[angerman]

Description

musl libc implements memcmp (src/string/memcmp.c) as a portable 1-byte scalar loop:

for (; n && *l == *r; n--, l++, r++);
return n ? *l - *r : 0;

In static musl Linux builds (hydraJobs.*.musl.*, projectCross.musl64, projectCross.aarch64-multiplatform-musl), large Data.ByteString, ShortByteString, Data.Text, and ByteArray# (compareByteArrays# / MO_Memcmp) comparisons run ~21× slower than native glibc builds that use SIMD memcmp, while small fixed-size MO_Memcmp comparisons incur an out-of-line C call.

Until haskell.nix is bumped past ican-group/haskell.nix#2640 (upstream GHC MR: https://gitlab.haskell.org/ghc/ghc/-/merge_requests/16986; see also ican-group/haskell.nix#2639), this PR applies ghc-9.6.7-simd-memcmp.patch to haskell-nix.compiler.ghc967 on musl targets in nix/pkgs.nix:

• Inline small constant-size MO_Memcmp in X86_64 & AArch64 NCGs (0..64 bytes): Unrolls constant-size MO_Memcmp inline in GHC.CmmToAsm.X86.CodeGen (XOR for 0; MOVZxL + SUB for 1; branchless load + BSWAP + CMP + SETCC + SUB for 2, 4, 8; equality scan + shared cold mismatch block with X86 CFG edge updates for 3, 5..7, 9..64) and GHC.CmmToAsm.AArch64.{CodeGen,Instr,Ppr} (LDR/REV/CMP/CSET/SUB), gated by ncgInlineThresholdMemcpy (-fmax-inline-memcpy-insns, default 32, preserving the 9.6.7 DynFlags/NCGConfig ABI). All loads stay strictly within [0, n).
• GHC 9.6.7 RTS (hs_memcmp + static MuslMemcmp.o): Exports hs_memcmp in libHSrts with SSE2 + runtime CPUID/XGETBV AVX2 (4×32B loop + vzeroupper) on x86_64 non-glibc/non-Bionic/non-uClibc Linux, ARMv8-A NEON (64B loop) on little-endian aarch64 non-glibc/non-Bionic/non-uClibc Linux, a bounded 64-bit word fallback on other non-glibc Linux architectures, and an n == 0-guarded delegation to platform memcmp elsewhere. All loads stay strictly within [0, n). Static RTS archives (!DYNAMIC && !SHARED) also provide MuslMemcmp.o (memcmp -> hs_memcmp) before -lc so C FFI callers resolve to hs_memcmp without extracting scalar libc.a(memcmp.lo).
• Out-of-line MO_Memcmp & Stage 1+ bytestring: Lowers dynamic/large MO_Memcmp (compareByteArrays#) to hs_memcmp across NCG and LLVM backends (while preserving inline 0 for n == 0 and "memcmp" for constant 1..32 in LLVM so SimplifyLibCalls unrolls them), and compiles Stage 1+ libraries/bytestring with -DHS_MEMCMP_IN_RTS so Data.ByteString calls hs_memcmp directly while Stage 0 bootstraps cleanly against the unpatched host RTS.
• Scoped to musl targets & content-addressed: Applied via lib.optionalAttrs final.stdenv.targetPlatform.isMusl on top of the existing haskell.nix 9.6.7 patch stack (12 patches on x86_64-musl, 16 patches on aarch64-musl), wrapped in builtins.path so unrelated repository commits do not invalidate the musl compiler store path. Stage 0 bootstrap compilers and all native Linux, macOS, and Windows (ucrt64) derivations and devShells remain bit-for-bit hash-identical.

Verification

1. Derivation invariance (noGitRev.drvPath):
   • hydraJobs.x86_64-linux.native.cardano-node.passthru.noGitRev.drvPath and packages.aarch64-darwin.cardano-node.passthru.noGitRev.drvPath are unchanged before and after this commit.
   • Patch-stack application verified cleanly against ghc-9.6.7 for both pkgsCross.musl64 (13 patches total) and pkgsCross.aarch64-multiplatform-musl (17 patches total).
2. NCG inline MO_Memcmp & end-to-end static musl build (x86_64 + aarch64):
   • Verified X86_64 (-dasm-lint enabled) and AArch64 NCG inline MO_Memcmp lowering and live execution across constant sizes 0..65, unaligned offsets, high-bit unsigned bytes, and back-to-back calls in the same Cmm basic block.
   • Built hydraJobs.x86_64-linux.musl.cardano-node (cardano-node 11.2.1 - linux-x86_64 - ghc-9.6, statically linked) and hydraJobs.x86_64-linux.musl.cardano-cli (cardano-cli 11.7.0.0), confirming hs_memcmp -> hs_memcmp_simd and memcmp -> hs_memcmp in the resulting static ELF binaries.
3. Microbenchmark (x86_64-linux, AMD Ryzen 5 3600, single core, stock unpatched musl):

Compiler (GHC 9.6.7 -O2, 13,466 × 64 KiB comparisons)	Data.ByteString	ShortByteString / compareByteArrays#
Native glibc	14.061 ms (1.00×)	13.779 ms (1.00×)
Unpatched static musl	298.485 ms (21.23× slower)	299.354 ms (21.73× slower)
Patched static musl (this PR)	14.147 ms (21.10× speedup, 1.01× glibc)	13.834 ms (21.64× speedup, 1.00× glibc)

Checklist

• Commit sequence broadly makes sense and commits have useful messages
• New tests are added if needed and existing tests are updated
• Any changes are noted in the CHANGELOG.md for affected package (N/A: Nix toolchain patch only)
• The version bounds in .cabal files are updated (N/A)
• CI passes
• Self-reviewed the diff