diff --git a/docs/content/docs/cli.mdx b/docs/content/docs/cli.mdx index 8b0353620..9fa5e9447 100644 --- a/docs/content/docs/cli.mdx +++ b/docs/content/docs/cli.mdx @@ -135,6 +135,9 @@ With `--sparse`, the profile leaves out the functions the runs never reached. Th
--detached-exit
On Linux, when the program calls process.exit(), hand the release of its memory to a helper process so the parent sees the exit sooner. The kernel otherwise frees the whole address space before the exit is reported, which takes tens of milliseconds per gigabyte. Opt in only where the trade-offs are acceptable: for up to about 100 ms after the exit, the helper keeps the executable mapped (overwriting the binary in place fails with "text file busy"), its memory stays charged to the cgroup, and the teardown CPU time no longer appears in the parent's resource usage. Valid only for executable builds; it has no effect on other platforms.
+
--target-cpu <baseline|x86-64-v2|x86-64-v3|x86-64-v4|native>
+
On x86-64 targets, compile the program's code for an x86-64 microarchitecture level. The default, baseline, runs on any x86-64 CPU. A higher level lets LLVM use more instructions, and the executable then runs only on CPUs that support that level. native uses the highest level the build machine supports. Valid for executable, assembly, and object output on x86-64 targets. See Target CPU.
+
--profile-generate <pattern>
On Linux, build an instrumented executable for profile-guided optimization. Each run writes a raw profile to the path pattern, where %p expands to the process ID; the LLVM_PROFILE_FILE environment variable overrides the pattern at run time. A relative pattern resolves from the current working directory. Instrumented programs run slower. Valid only for executable builds with --optimization=release or speed.
@@ -228,6 +231,43 @@ $ scriptc build app.ts --optimization=speed --profile-use app.profdata Train on workloads like the ones the program will run: code the training never reached is optimized as without a profile. The profile can be committed and reused as the program changes. Functions whose control flow changed since the profile was collected lose their profile, and the rest keep theirs, so regenerate it when much of the program has changed. [Build parallelism](/docs/how-it-works#build-parallelism) describes what a profile adds to build time. +## Target CPU + +By default, x86-64 executables run on any x86-64 CPU, so the compiler uses only the baseline instruction set (SSE2). `--target-cpu` compiles the program's code for a higher x86-64 microarchitecture level. Each level includes the ones below it: + + + + + + + + + + + + + + + + + + + + + + +
LevelAdds
x86-64-v2SSE3, SSSE3, SSE4.1, SSE4.2, POPCNT
x86-64-v3AVX, AVX2, BMI1, BMI2, F16C, FMA, LZCNT, MOVBE
x86-64-v4AVX-512 (F, BW, CD, DQ, VL)
+ +```console +$ scriptc build app.ts --target-cpu=x86-64-v3 -o app +``` + +Loops over typed arrays and numbers gain the most, because LLVM vectorizes them with wider registers. Some loops run slower at a higher level, so measure the program on the machines it will run on. + +An executable built for a level runs only on CPUs that support that level. On an older CPU, it crashes with an illegal-instruction signal (`SIGILL`) when it reaches an instruction the CPU lacks. A named level is portable to every machine of that level. `native` selects the highest level the build machine supports, so use it only for executables that run on the machine that built them or on identical machines. + +The level applies to the program's own code. The runtime pack keeps its baseline code. Builds for different levels never share cached artifacts, and `baseline` builds are the same as builds without the option. + ## LLVM backend scriptc generates LLVM IR and compiles it to native code. Executable builds link the result with the matching precompiled runtime pack. Use --emit=llvm to inspect the generated code. diff --git a/docs/content/limitations/builtin-apis/filesystem-url-paths.mdx b/docs/content/limitations/builtin-apis/filesystem-url-paths.mdx index d5079ed5a..9b3580647 100644 --- a/docs/content/limitations/builtin-apis/filesystem-url-paths.mdx +++ b/docs/content/limitations/builtin-apis/filesystem-url-paths.mdx @@ -1,3 +1,3 @@ -- **Filesystem URL paths:** Synchronous `fs` functions accept `file:` URL paths and convert them as `fileURLToPath` does. - `existsSync` and the `fs.promises` functions refuse URL paths; pass `fileURLToPath(url)` instead. - A URL path after a string path, or with options that are not literals, is refused, because Node checks those arguments before it converts the URL. +- **Filesystem URL paths:** `fs` and `fs.promises` functions accept `file:` URL paths and convert them as `fileURLToPath` does; `existsSync` answers `false` and the `fs.promises` functions reject for a URL that does not convert. + A URL path with options that are not literals is refused where Node checks those options before it converts the URL, except a `readFileSync` encoding. + `fs.promises` functions also refuse a string path before a URL path; pass `fileURLToPath(url)` instead. diff --git a/docs/content/limitations/untyped-values.mdx b/docs/content/limitations/untyped-values.mdx index 6a881aec3..05245a396 100644 --- a/docs/content/limitations/untyped-values.mdx +++ b/docs/content/limitations/untyped-values.mdx @@ -15,3 +15,6 @@ JavaScript argument and return conversions preserve native array references; oth Supported operations on `unknown` include truthiness, `typeof` narrowing, property access, `+`, `switch`, and `throw`. Cast before using other operations. + +In JavaScript, a call whose untyped argument could select an overload with another result type returns a checked value, whichever overload the type checker picks, and so do the unannotated bindings and functions that hold or return it. +In TypeScript, an `any` argument keeps the overload the type checker picks, and a value that selects another overload at runtime throws a `TypeError` where the result is used. diff --git a/packages/cli/src/bootstrap.ts b/packages/cli/src/bootstrap.ts index 242a7c6f7..f14866a9c 100644 --- a/packages/cli/src/bootstrap.ts +++ b/packages/cli/src/bootstrap.ts @@ -70,6 +70,8 @@ async function tryFastPath(): Promise { values["profile-generate"] !== undefined || values["profile-use"] !== undefined || values.sparse || + // The full compiler resolves `native` against the build host. + values["target-cpu"] !== undefined || (values["external-types"] ?? []).length > 0 ) return null; @@ -200,6 +202,7 @@ async function coverageFastPath( values.emit !== undefined || values.strip || values["detached-exit"] || + values["target-cpu"] !== undefined || values.lib || values.profile !== undefined || values.ffi !== undefined || diff --git a/packages/cli/test/target-cpu.test.ts b/packages/cli/test/target-cpu.test.ts new file mode 100644 index 000000000..ca5243564 --- /dev/null +++ b/packages/cli/test/target-cpu.test.ts @@ -0,0 +1,100 @@ +import { execFile } from "node:child_process"; +import { createRequire } from "node:module"; +import { mkdtemp, readFile, rm, writeFile } from "node:fs/promises"; +import { dirname, join } from "node:path"; +import { promisify } from "node:util"; +import { expect, test } from "vitest"; +import { hostTargetCpu } from "../../compiler/src/backend/target-cpu.js"; + +const execFileAsync = promisify(execFile); +const require = createRequire(import.meta.url); +const repoRoot = join(import.meta.dirname, "../../.."); +const cliEntry = join(repoRoot, "packages/cli/src/main.ts"); +const bootstrapEntry = join(repoRoot, "packages/cli/dist/bootstrap.js"); +const tsxLoader = join(dirname(require.resolve("tsx/package.json")), "dist/loader.mjs"); +const host = hostTargetCpu(); +const runsV3 = host === "x86-64-v3" || host === "x86-64-v4"; + +const program = `const a = new Int32Array(4096); +for (let i = 0; i < a.length; i++) a[i] = (i * 2654435761) | 0; +let s = 0; +for (let r = 0; r < 64; r++) for (let i = 0; i < a.length; i++) s = (s + a[i]!) | 0; +console.log(s); +`; + +test.runIf(process.platform === "linux" && process.arch === "x64")( + "--target-cpu compiles the program for the level without mixing cache variants", + async () => { + const dir = await mkdtemp(join("/tmp", "scriptc-target-cpu-")); + const entry = join(dir, "main.ts"); + const outPath = join(dir, "program"); + const llvmPath = join(dir, "main.ll"); + const env = { ...process.env, SCRIPTC_CACHE_DIR: join(dir, "cache") }; + const build = async (args: string[], bootstrap = false): Promise => { + await execFileAsync( + process.execPath, + [ + ...(bootstrap ? [bootstrapEntry] : ["--import", tsxLoader, cliEntry]), + "build", + entry, + "-o", + outPath, + ...args, + ], + { env, maxBuffer: 4 * 1024 * 1024 }, + ); + return readFile(llvmPath, "utf8"); + }; + try { + await writeFile(entry, program); + const expected = `${(await execFileAsync(process.execPath, [entry])).stdout}`; + const baseline = await build([]); + expect(baseline).not.toContain('"target-cpu"'); + expect(await build(["--target-cpu=baseline"])).toBe(baseline); + const v3 = await build(["--target-cpu=x86-64-v3"]); + expect(v3).toMatch(/^attributes #0 = \{ sanitize_address "target-cpu"="x86-64-v3" \}$/m); + if (runsV3) expect((await execFileAsync(outPath)).stdout).toBe(expected); + // The bootstrap's cached route builds without a level. + expect(await build([], true)).toBe(baseline); + expect((await execFileAsync(outPath)).stdout).toBe(expected); + await expect(build(["--target-cpu=haswell"])).rejects.toThrow(/unknown target CPU "haswell"/); + } finally { + await rm(dir, { recursive: true, force: true }); + } + }, + 240_000, +); + +test("--target-cpu levels are refused for arm64 and WASI targets", async () => { + const dir = await mkdtemp(join("/tmp", "scriptc-target-cpu-refusal-")); + const entry = join(dir, "main.ts"); + try { + await writeFile(entry, program); + for (const target of ["aarch64-unknown-linux-gnu", "wasm32-wasi"]) { + for (const kind of ["llvm", "asm", "obj", "exe"]) { + const build = execFileAsync( + process.execPath, + [ + "--import", + tsxLoader, + cliEntry, + "build", + entry, + `--emit=${kind}`, + "-o", + join(dir, `out-${kind}`), + "--target-cpu=x86-64-v3", + ], + { + env: { ...process.env, SCRIPTC_TARGET: target, SCRIPTC_CACHE_DIR: join(dir, "cache") }, + }, + ); + await expect(build, `${target} ${kind}`).rejects.toThrow( + /--target-cpu=x86-64-v3 requires an x86-64 target/, + ); + } + } + } finally { + await rm(dir, { recursive: true, force: true }); + } +}, 240_000); diff --git a/packages/compiler/ambient/scriptc-node-fallback.d.ts b/packages/compiler/ambient/scriptc-node-fallback.d.ts index 0be5bdca2..b97548e3b 100644 --- a/packages/compiler/ambient/scriptc-node-fallback.d.ts +++ b/packages/compiler/ambient/scriptc-node-fallback.d.ts @@ -1590,7 +1590,7 @@ declare module "node:fs" { export function writeFileSync(path: string | URL, data: string, encoding: "utf8" | "utf-8"): void; export function writeFileSync(path: string | URL, data: string | Uint8Array): void; export function appendFileSync(path: string | URL, data: string | Uint8Array): void; - export function existsSync(path: string): boolean; + export function existsSync(path: string | URL): boolean; export function mkdirSync(path: string | URL): void; export function mkdirSync( path: string | URL, @@ -2231,7 +2231,7 @@ declare module "fs/promises" { stat(): Promise; [Symbol.asyncDispose](): Promise; } - export function open(path: string, flags?: string, mode?: number): Promise; + export function open(path: string | URL, flags?: string, mode?: number): Promise; export function link( existingPath: string | Uint8Array | URL, newPath: string | Uint8Array | URL, @@ -2268,14 +2268,14 @@ declare module "fs/promises" { atime: number | string | Date, mtime: number | string | Date, ): Promise; - export function readFile(path: string, encoding: "utf8" | "utf-8"): Promise; + export function readFile(path: string | URL, encoding: "utf8" | "utf-8"): Promise; export function readFile( - path: string, + path: string | URL, options: BufferEncoding | { encoding: BufferEncoding }, ): Promise; - export function readFile(path: string): Promise; + export function readFile(path: string | URL): Promise; export function writeFile( - path: string, + path: string | URL, data: string, options?: { mode?: number; @@ -2284,14 +2284,18 @@ declare module "fs/promises" { [option: string]: unknown; }, ): Promise; - export function writeFile(path: string, data: string, encoding: "utf8" | "utf-8"): Promise; + export function writeFile( + path: string | URL, + data: string, + encoding: "utf8" | "utf-8", + ): Promise; export function mkdir( - path: string, + path: string | URL, options?: { recursive?: boolean; mode?: number }, ): Promise; - export function readdir(path: string): Promise; + export function readdir(path: string | URL): Promise; export function readdir( - path: string, + path: string | URL, options: { encoding?: "utf8" | "utf-8"; withFileTypes: true; @@ -2308,12 +2312,12 @@ declare module "fs/promises" { pattern: string | readonly string[], options: import("node:fs").GlobOptionsWithFileTypes, ): AsyncIterableIterator; - export function rm(path: string): Promise; + export function rm(path: string | URL): Promise; export function rm( - path: string, + path: string | URL, options: { recursive?: boolean; force?: boolean; maxRetries?: number; retryDelay?: number }, ): Promise; - export function stat(path: string): Promise; + export function stat(path: string | URL): Promise; export function statfs( path: string | Uint8Array | URL, options?: import("node:fs").StatFsOptions & { bigint?: false }, @@ -2326,21 +2330,21 @@ declare module "fs/promises" { path: string | Uint8Array | URL, options?: import("node:fs").StatFsOptions, ): Promise; - export function realpath(path: string): Promise; - export function lstat(path: string): Promise; - export function unlink(path: string): Promise; - export function chmod(path: string, mode: number): Promise; - export function chown(path: string, uid: number, gid: number): Promise; - export function lchown(path: string, uid: number, gid: number): Promise; - export function access(path: string, mode?: number): Promise; - export function truncate(path: string, len?: number): Promise; - export function appendFile(path: string, data: string | Uint8Array): Promise; - export function copyFile(src: string, dest: string): Promise; - export function mkdtemp(prefix: string): Promise; - export function rmdir(path: string): Promise; - export function rename(oldPath: string, newPath: string): Promise; + export function realpath(path: string | URL): Promise; + export function lstat(path: string | URL): Promise; + export function unlink(path: string | URL): Promise; + export function chmod(path: string | URL, mode: number): Promise; + export function chown(path: string | URL, uid: number, gid: number): Promise; + export function lchown(path: string | URL, uid: number, gid: number): Promise; + export function access(path: string | URL, mode?: number): Promise; + export function truncate(path: string | URL, len?: number): Promise; + export function appendFile(path: string | URL, data: string | Uint8Array): Promise; + export function copyFile(src: string | URL, dest: string | URL): Promise; + export function mkdtemp(prefix: string | URL): Promise; + export function rmdir(path: string | URL): Promise; + export function rename(oldPath: string | URL, newPath: string | URL): Promise; export function opendir( - path: string, + path: string | URL, options?: import("node:fs").OpenDirOptions, ): Promise; } diff --git a/packages/compiler/src/backend/llvm/bytes-loop-versions.test.ts b/packages/compiler/src/backend/llvm/bytes-loop-versions.test.ts index de71a6ef0..aef39ef75 100644 --- a/packages/compiler/src/backend/llvm/bytes-loop-versions.test.ts +++ b/packages/compiler/src/backend/llvm/bytes-loop-versions.test.ts @@ -385,7 +385,10 @@ test("while loop counters below a bound version their accesses", () => { ]); const declared = (loop: IrStmt): IrFunction => fn([{ kind: "varDecl", localId: "i", init: ref("k"), loc }, loop]); - const versionOf = (loop: IrStmt & { kind: "while" }): BytesLoopVersion | null => { + const versionOf = ( + loop: IrStmt & { kind: "while" }, + workers = false, + ): BytesLoopVersion | null => { const f = declared(loop); const ranges = analyzeIntegerRanges(f); return matchBytesLoopVersion(loop, null, { @@ -396,7 +399,7 @@ test("while loop counters below a bound version their accesses", () => { initialized: findInitializedByteLoopBindings(f).get(loop) ?? new Set(), proven: new Set(), scalarStrings: new Set(), - workers: false, + workers, owned: new Set(), }); }; @@ -413,6 +416,11 @@ test("while loop counters below a bound version their accesses", () => { // An access after the counter advances in the same pass stays checked. const advanced = whileLoop([step, set(ref("out", f64s), ref("i"), num(0))]); expect(versionOf(advanced)).toBeNull(); + // In a worker program, a while loop versions for its counter's accesses, + // and not for thread-owned receivers alone. + expect(versionOf(copy, true)!.ownedReceivers).toHaveLength(2); + const job = whileLoop([set(ref("out", f64s), ref("k"), get(ref("src", f64s), ref("k"))), step]); + expect(versionOf(job, true)).toBeNull(); const mod: IrModule = { irVersion: 17, sourceFile: loc.file, diff --git a/packages/compiler/src/backend/llvm/bytes-loop-versions.ts b/packages/compiler/src/backend/llvm/bytes-loop-versions.ts index 2013d03e4..1b51deee1 100644 --- a/packages/compiler/src/backend/llvm/bytes-loop-versions.ts +++ b/packages/compiler/src/backend/llvm/bytes-loop-versions.ts @@ -581,6 +581,11 @@ export function matchBytesLoopVersion( pairedKept.length === 0 ) return null; + // A while loop versions only for accesses at its counters. Its other + // checks (thread-owned receivers, static indices) save a test or two per + // pass, while a second copy of its body doubles the calls it makes, which + // can keep a worker's job loop from inlining the function that runs a job. + if (loop.kind === "while" && boundedKept.length === 0) return null; const bounds = new Set(); const ownedNodes = new Set(); const used = new Set(); diff --git a/packages/compiler/src/backend/llvm/emit-module.ts b/packages/compiler/src/backend/llvm/emit-module.ts index 518bbf270..5aa18bf9f 100644 --- a/packages/compiler/src/backend/llvm/emit-module.ts +++ b/packages/compiler/src/backend/llvm/emit-module.ts @@ -64,6 +64,7 @@ import { literalEqualityDefinition } from "./string-literal-equality.js"; import { STRING_READ_HELPERS, stringScanDecl } from "./string-reads.js"; import { emitLibraryIdentityLines } from "../library-identity-markers.js"; import { InternalCompilerError } from "../../errors.js"; +import { targetCpuAttribute } from "../target-cpu.js"; /** The function bodies emitted in `chunks` contiguous ranges, each by its * own emitter (LlEmitter.chunkEmitter), then merged in function order: the @@ -191,6 +192,20 @@ function readRange(dir: string, range: number): { result: ChunkResult; texts: st } } +/** The attribute groups of the emitted functions: #0 for every function, + * #1 for coroutine bodies and #2 for out-of-line record clones. Each carries + * the build's `--target-cpu` level. */ +function functionAttributes(emitter: LlEmitter, hasNoInlineRecordClone: boolean): string[] { + const cpu = targetCpuAttribute(emitter.targetCpu); + return [ + `attributes #0 = { sanitize_address${cpu} }`, + ...(emitter.usesCoroutines + ? [`attributes #1 = { sanitize_address presplitcoroutine${cpu} }`] + : []), + ...(hasNoInlineRecordClone ? [`attributes #2 = { noinline sanitize_address${cpu} }`] : []), + ]; +} + export function emitModuleParts(emitter: LlEmitter): string[] { // Function bodies first (the literal/unit/fn-value tables fill as they // emit), then the file assembles around them — the runtime ABI’s order. @@ -972,9 +987,7 @@ export function emitModuleParts(emitter: LlEmitter): string[] { // symbols specified by the library IR instead. for (const line of emitter.emitLibDefs(globals, globalReleaseLines, stamps)) out.push(line); out.push(...emitter.inlineRcTail(flushedDecls)); - out.push(`attributes #0 = { sanitize_address }`); - if (emitter.usesCoroutines) out.push(`attributes #1 = { sanitize_address presplitcoroutine }`); - if (hasNoInlineRecordClone) out.push(`attributes #2 = { noinline sanitize_address }`); + out.push(...functionAttributes(emitter, hasNoInlineRecordClone)); out.push(``); out.push(...emitter.fieldAliasMetadata()); if (emitter.debug !== null) out.push(emitter.debug.render()); @@ -1122,9 +1135,7 @@ export function emitModuleParts(emitter: LlEmitter): string[] { // sanitize_address is inert under the plain pipeline; the sanitized // lane's -fsanitize=address link activates instrumentation over the // emitted functions too (the runtime TUs get theirs from clang). - `attributes #0 = { sanitize_address }`, - ...(emitter.usesCoroutines ? [`attributes #1 = { sanitize_address presplitcoroutine }`] : []), - ...(hasNoInlineRecordClone ? [`attributes #2 = { noinline sanitize_address }`] : []), + ...functionAttributes(emitter, hasNoInlineRecordClone), ``, ); if (emitter.mod.workers) { diff --git a/packages/compiler/src/backend/llvm/emitter.test.ts b/packages/compiler/src/backend/llvm/emitter.test.ts index 23ca9d9d6..876476cc1 100644 --- a/packages/compiler/src/backend/llvm/emitter.test.ts +++ b/packages/compiler/src/backend/llvm/emitter.test.ts @@ -201,6 +201,16 @@ test("detached exit defines the runtime opt-in symbol only when requested", () = expect(detached.replace("@scr_detached_exit = constant i8 1\n", "")).toBe(plain); }); +test("a target CPU level reaches every function attribute group and nothing else", () => { + const module = moduleFor([call()]); + const plain = emitLlvmModule(module); + expect(plain).not.toContain("target-cpu"); + const v3 = emitLlvmModule(module, { targetCpu: "x86-64-v3" }); + expect(v3).toMatch(/^attributes #0 = \{ sanitize_address "target-cpu"="x86-64-v3" \}$/m); + expect(v3.replaceAll(' "target-cpu"="x86-64-v3"', "")).toBe(plain); + expect(emitLlvmModule(module, { targetCpu: "baseline" })).toBe(plain); +}); + test("inline RC inlines every fast path at speed and keeps box releases as calls in release", () => { for (const boxed of [false, true]) { const module = moduleFor(Array.from({ length: 4 }, call), { boxed }); diff --git a/packages/compiler/src/backend/llvm/emitter.ts b/packages/compiler/src/backend/llvm/emitter.ts index 13011f21e..fe50e974f 100644 --- a/packages/compiler/src/backend/llvm/emitter.ts +++ b/packages/compiler/src/backend/llvm/emitter.ts @@ -1,3 +1,4 @@ +import type { TargetCpu } from "../target-cpu.js"; import { relieveHeapPressure } from "../../heap-pressure.js"; import { wasmBoundsChecks } from "./wasm-modules.js"; import { jsTruthy } from "./jsval.js"; @@ -482,6 +483,10 @@ export interface LlvmTargetOptions { * store the instance's address over 8 in 32 bits (classes.ts * compactRefLoadLines). Elsewhere they keep 8-byte pointers. */ lowRegion?: boolean; + /** The x86-64 level the program's code may require (`--target-cpu`): + * every emitted function carries it as its `target-cpu`. Unset, code + * runs on any x86-64 CPU. */ + targetCpu?: TargetCpu; } /** LlvmTargetOptions.lowRegion for an executable: Linux targets without @@ -1048,6 +1053,7 @@ export class LlEmitter { readonly emitLibraryIdentity: boolean; readonly runtimeAbiMarker: boolean; readonly detachedExit: boolean; + readonly targetCpu: TargetCpu | undefined; readonly wasmBoundsChecks: boolean; /** Interned string literals: UTF-8 text → { symbol, byte length } — * first-use order, the runtime ABI’s determinism discipline. */ @@ -1630,6 +1636,7 @@ export class LlEmitter { this.emitLibraryIdentity = options.emitLibraryIdentity !== false; this.runtimeAbiMarker = options.runtimeAbiMarker === true; this.detachedExit = options.detachedExit === true; + this.targetCpu = options.targetCpu; this.wasmBoundsChecks = options.wasmBoundsChecks === true || wasmBoundsChecks({ @@ -3801,8 +3808,17 @@ export class LlEmitter { /** A union value's tag (i32): the box's tag word, or the NULL test of a * nullable union's pointer. */ + /** A box's tag and payload words are written only where the box is + * built (runtime constructors and untagged stack stores), so their loads + * carry their own alias tags: element and field stores cannot change + * them, and loops keep them hoisted. */ unionTag(uName: string, unionId: string): string { - return emitUnionTag(this.B, this.nullableUnions.get(unionId), uName); + return emitUnionTag( + this.B, + this.nullableUnions.get(unionId), + uName, + this.memoryAliasAttachment("union.tag"), + ); } /** Guard for helpers that build or read a union's `%ScrUnion` box @@ -3818,7 +3834,12 @@ export class LlEmitter { /** The BORROWED payload pointer of a ref arm (scr_union_peek inlined — * the runtime's is a static inline); a nullable union is its payload. */ unionPeek(uName: string, unionId: string): string { - return emitUnionPeek(this.B, this.nullableUnions.get(unionId), uName); + return emitUnionPeek( + this.B, + this.nullableUnions.get(unionId), + uName, + this.memoryAliasAttachment("union.payload"), + ); } /** Emits `switch` over a union's tag with one block per arm; each arm diff --git a/packages/compiler/src/backend/llvm/expr-bytes.ts b/packages/compiler/src/backend/llvm/expr-bytes.ts index 41ddf2b9d..23025d00c 100644 --- a/packages/compiler/src/backend/llvm/expr-bytes.ts +++ b/packages/compiler/src/backend/llvm/expr-bytes.ts @@ -4,7 +4,7 @@ import { emitBorrowedInput } from "./borrowed-inputs.js"; import { BYTES_ELEMENT_SIZE, F64, type IrBytesElem, type IrExpr } from "../../ir/ir.js"; import type { LlvmEmitterContext, LlValue } from "./expr-context.js"; import { F64_INF, exactIndexLines, f64Lit, sizeToDouble } from "./common.js"; -import { exactInteger, widenInteger } from "./integer-values.js"; +import { alwaysExactInteger, exactInteger } from "./integer-values.js"; import { byteNumberAccess } from "../../ir/byte-numbers.js"; import { emitByteNumber } from "./byte-numbers.js"; import { integerOrNaNIndex, integerOrNaNRange } from "./wrapping-sums.js"; @@ -86,17 +86,18 @@ export function emitBytesIndex( ): string { const B = host.B; const index = value.name; - const integer = exactInteger(host, value, expr); - if (inBounds && integer) { - const wide = widenInteger(host, integer); - if (host.sizeType === "i64") return wide; + // The exact integer, or a wide view that always equals the double (such + // as `i + 1` of a speculated counter). + const exact = alwaysExactInteger(host, value, exactInteger(host, value, expr)); + if (inBounds && exact !== null) { + if (host.sizeType === "i64") return exact; const narrow = B.tmp(); - B.line(`${narrow} = trunc i64 ${wide} to ${host.sizeType}`); + B.line(`${narrow} = trunc i64 ${exact} to ${host.sizeType}`); return narrow; } const len = emitBytesHeaderLength(host, receiver); - if (integer) { - const wideIndex = widenInteger(host, integer); + if (exact !== null) { + const wideIndex = exact; const wideLength = host.sizeType === "i64" ? len : B.tmp(); if (host.sizeType !== "i64") B.line(`${wideLength} = zext ${host.sizeType} ${len} to i64`); const inRange = B.tmp(); @@ -936,6 +937,23 @@ export function emitBytesIntrinsic( B.line(`${t} = fcmp one double ${idx.name}, ${f64Lit(-1)}`); return { name: t, type: e.type }; } + case "indexOfElem": + case "lastIndexOfElem": + return call( + "scr_bytes_index_of_elem", + "double (ptr, double, double, i1 zeroext)", + `ptr ${r.name}, double ${args[0]!.name}, double ${args[1]!.name}, i1 ${method === "lastIndexOfElem"}`, + false, + false, + ); + case "includesElem": + return call( + "scr_bytes_includes_elem", + "zeroext i1 (ptr, double, double)", + `ptr ${r.name}, double ${args[0]!.name}, double ${args[1]!.name}`, + false, + false, + ); case "fillElem": // Per-element TypedArray fill (non-u8): slice-style index // defaults, never throws; the receiver comes back +1. diff --git a/packages/compiler/src/backend/llvm/expr-primitives.ts b/packages/compiler/src/backend/llvm/expr-primitives.ts index 2c6ce35db..04af0fcef 100644 --- a/packages/compiler/src/backend/llvm/expr-primitives.ts +++ b/packages/compiler/src/backend/llvm/expr-primitives.ts @@ -13,6 +13,7 @@ import { emitStringInputs } from "./string-lifetimes.js"; import { emitStringParts, numberPart, stringParts } from "./string-construction.js"; import { emitBorrowedInputs } from "./borrowed-inputs.js"; import { + alwaysExactInteger, exactInteger, int32Number, integerNumber, @@ -46,6 +47,16 @@ import { } from "./dense-array-access.js"; import { emitWrappingSum, markWrappingOperands } from "./wrapping-sums.js"; +/** The signed integer predicates of the number compares. */ +const INTEGER_COMPARES: Record = { + "<": "slt", + "<=": "sle", + ">": "sgt", + ">=": "sge", + "===": "eq", + "!==": "ne", +}; + export function emitLiteralExpr( host: LlvmEmitterContext, e: ExprOf<"numLit" | "boolLit" | "strLit" | "moduleNsRef" | "unitLit" | "varRef">, @@ -318,15 +329,7 @@ export function emitOperatorExpr( if (integer) return integer; } if (cmp[e.op] !== undefined) { - const integerCmp: Record = { - "<": "slt", - "<=": "sle", - ">": "sgt", - ">=": "sge", - "===": "eq", - "!==": "ne", - }; - B.line(`${t} = icmp ${integerCmp[e.op]} i64 ${left}, ${right}`); + B.line(`${t} = icmp ${INTEGER_COMPARES[e.op]} i64 ${left}, ${right}`); return { name: t, type: e.type }; } } @@ -356,7 +359,17 @@ export function emitOperatorExpr( if (wideInteger) value.wideInteger = wideInteger; return value; } - } else B.line(`${t} = fcmp ${cmp[e.op]} double ${l.name}, ${r.name}`); + } else { + // An operand with only a wide view, such as `i + 1` of a speculated + // counter: compare the integers when both always equal the doubles. + const left = alwaysExactInteger(host, l, li), + right = left === null ? null : alwaysExactInteger(host, r, ri); + if (left !== null && right !== null) { + B.line(`${t} = icmp ${INTEGER_COMPARES[e.op]} i64 ${left}, ${right}`); + return { name: t, type: e.type }; + } + B.line(`${t} = fcmp ${cmp[e.op]} double ${l.name}, ${r.name}`); + } } else if (bit[e.op] !== undefined) { if (e.left.type.kind !== "f64" || e.right.type.kind !== "f64") { throw new LlvmUnsupportedError( diff --git a/packages/compiler/src/backend/llvm/expr-stream-bridges.ts b/packages/compiler/src/backend/llvm/expr-stream-bridges.ts index 1acc75cec..46c2d884f 100644 --- a/packages/compiler/src/backend/llvm/expr-stream-bridges.ts +++ b/packages/compiler/src/backend/llvm/expr-stream-bridges.ts @@ -993,6 +993,28 @@ export function streamTypedRefMaterializeAdapter( : `call void @scr_error_dyn_place_name(ptr ${out}, ptr ${host.cstr(nameAfter)}, ${host.sizeType} ${Buffer.byteLength(nameAfter, "utf8")})`, ); } + // A JavaScript Error subclass's property bag records, once a dynamic + // write has committed the view, where `name` and `message` were created + // (zod's error factory assigns `name` before it defines `message`); the + // slots hold their values. A `name` the class's construction sets is + // already where errorNameAfter places it among the fields. + if ( + meta.def.errorBagOrder && + meta.root.def.name === "%Error" && + meta.def.fields.some((field) => field.name === DYN_CLASS_PROPERTIES) + ) { + host.declare(`declare void @scr_error_dyn_follow_bag(ptr, ptr, ptr, i1)`); + const { index } = classFieldIndex(meta, DYN_CLASS_PROPERTIES); + const slot = B.tmp(), + bag = B.tmp(); + B.line( + `${slot} = getelementptr inbounds %${classStructSym(t.className)}, ptr %p, i64 0, i32 ${index}`, + ); + B.line(`${bag} = load ptr, ptr ${slot}`); + B.line( + `call void @scr_error_dyn_follow_bag(ptr ${out}, ptr ${bag}, ptr %p, i1 ${nameAfter === undefined ? "false" : "true"})`, + ); + } // A frozen, sealed or non-extensible instance's view says so. if (withAttributes) { host.declare(`declare void @scr_dyn_class_view_attributes(ptr, double)`); diff --git a/packages/compiler/src/backend/llvm/integer-values.ts b/packages/compiler/src/backend/llvm/integer-values.ts index ddbcdd8da..90954f4c6 100644 --- a/packages/compiler/src/backend/llvm/integer-values.ts +++ b/packages/compiler/src/backend/llvm/integer-values.ts @@ -65,6 +65,22 @@ function magnitude(range: IntegerRange): number { return Math.max(Math.abs(range.min), Math.abs(range.max)); } +/** The bounds of a sum and a product of two magnitude bounds, as doubles. + * Below 2^53 the double result is the integer; above it the result is past + * every exactness limit either way, and rounding moves it far less than the + * margin from WIDE_BOUND to the i64 range. A result of exactly 2^53 may + * stand for 2^53 + 1 (2^53 - 1 + 2 rounds to 2^53), so it moves to the next + * double up unless the sum is exactly 2^53. */ +function sumBound(a: number, b: number): number { + const sum = a + b; + return sum === EXACT_DOUBLES && EXACT_DOUBLES - a !== b ? sum + 2 : sum; +} + +function productBound(a: number, b: number): number { + const product = a * b; + return product === EXACT_DOUBLES ? product + 2 : product; +} + /** The wide view of `l op r` for `+`, `-` and `*` when the exact operands * have no proven result range (their product or sum may pass 2^53). * @@ -98,14 +114,14 @@ export function wideArithmetic( if (op === "*") { if (lv.below !== Infinity || rv.below !== Infinity) return null; below = EXACT_DOUBLES; - bound = lv.bound * rv.bound; + bound = productBound(lv.bound, rv.bound); } else if (lv.below === Infinity && rv.below === Infinity) { below = EXACT_DOUBLES; - bound = lv.bound + rv.bound; + bound = sumBound(lv.bound, rv.bound); } else if (lv.below === Infinity || rv.below === Infinity) { const [exact, wide] = lv.below === Infinity ? [lv, rv] : [rv, lv]; below = Math.min(wide.below - exact.bound, EXACT_DOUBLES); - bound = exact.bound + wide.bound; + bound = sumBound(exact.bound, wide.bound); } else return null; if (!(bound <= WIDE_BOUND) || below < INDEX_LIMIT) return null; const operand = (value: LlValue, integer: LlInteger | null): string => @@ -134,6 +150,20 @@ export function indexInteger( return wide && wide.below >= INDEX_LIMIT ? wide.name : null; } +/** An i64 equal to a value's double wherever the value can be: its exact + * integer, or a wide view whose magnitude bound is within the magnitude below + * which the view is exact. Every integer of magnitude at most 2^53 is a + * double, so a bound equal to that limit is exact too. Null otherwise. */ +export function alwaysExactInteger( + host: LlvmEmitterContext, + value: LlValue, + integer: LlInteger | null, +): string | null { + if (integer) return widenInteger(host, integer); + const wide = value.wideInteger; + return wide && wide.bound <= wide.below ? wide.name : null; +} + export function widenInteger(host: LlvmEmitterContext, value: LlInteger): string { if (value.type === "i64") return value.name; const wide = host.B.tmp(); diff --git a/packages/compiler/src/backend/llvm/lib-shared.ts b/packages/compiler/src/backend/llvm/lib-shared.ts index e41a386eb..d58431d29 100644 --- a/packages/compiler/src/backend/llvm/lib-shared.ts +++ b/packages/compiler/src/backend/llvm/lib-shared.ts @@ -658,6 +658,7 @@ export const LIB_FN_SYMS: Record = { "process.stderrWriteBytes": "scr_process_stderr_write_bytes", "insp.buffer": "scr_insp_buffer", "insp.typedArray": "scr_insp_typed_array", + "insp.bytes": "scr_insp_bytes", // util.inspect (scr_inspect.c): all args borrowed; string results +1; // the begin/entry/key/moreItems/end accumulator drives the emitted // container walks. None of these throw. @@ -713,6 +714,9 @@ export const LIB_FN_SYMS: Record = { "url.hash": "scr_url_hash", "url.searchParams": "scr_url_search_params", "url.fileURLToPathUrl": "scr_url_to_path", + "url.fileURLToPathValue": "scr_url_to_path_value", + "fs.checkPath": "scr_fs_check_path", + "fs.checkReadEncoding": "scr_fs_check_read_encoding", "url.fileURLToPathChecked": "scr_url_checked_to_path", "url.fileURLToPathOptions": "scr_url_checked_to_path_options", "url.fileURLToPathBuffer": "scr_url_checked_to_path_buffer", diff --git a/packages/compiler/src/backend/llvm/reference-effects.ts b/packages/compiler/src/backend/llvm/reference-effects.ts index 6ebe1222b..43a9176d5 100644 --- a/packages/compiler/src/backend/llvm/reference-effects.ts +++ b/packages/compiler/src/backend/llvm/reference-effects.ts @@ -181,6 +181,9 @@ function expressionPreservesEdges( case "bytesIntrinsic": return ( e.method === "get" || + e.method === "indexOfElem" || + e.method === "lastIndexOfElem" || + e.method === "includesElem" || e.method === "length" || e.method === "byteLength" || e.method === "byteOffset" || diff --git a/packages/compiler/src/backend/llvm/union-repr.ts b/packages/compiler/src/backend/llvm/union-repr.ts index 959f9e51e..8c6fb4c85 100644 --- a/packages/compiler/src/backend/llvm/union-repr.ts +++ b/packages/compiler/src/backend/llvm/union-repr.ts @@ -8,8 +8,14 @@ import type { BlockBuilder } from "./blocks.js"; import { NULLABLE_NULL, shapeTagFn, type PointerUnion } from "./nullable-unions.js"; -/** The i32 tag of a union value. */ -export function emitUnionTag(B: BlockBuilder, nullable: PointerUnion | null, u: string): string { +/** The i32 tag of a union value. `alias` is the tag word's alias + * attachment (", !tbaa !N"), when the caller has one. */ +export function emitUnionTag( + B: BlockBuilder, + nullable: PointerUnion | null, + u: string, + alias = "", +): string { const t = B.tmp(); if (nullable?.repr === "shapes") { B.line(`${t} = call i32 ${shapeTagFn(nullable.unionId)}(ptr ${u})`); @@ -31,7 +37,7 @@ export function emitUnionTag(B: BlockBuilder, nullable: PointerUnion | null, u: } const p = B.tmp(); B.line(`${p} = getelementptr inbounds %ScrUnion, ptr ${u}, i64 0, i32 1`); - B.line(`${t} = load i32, ptr ${p}`); + B.line(`${t} = load i32, ptr ${p}${alias}`); return t; } @@ -81,12 +87,17 @@ export function emitNullableIsTag( } /** The BORROWED payload pointer of a reference arm. */ -export function emitUnionPeek(B: BlockBuilder, nullable: PointerUnion | null, u: string): string { +export function emitUnionPeek( + B: BlockBuilder, + nullable: PointerUnion | null, + u: string, + alias = "", +): string { if (nullable) return u; const p = B.tmp(); const t = B.tmp(); B.line(`${p} = getelementptr inbounds %ScrUnion, ptr ${u}, i64 0, i32 5`); - B.line(`${t} = load ptr, ptr ${p}`); + B.line(`${t} = load ptr, ptr ${p}${alias}`); return t; } diff --git a/packages/compiler/src/backend/target-cpu.test.ts b/packages/compiler/src/backend/target-cpu.test.ts new file mode 100644 index 000000000..428caee32 --- /dev/null +++ b/packages/compiler/src/backend/target-cpu.test.ts @@ -0,0 +1,51 @@ +import { expect, test } from "vitest"; +import { + isTargetCpuOption, + targetCpuAttribute, + targetCpuField, + targetCpuKeyParts, + targetCpuTargetRefusal, + x86Level, +} from "./target-cpu.js"; + +const V2 = ["cx16", "lahf_lm", "popcnt", "pni", "sse4_1", "sse4_2", "ssse3"]; +const V3 = ["avx", "avx2", "bmi1", "bmi2", "f16c", "fma", "abm", "movbe", "xsave"]; +const V4 = ["avx512f", "avx512bw", "avx512cd", "avx512dq", "avx512vl"]; +const level = (flags: string[]) => x86Level((flag) => flags.includes(flag)); + +test("a host's level is the highest one whose features and lower levels' features it has", () => { + expect(level([])).toBe("baseline"); + expect(level(V2)).toBe("x86-64-v2"); + expect(level([...V2, ...V3])).toBe("x86-64-v3"); + expect(level([...V2, ...V3, ...V4])).toBe("x86-64-v4"); + // AVX-512 without BMI2 is still v2. + expect(level([...V2, ...V3.filter((f) => f !== "bmi2"), ...V4])).toBe("x86-64-v2"); + expect(level([...V3, ...V4])).toBe("baseline"); +}); + +test("the baseline adds no key part, attribute or option field", () => { + for (const cpu of [undefined, "baseline"] as const) { + expect(targetCpuKeyParts(cpu)).toEqual([]); + expect(targetCpuAttribute(cpu)).toBe(""); + expect(targetCpuField(cpu)).toEqual({}); + expect(targetCpuTargetRefusal(cpu, "linux", "arm64")).toBeNull(); + } + expect(targetCpuKeyParts("x86-64-v3")).toEqual(["target-cpu-x86-64-v3"]); + expect(targetCpuAttribute("x86-64-v4")).toBe(' "target-cpu"="x86-64-v4"'); + expect(targetCpuField("x86-64-v2")).toEqual({ targetCpu: "x86-64-v2" }); +}); + +test("a level is refused for targets that are not x86-64", () => { + expect(targetCpuTargetRefusal("x86-64-v3", "linux", "x64")).toBeNull(); + expect(targetCpuTargetRefusal("x86-64-v3", "darwin", "arm64")).toBe( + "--target-cpu=x86-64-v3 requires an x86-64 target", + ); + expect(targetCpuTargetRefusal("x86-64-v3", "wasi", "x64")).not.toBeNull(); +}); + +test("the option accepts the levels and native", () => { + for (const value of ["baseline", "x86-64-v2", "x86-64-v3", "x86-64-v4", "native"]) + expect(isTargetCpuOption(value)).toBe(true); + for (const value of ["x86-64", "haswell", "", undefined]) + expect(isTargetCpuOption(value)).toBe(false); +}); diff --git a/packages/compiler/src/backend/target-cpu.ts b/packages/compiler/src/backend/target-cpu.ts new file mode 100644 index 000000000..51db87bbb --- /dev/null +++ b/packages/compiler/src/backend/target-cpu.ts @@ -0,0 +1,118 @@ +import { execFileSync } from "node:child_process"; +import { readFileSync } from "node:fs"; + +/** The x86-64 microarchitecture levels an executable may require + * (`--target-cpu`). `baseline` is the x86-64 baseline every release build + * targets; the other levels are LLVM's `x86-64-v2` to `x86-64-v4`. */ +export type TargetCpu = "baseline" | "x86-64-v2" | "x86-64-v3" | "x86-64-v4"; + +/** The values `--target-cpu` accepts: a level, or `native` for the highest + * level the build host supports. */ +export const TARGET_CPU_OPTIONS = ["baseline", "x86-64-v2", "x86-64-v3", "x86-64-v4", "native"]; + +export function isTargetCpuOption(value: unknown): value is TargetCpu | "native" { + return typeof value === "string" && TARGET_CPU_OPTIONS.includes(value); +} + +/** The CPU features each level adds to the one below it, as Linux names + * them in /proc/cpuinfo (`pni` is SSE3, `abm` includes LZCNT). */ +const LEVEL_FLAGS: readonly [TargetCpu, readonly string[]][] = [ + ["x86-64-v2", ["cx16", "lahf_lm", "popcnt", "pni", "sse4_1", "sse4_2", "ssse3"]], + ["x86-64-v3", ["avx", "avx2", "bmi1", "bmi2", "f16c", "fma", "abm", "movbe", "xsave"]], + ["x86-64-v4", ["avx512f", "avx512bw", "avx512cd", "avx512dq", "avx512vl"]], +]; + +/** The macOS `machdep.cpu` spellings of the same features. */ +const DARWIN_FLAGS: Readonly> = { + cx16: "CX16", + lahf_lm: "LAHF", + popcnt: "POPCNT", + pni: "SSE3", + sse4_1: "SSE4.1", + sse4_2: "SSE4.2", + ssse3: "SSSE3", + avx: "AVX1.0", + avx2: "AVX2", + bmi1: "BMI1", + bmi2: "BMI2", + f16c: "F16C", + fma: "FMA", + abm: "LZCNT", + movbe: "MOVBE", + xsave: "XSAVE", + avx512f: "AVX512F", + avx512bw: "AVX512BW", + avx512cd: "AVX512CD", + avx512dq: "AVX512DQ", + avx512vl: "AVX512VL", +}; + +/** The highest level whose features (and every lower level's) are all in + * `has`. */ +export function x86Level(has: (flag: string) => boolean): TargetCpu { + let level: TargetCpu = "baseline"; + for (const [next, flags] of LEVEL_FLAGS) { + if (!flags.every(has)) break; + level = next; + } + return level; +} + +/** The highest level the build host supports, or null when the host is not + * an x86-64 Linux or macOS machine. */ +export function hostTargetCpu( + platform: string = process.platform, + arch: string = process.arch, +): TargetCpu | null { + if (arch !== "x64") return null; + try { + if (platform === "linux") { + const info = readFileSync("/proc/cpuinfo", "utf8"); + const line = /^flags\s*:(.*)$/m.exec(info); + if (!line) return null; + const flags = new Set(line[1]!.trim().split(/\s+/)); + return x86Level((flag) => flags.has(flag)); + } + if (platform === "darwin") { + const text = execFileSync( + "sysctl", + ["-n", "machdep.cpu.features", "machdep.cpu.leaf7_features", "machdep.cpu.extfeatures"], + { encoding: "utf8" }, + ); + const flags = new Set(text.trim().split(/\s+/)); + return x86Level((flag) => flags.has(DARWIN_FLAGS[flag]!)); + } + } catch { + return null; + } + return null; +} + +/** Cache-key parts for a target level; empty for the baseline so existing + * keys are unchanged. */ +export function targetCpuKeyParts(cpu: TargetCpu | undefined): string[] { + return cpu === undefined || cpu === "baseline" ? [] : [`target-cpu-${cpu}`]; +} + +/** The function attribute that makes LLVM generate code for the level, or + * "" for the baseline. */ +export function targetCpuAttribute(cpu: TargetCpu | undefined): string { + return cpu === undefined || cpu === "baseline" ? "" : ` "target-cpu"="${cpu}"`; +} + +/** The options field for a level: omitted for the baseline. */ +export function targetCpuField(cpu: TargetCpu | undefined): { targetCpu?: TargetCpu } { + return cpu === undefined || cpu === "baseline" ? {} : { targetCpu: cpu }; +} + +/** The error for a level on a target that is not x86-64, or null. */ +export function targetCpuTargetRefusal( + cpu: TargetCpu | undefined, + platform: string, + architecture: string, +): string | null { + if (cpu === undefined || cpu === "baseline") return null; + if (architecture !== "x64" || platform === "wasi") + return `--target-cpu=${cpu} requires an x86-64 target`; + return null; +} diff --git a/packages/compiler/src/cli/command.ts b/packages/compiler/src/cli/command.ts index 3285aa7fa..57b15349e 100644 --- a/packages/compiler/src/cli/command.ts +++ b/packages/compiler/src/cli/command.ts @@ -15,6 +15,13 @@ import { buildEnvelope } from "../diagnostics/envelope.js"; import { renderWarnings } from "../diagnostics/render.js"; import { isPerformanceCode, renderPerformanceWarnings } from "../diagnostics/dynamic-tier.js"; import { isNativeOptimization } from "../backend/optimization.js"; +import { + hostTargetCpu, + isTargetCpuOption, + TARGET_CPU_OPTIONS, + targetCpuField, + type TargetCpu, +} from "../backend/target-cpu.js"; import { setProvenanceSources } from "../frontend/provenance-registry.js"; import type { CliHost, NativeCacheWarmProfile } from "./host.js"; import { resolveOutputOptions } from "./output-options.js"; @@ -96,6 +103,7 @@ async function main(args: string[], host: CliHost): Promise { values.strip || values["windows-subsystem"] !== undefined || values["detached-exit"] || + values["target-cpu"] !== undefined || values["profile-generate"] !== undefined || values["profile-use"] !== undefined || values.sparse || @@ -154,6 +162,7 @@ async function main(args: string[], host: CliHost): Promise { values.sanitize || values["windows-subsystem"] !== undefined || values["detached-exit"] || + values["target-cpu"] !== undefined || values["profile-generate"] !== undefined || values["profile-use"] !== undefined || (values["npm-static"] ?? []).length > 0 || @@ -210,6 +219,7 @@ async function main(args: string[], host: CliHost): Promise { values.ffi !== undefined || values["windows-subsystem"] !== undefined || values["detached-exit"] || + values["target-cpu"] !== undefined || values["profile-generate"] !== undefined || values["profile-use"] !== undefined || (values["npm-static"] ?? []).length > 0 || @@ -218,7 +228,7 @@ async function main(args: string[], host: CliHost): Promise { values["explain-dynamic"] !== undefined ) { fail( - "scriptc build --lib takes no --backend/--emit/--print/--optimization/--strip/--windows-subsystem/--detached-exit/--profile-generate/--profile-use/--npm-static/--ffi/--external-types/--tier/--explain-dynamic: the profile pins the emission and optimization, npm imports are judged automatically, outbound FFI belongs to executable builds, external type mappings belong to coverage, and library artifacts are static-tier only", + "scriptc build --lib takes no --backend/--emit/--print/--optimization/--strip/--windows-subsystem/--detached-exit/--target-cpu/--profile-generate/--profile-use/--npm-static/--ffi/--external-types/--tier/--explain-dynamic: the profile pins the emission and optimization, npm imports are judged automatically, outbound FFI belongs to executable builds, external type mappings belong to coverage, and library artifacts are static-tier only", ); } const profilePath = resolve(profileArg); @@ -261,6 +271,9 @@ async function main(args: string[], host: CliHost): Promise { if (command === "coverage" && values.strip) { fail(`--strip is a build/run option\n\n${USAGE}`); } + if (command === "coverage" && values["target-cpu"] !== undefined) { + fail(`--target-cpu is a build/run option\n\n${USAGE}`); + } if (command === "coverage" && values["detached-exit"]) { fail(`--detached-exit is a build/run option\n\n${USAGE}`); } @@ -377,6 +390,19 @@ async function main(args: string[], host: CliHost): Promise { if (windowsSubsystem !== undefined && command === "coverage") { fail(`--windows-subsystem is only supported for executable builds\n\n${USAGE}`); } + const targetCpuOption = values["target-cpu"]; + if (targetCpuOption !== undefined && !isTargetCpuOption(targetCpuOption)) { + fail( + `unknown target CPU "${targetCpuOption}" (supported: ${TARGET_CPU_OPTIONS.join(", ")})\n\n${USAGE}`, + ); + } + let targetCpu: TargetCpu | undefined = targetCpuOption === "native" ? undefined : targetCpuOption; + if (targetCpuOption === "native") { + const level = hostTargetCpu(); + if (level === null) + fail(`--target-cpu=native needs an x86-64 Linux or macOS build host; name a level instead`); + targetCpu = level; + } const output = command === "coverage" ? null @@ -498,6 +524,7 @@ async function main(args: string[], host: CliHost): Promise { ...(values.strip ? { strip: true } : {}), ...(windowsSubsystem !== undefined ? { windowsSubsystem } : {}), ...(values["detached-exit"] ? { detachedExit: true } : {}), + ...targetCpuField(targetCpu), // The instrumented program resolves a relative pattern against its own // working directory; resolve it here, where the user wrote it. ...(values["profile-generate"] === undefined diff --git a/packages/compiler/src/cli/usage.ts b/packages/compiler/src/cli/usage.ts index c4d07656a..01afbdd41 100644 --- a/packages/compiler/src/cli/usage.ts +++ b/packages/compiler/src/cli/usage.ts @@ -54,6 +54,12 @@ Options: a helper process so the parent sees the exit sooner. The helper keeps the executable mapped and its memory charged for up to ~100 ms after exit (opt in; --emit=exe only) + --target-cpu + x86-64: compile the program's code for a CPU level + (default: baseline, which runs on any x86-64 CPU). The + executable then requires that level: v3 adds AVX2, BMI2 + and FMA, v4 adds AVX-512. native uses the build host's + level. The runtime keeps its baseline code --profile-generate Linux: instrument the executable for profile-guided optimization. Each run writes a raw profile to the @@ -123,6 +129,7 @@ export const CLI_OPTIONS = { strip: { type: "boolean", default: false }, "windows-subsystem": { type: "string" }, "detached-exit": { type: "boolean", default: false }, + "target-cpu": { type: "string" }, "profile-generate": { type: "string" }, "profile-use": { type: "string" }, sparse: { type: "boolean", default: false }, diff --git a/packages/compiler/src/compile-types.ts b/packages/compiler/src/compile-types.ts index 9ecd75974..c50931c6a 100644 --- a/packages/compiler/src/compile-types.ts +++ b/packages/compiler/src/compile-types.ts @@ -1,3 +1,4 @@ +import type { TargetCpu } from "./backend/target-cpu.js"; import type { WindowsSubsystem } from "./backend/targets.js"; import type { NativeLinkInfo } from "./backend/native-link-info.js"; import type { NativeOptimization } from "./backend/optimization.js"; @@ -36,6 +37,12 @@ export interface CompileBaseOptions { * executable mapped and the memory charged for up to ~100 ms after exit. * Executables only; no effect on other platforms. */ detachedExit?: boolean; + /** The x86-64 microarchitecture level the executable's code may require + * (`x86-64-v2` to `x86-64-v4`). The program's functions are compiled for + * that level, and the executable then runs only on CPUs that support it. + * Unset or `baseline`, the code runs on any x86-64 CPU. x86-64 targets + * only; the runtime pack keeps its baseline code. */ + targetCpu?: TargetCpu; /** Instrument the executable for profile-guided optimization. Each run * writes a raw profile to this path pattern (`%p` expands to the process * id; `LLVM_PROFILE_FILE` overrides the pattern at run time). Linux diff --git a/packages/compiler/src/executable/executable-cache.test.ts b/packages/compiler/src/executable/executable-cache.test.ts index e066c0412..e1235d8ac 100644 --- a/packages/compiler/src/executable/executable-cache.test.ts +++ b/packages/compiler/src/executable/executable-cache.test.ts @@ -236,6 +236,8 @@ test("early executable keys isolate compile modes, paths, implementation, checke { ...f.options, optimization: "dev" }, { ...f.options, optimization: "speed" }, { ...f.options, detachedExit: true }, + { ...f.options, targetCpu: "x86-64-v3" }, + { ...f.options, targetCpu: "x86-64-v4" }, { ...f.options, profile: "profile-generate:/tmp/%p.profraw" }, { ...f.options, profile: "profile-use:digest" }, { ...f.options, checker: "tsc-ts-executable-identity" }, diff --git a/packages/compiler/src/executable/executable-cache.ts b/packages/compiler/src/executable/executable-cache.ts index 96922ca13..50f783761 100644 --- a/packages/compiler/src/executable/executable-cache.ts +++ b/packages/compiler/src/executable/executable-cache.ts @@ -18,6 +18,7 @@ import { compilerReleaseVersion } from "../library/sidecar.js"; import { nativeArtifactDependenciesStillMatch } from "../backend/native/artifact-stamps.js"; import { type NativeArtifactDependency } from "../backend/native/contracts.js"; import { optimizationKeyParts } from "../backend/optimization.js"; +import { type TargetCpu, targetCpuKeyParts } from "../backend/target-cpu.js"; import { privateSiblingPath } from "../backend/build-cache.js"; import type { CompilerImplementationDependency } from "../library/compiler-self-identity.js"; import { @@ -139,6 +140,8 @@ export interface EarlyExecutableCacheOptions { windowsSubsystem?: "gui"; /** Omitted unless --detached-exit is set. */ detachedExit?: true; + /** Omitted for the x86-64 baseline (`--target-cpu`). */ + targetCpu?: TargetCpu; /** Omitted without --profile-generate/--profile-use; otherwise the * pattern, or the digest of the profile's bytes. */ profile?: string; @@ -281,6 +284,7 @@ function cacheKey(options: EarlyExecutableCacheOptions): string { ...(options.strip ? ["strip"] : []), ...(options.windowsSubsystem === "gui" ? ["windows-subsystem-gui"] : []), ...(options.detachedExit ? ["detached-exit"] : []), + ...targetCpuKeyParts(options.targetCpu), ...(options.profile !== undefined ? [options.profile] : []), ...(options.checker !== undefined ? [`checker:${options.checker}`] : []), ...(options.youngNursery ? ["young-nursery"] : []), @@ -323,6 +327,7 @@ function routeKey(options: Omit