Added structure alignment in headers; extended the test system to support alignment tests

module.cpp

@@ -1486,22 +1486,49 @@ lEmitStructDecl(const StructType *st, std::vector<const StructType *> *emittedSt
     fprintf(file, "#ifndef __ISPC_STRUCT_%s__\n",st->GetCStructName().c_str());
     fprintf(file, "#define __ISPC_STRUCT_%s__\n",st->GetCStructName().c_str());
 
-    fprintf(file, "struct %s", st->GetCStructName().c_str());
+    char sSOA[48];
+    bool pack, needsAlign = false;
+    llvm::Type *stype = st->LLVMType(g->ctx);
+    llvm::DataLayout *DL = g->target->getDataLayout();
+
+    if (!(pack = llvm::dyn_cast<llvm::StructType>(stype)->isPacked()))
+        for (int i = 0; !needsAlign && (i < st->GetElementCount()); ++i) {
+            const Type *ftype = st->GetElementType(i)->GetAsNonConstType();
+            needsAlign |= ftype->IsVaryingType()
+                       && (CastType<StructType>(ftype) == NULL);
+        }
     if (st->GetSOAWidth() > 0)
         // This has to match the naming scheme in
         // StructType::GetCDeclaration().
-        fprintf(file, "_SOA%d", st->GetSOAWidth());
-    fprintf(file, " {\n");
-
+        sprintf(sSOA, "_SOA%d", st->GetSOAWidth());
+    else
+        *sSOA = '\0';
+    if (!needsAlign)
+        fprintf(file, "%sstruct %s%s {\n", (pack)? "packed " : "",
+                      st->GetCStructName().c_str(), sSOA);
+    else {
+        unsigned uABI = DL->getABITypeAlignment(stype);
+        fprintf(file, "__ISPC_ALIGNED_STRUCT__(%u) %s%s {\n", uABI,
+                      st->GetCStructName().c_str(), sSOA);
+    }
     for (int i = 0; i < st->GetElementCount(); ++i) {
-        const Type *type = st->GetElementType(i)->GetAsNonConstType();
-        std::string d = type->GetCDeclaration(st->GetElementName(i));
-	// Don't expand struct members as their insides will be expanded.
-        if (type->IsVaryingType() && (CastType<StructType>(type) == NULL)) {
-          fprintf(file, "    %s[%d];\n", d.c_str(), g->target->getVectorWidth());
+        const Type *ftype = st->GetElementType(i)->GetAsNonConstType();
+        std::string d = ftype->GetCDeclaration(st->GetElementName(i));
+
+        fprintf(file, "    ");
+        if (needsAlign && ftype->IsVaryingType() &&
+           (CastType<StructType>(ftype) == NULL)) {
+            unsigned uABI = DL->getABITypeAlignment(ftype->LLVMType(g->ctx));
+            fprintf(file, "__ISPC_ALIGN__(%u) ", uABI);
+        }
+        // Don't expand arrays, pointers and structures:
+        // their insides will be expanded automatically.
+        if (!ftype->IsArrayType() && !ftype->IsPointerType() &&
+            ftype->IsVaryingType() && (CastType<StructType>(ftype) == NULL)) {
+            fprintf(file, "%s[%d];\n", d.c_str(), g->target->getVectorWidth());
         }
         else {
-          fprintf(file, "    %s;\n", d.c_str());
+            fprintf(file, "%s;\n", d.c_str());
         }
     }
     fprintf(file, "};\n");
@@ -1515,8 +1542,22 @@ lEmitStructDecl(const StructType *st, std::vector<const StructType *> *emittedSt
 static void
 lEmitStructDecls(std::vector<const StructType *> &structTypes, FILE *file, bool emitUnifs=true) {
     std::vector<const StructType *> emittedStructs;
+
+    fprintf(file,
+            "\n#ifndef __ISPC_ALIGN__\n"
+            "#if defined(__clang__) || !defined(_MSC_VER)\n"
+            "// Clang, GCC, ICC\n"
+            "#define __ISPC_ALIGN__(s) __attribute__((aligned(s)))\n"
+            "#define __ISPC_ALIGNED_STRUCT__(s) struct __ISPC_ALIGN__(s)\n"
+            "#else\n"
+            "// Visual Studio\n"
+            "#define __ISPC_ALIGN__(s) __declspec(align(s))\n"
+            "#define __ISPC_ALIGNED_STRUCT__(s) __ISPC_ALIGN__(s) struct\n"
+            "#endif\n"
+            "#endif\n\n");
+
     for (unsigned int i = 0; i < structTypes.size(); ++i)
-      lEmitStructDecl(structTypes[i], &emittedStructs, file, emitUnifs);
+        lEmitStructDecl(structTypes[i], &emittedStructs, file, emitUnifs);
 }
 
 

run_tests.py

@@ -194,8 +194,8 @@ def run_test(testname):
 
         # We need to figure out the signature of the test
         # function that this test has.
-        sig2def = { "f_v(" : 0, "f_f(" : 1, "f_fu(" : 2, "f_fi(" : 3, 
-                    "f_du(" : 4, "f_duf(" : 5, "f_di(" : 6 }
+        sig2def = { "f_v(" : 0, "f_f(" : 1, "f_fu(" : 2, "f_fi(" : 3,
+                    "f_du(" : 4, "f_duf(" : 5, "f_di(" : 6, "f_sz" : 7 }
         file = open(filename, 'r')
         match = -1
         for line in file:
@@ -310,16 +310,16 @@ def run_test(testname):
                   ispc_cmd = ispc_exe_rel + " --woff %s -o %s -O3 --emit-asm --target=%s" % \
                          (filename4ptx, obj_name, options.target)
 
-
-             
-
         # compile the ispc code, make the executable, and run it...
+        ispc_cmd += " -h " + filename + ".h"
+        cc_cmd += " -DTEST_HEADER=<" + filename + ".h>"
         (compile_error, run_error) = run_cmds([ispc_cmd, cc_cmd], 
                                               options.wrapexe + " " + exe_name, \
                                               testname, should_fail)
 
         # clean up after running the test
         try:
+            os.unlink(filename + ".h")
             if not options.save_bin:
                 if not run_error:
                     os.unlink(exe_name)

test_static.cpp

@@ -51,6 +51,16 @@
 #include <malloc.h>
 #endif
 
+#if (TEST_SIG == 7)
+    #define varying_f_sz f_sz
+    #define v1_varying_f_sz f_sz
+    #define v2_varying_f_sz f_sz
+    #define v4_varying_f_sz f_sz
+    #define v8_varying_f_sz f_sz
+    #define v16_varying_f_sz f_sz
+    #include TEST_HEADER
+#endif
+
 extern "C" {
     extern int width();
     extern void f_v(float *result);
@@ -103,7 +113,7 @@ void *ISPCAlloc(void **handle, int64_t size, int32_t alignment) {
 
 
 #if defined(_WIN32) || defined(_WIN64)
-#define ALIGN
+#define ALIGN __declspec(align(64))
 #else
 #define ALIGN __attribute__((aligned(64)))
 #endif
@@ -112,11 +122,11 @@ int main(int argc, char *argv[]) {
     int w = width();
     assert(w <= 64);
 
-    float returned_result[64] ALIGN;
-    float vfloat[64] ALIGN;
-    double vdouble[64] ALIGN;
-    int vint[64] ALIGN;
-    int vint2[64] ALIGN;
+    ALIGN float returned_result[64];
+    ALIGN float vfloat[64];
+    ALIGN double vdouble[64];
+    ALIGN int vint[64];
+    ALIGN int vint2[64];
 
     for (int i = 0; i < 64; ++i) {
         returned_result[i] = -1e20;
@@ -142,6 +152,8 @@ int main(int argc, char *argv[]) {
     f_duf(returned_result, vdouble, 5.f);
 #elif (TEST_SIG == 6)
     f_di(returned_result, vdouble, vint2);
+#elif (TEST_SIG == 7)
+    *returned_result = sizeof(ispc::f_sz);
 #else
 #error "Unknown or unset TEST_SIG value"
 #endif

tests/align1.ispc

@@ -0,0 +1,22 @@
+struct InnerUniform {
+    uniform int i;
+};
+
+struct f_sz {
+    uniform int type, temp[5];
+    int k[3];
+    InnerUniform u;
+};
+
+
+
+export void keep_struct_declared(varying f_sz * uniform s) {
+}
+
+export uniform int width() {
+    return 1;
+}
+
+export void result(uniform float RET[]) {
+    RET[0] = sizeof(f_sz);
+}

tests/align2.ispc

@@ -0,0 +1,21 @@
+struct InnerVec {
+    int i;
+};
+
+struct f_sz {
+    InnerVec u;
+    uniform int tail;
+};
+
+
+
+export void keep_struct_declared(varying f_sz * uniform s) {
+}
+
+export uniform int width() {
+    return 1;
+}
+
+export void result(uniform float RET[]) {
+    RET[0] = sizeof(f_sz);
+}