊含义：表示该类型没有实现目标接口。如果 DCE 把匹配到的方法函数指针清成 0，类型方法签名虽然仍然匹配，但生成出来的 itab 会被当成“不完整实现”，不会进入 itab cache。

这会导致同一个 concrete type 到同一个接口的转换生成多个不同 itab。典型表现是 reflect.Type 作为 map key 时出现不一致：

t0 == t1 // true
m[t1]    // false

最终会影响标准库里这类逻辑，例如 testing.F.Add(int) 中的：

supportedTypes[reflect.TypeOf(5)]

所以现在的策略是：dead method slot 仍保留非零函数入口，指向 runtime.unreachableMethod。如果该方法真的在运行时被调用，会像 Go 一样明确 throw；同时 itab 的实现判断和缓存语义不会被破坏。

[luoliwoshang]

Additional experiment on _demo/go: temporarily degraded the DCE interface-method rule to only match reachable interface method demands by method name + method type (MethodSig), without checking whether the concrete type fully implements the demanded interface. Metadata collection and cache inputs were kept unchanged; only the analyzer matching rule was changed for this experiment.

Summary over the same demo set:

matched demos:          55
failed demos:           2 (defer/setjmp, defer/test)
No DCE total:           59,670,624 bytes
Precise DCE total:      43,519,232 bytes
Name+Sig DCE total:     43,602,096 bytes

Precise DCE saved:      16,151,392 bytes (27.07%)
Name+Sig DCE saved:     16,068,528 bytes (26.93%)
Extra vs precise DCE:   82,864 bytes
Lost precise saving:    0.51% of the precise-DCE saved bytes

Top cases where name+sig keeps more than the precise rule:

Demo	Precise DCE	Name+Sig DCE	Extra vs precise
gobuild	1,776,784	1,795,344	18,560
timedur	963,440	981,184	17,744
reflectmake	1,443,568	1,447,424	3,856
abimethod	912,224	915,056	2,832
mimeheader	1,005,856	1,008,192	2,336
goimporter-1389	2,875,648	2,877,904	2,256
gotypes	2,458,912	2,460,928	2,016
embedunexport-1598	1,762,912	1,764,816	1,904

Conclusion from this demo set: the more precise interface implementation check does reduce output size, but the measured delta is small here: about 83KB total, or 0.51% of the bytes saved by precise DCE.