Compiler Regression in Go 1.27.0

“cmd/compile: arm64 "FLDPS ...: constant is not in pool" building a large-offset float-pair load (regression in go1.27.0) ### Go version `go1.27.0 darwin/arm64` (also reproduces with any host GOOS/GOARCH when cross-compiling `GOARCH=arm64`). ### What did you do? Minimal reproducer — compile-only, no run needed: `go.mod` ``` module repro go 1.21 ``` `repro.go` ```go package repro // A large leading region so the trailing float pair lands at a multi-MB offset. type Big struct { _ [28560000]byte // ~27 MB F [2]float32 } var G Big func Load() [2]float32 { return G.F } func Store(v [2]float32) { G.F = v } ``` Build for arm64: ``` GOOS=linux GOARCH=arm64 go build ./... ``` > The `go 1.21` directive is intentional: it lets the identical module be compiled by both go1.26.6 and go1.27.0, isolating the compiler backend as the only variable. The bug is independent of language version — it is a code-generation issue, not a language-feature one. ### What did you expect to see? A clean compile — as with go1.26.6, and as with GOARCH=amd64 on go1.27.0. ### What did you see instead? ``` # repro omovlit add 28560000 (0x1b3ca80) omovlit add 28560000 (0x1b3ca80) <autogenerated>:1: 00000 (<autogenerated>:1) FLDPS 28560000(R1), (F0, F1): constant is not in pool <autogenerated>:1: 00012 (<autogenerated>:1) FLDPS 28560000(R0), (F2, F3): constant is not in pool ``` `go build` exits 1. The arm64 backend materialises the large field offset (`omovlit`) and emits a paired float load `FLDPS`, but the offset constant is not placed in the constant pool, so assembly fails. ### Regression matrix (same reproducer, same source) | toolchain | GOARCH=arm64 | GOARCH=amd64 | |---|---|---| | **go1.27.0** | **FAIL** (rc=1, error above) | ok (rc=0) | | go1.26.6 | ok (rc=0) | ok (rc=0) | So this is **new in go1.27.0** and **arm64-only**. The trigger is the *offset magnitude* — a paired float field (`[2]float32`; `[2]float64`/`complex128` behave analogously) reached at a large offset from the base of an addressable value. Likely area: `cmd/internal/obj/arm64` constant-pool handling (`addpool`/`span7`) for the FLDPS/FSTPS large-displacement path. ### Real-world impact This broke compilation of a package containing a large memory-mapped shared-memory struct (a ~27 MB segment with float fields) that had compiled cleanly through go1.26.x. Because the struct layout is fixed by an external ABI, the only workaround was to stay on go1.26.6. ### Notes - A GitHub issue search for `"constant is not in pool"` returned no existing report. - amd64 is unaffected; the pure-integer path is unaffected (the failure is specific to the paired-float load `FLDPS`). ”