runcpu: support -fopenmp (dynamic linking)

assets/specTemplate.cfg

@@ -25,7 +25,7 @@ verify_binaries      = 0
 intrate,fprate:
    copies           = 1   # EDIT to change number of copies (see above)
 intspeed,fpspeed:
-   threads          = 1   # EDIT to change number of OpenMP threads (see above)
+   threads          = @@OpenMPThreads@@   # EDIT to change number of OpenMP threads (see above)
 
 
 #------- Compilers ------------------------------------------------------------
@@ -115,4 +115,4 @@ default=base:
 
 intrate,intspeed=base:
     EXTRA_COPTIMIZE = -fno-strict-aliasing
-    LDCFLAGS        = -z muldefs
+    LDCFLAGS        = -z muldefs

src/bin/runcpu.ts

@@ -31,7 +31,7 @@ const argv = await yargs(hideBin(process.argv))
     })
     .option('optimize-profile', {
         type: 'string',
-        choices: ['O2', 'base'],
+        choices: ['O2', 'base', 'openmp'],
         demandOption: true,
     })
     .option('allow-misaligned', {
@@ -99,6 +99,12 @@ const optimizeProfiles = {
         "-O3",
         "-flto=thin",
     ],
+    openmp: [
+        "-O3",
+        "-flto=thin",
+        "-DSPEC_OPENMP",
+        "-fopenmp",
+    ],
     O2: [
         "-O2",
     ]
@@ -132,20 +138,25 @@ const specProc = await spawnSPECWithID(
             ...argv.compilerSuite == "llvm" ? [sunway.mcpu(sunwayGeneration)] : [], // Only add -mcpu for llvm
         ],
         ldflags: [
-            '-static',
             `-L${path.resolve(prefix, 'lib')}`,
             ...optimizeFlags,
             "-fuse-ld=lld",
             ...["--sysroot", sysroot],
-            ...argv.optimizeProfile === "base" ? [
+            ...["base", "openmp"].includes(argv.optimizeProfile) ? [
                 ...argv.simd ? ["-Wl,-plugin-opt,-mattr=+simd"] : [],
                 ...argv.allowMisaligned ? [`-Wl,-plugin-opt,${sw64UnalignedFlag}`] : [],
                 ...sunwayGeneration === "8a" ? ["-Wl,-plugin-opt,-mattr=+core4"] : [],
             ] : [],
+            ...argv.optimizeProfile === "openmp" ? [
+                `-Wl,-rpath=${argv.compilerPrefix}/lib`,
+            ] : ['-static'],
         ],
         compilerPaths: toolchain(argv.compilerPrefix, argv.compilerSuite as ToolchainSuite),
-        libs: argv.compilerSuite === "llvm" ? cflangLibs(prefix) : [],
+        libs: (argv.compilerSuite === "llvm" || argv.compilerSuite !== "openmp") ? cflangLibs(prefix) : [],
         specialFlags: getSpecialFlags(),
+        openmp: {
+            threads: (argv.optimizeProfile === "openmp" ? 64 : 1),
+        }
     },
     {
         setprocgroup: true,

src/spec/index.ts

@@ -64,6 +64,9 @@ export interface ConfigOptions {
     optimize: string[];
     ldflags: string[];
     libs: string[];
+    openmp: {
+        threads: number;
+    };
     compilerPaths: {
         CXX: string,
         CC: string,
@@ -83,6 +86,7 @@ export const renderConfig = (options: ConfigOptions): string => {
             .replace("@@CompilerFC@@", options.compilerPaths.FC)
             .replace("@@LIBS@@", options.libs.join(" "))
             .replace("@@SpecialFlags@@", options.specialFlags)
+            .replace("@@OpenMPThreads@@", options.openmp.threads.toString())
         ;
 };