Add --opt=force-aligned-memory option.

This forces all vector loads/stores to be done assuming that the given
pointer is aligned to the vector size, thus allowing the use of sometimes
more-efficient instructions.  (If it isn't the case that the memory is
aligned, the program will fail!).

ispc.cpp

@@ -640,6 +640,7 @@ Opt::Opt() {
     unrollLoops = true;
     disableAsserts = false;
     disableFMA = false;
+    forceAlignedMemory = false;
     disableMaskAllOnOptimizations = false;
     disableHandlePseudoMemoryOps = false;
     disableBlendedMaskedStores = false;

ispc.h

@@ -311,6 +311,12 @@ struct Opt {
         that support them). */
     bool disableFMA;
 
+    /** Always generate aligned vector load/store instructions; this
+        implies a guarantee that all dynamic access through pointers that
+        becomes a vector load/store will be a cache-aligned sequence of
+        locations. */
+    bool forceAlignedMemory;
+
     /** If enabled, disables the various optimizations that kick in when
         the execution mask can be determined to be "all on" at compile
         time. */

main.cpp

@@ -119,6 +119,7 @@ usage(int ret) {
     printf("        disable-loop-unroll\t\tDisable loop unrolling.\n");
     printf("        fast-masked-vload\t\tFaster masked vector loads on SSE (may go past end of array)\n");
     printf("        fast-math\t\t\tPerform non-IEEE-compliant optimizations of numeric expressions\n");
+    printf("        force-aligned-memory\t\tAlways issue \"aligned\" vector load and store instructions\n");
 #ifndef ISPC_IS_WINDOWS
     printf("    [--pic]\t\t\t\tGenerate position-independent code\n");
 #endif // !ISPC_IS_WINDOWS
@@ -336,6 +337,8 @@ int main(int Argc, char *Argv[]) {
                 g->opt.unrollLoops = false;
             else if (!strcmp(opt, "disable-fma"))
                 g->opt.disableFMA = true;
+            else if (!strcmp(opt, "force-aligned-memory"))
+                g->opt.forceAlignedMemory = true;
 
             // These are only used for performance tests of specific
             // optimizations

module.cpp

@@ -1783,6 +1783,8 @@ Module::execPreprocessor(const char* infilename, llvm::raw_string_ostream* ostre
         opts.addMacroDef("ISPC_TARGET_HAS_HALF");
     if (g->target.hasTranscendentals)
         opts.addMacroDef("ISPC_TARGET_HAS_TRANSCENDENTALS");
+    if (g->opt.forceAlignedMemory)
+        opts.addMacroDef("ISPC_FORCE_ALIGNED_MEMORY");
 
     opts.addMacroDef("ISPC_MAJOR_VERSION=1");
     opts.addMacroDef("ISPC_MINOR_VERSION=3");

opt.cpp

@@ -791,7 +791,11 @@ IntrinsicsOpt::runOnBasicBlock(llvm::BasicBlock &bb) {
                                               llvm::PointerType::get(returnType, 0), 
                                               name, callInst);
                     lCopyMetadata(castPtr, callInst);
-                    int align = callInst->getCalledFunction() == avxMaskedLoad32 ? 4 : 8;
+                    int align;
+                    if (g->opt.forceAlignedMemory)
+                        align = 0;
+                    else
+                        align = callInst->getCalledFunction() == avxMaskedLoad32 ? 4 : 8;
                     name = LLVMGetName(callInst->getArgOperand(0), "_load");
                     llvm::Instruction *loadInst = 
                         new llvm::LoadInst(castPtr, name, false /* not volatile */,
@@ -829,7 +833,11 @@ IntrinsicsOpt::runOnBasicBlock(llvm::BasicBlock &bb) {
 
                     llvm::StoreInst *storeInst = 
                         new llvm::StoreInst(rvalue, castPtr, (llvm::Instruction *)NULL);
-                    int align = callInst->getCalledFunction() == avxMaskedStore32 ? 4 : 8;
+                    int align;
+                    if (g->opt.forceAlignedMemory)
+                        align = 0;
+                    else
+                        align = callInst->getCalledFunction() == avxMaskedStore32 ? 4 : 8;
                     storeInst->setAlignment(align);
                     lCopyMetadata(storeInst, callInst);
                     llvm::ReplaceInstWithInst(callInst, storeInst);
@@ -2553,7 +2561,7 @@ lImproveMaskedStore(llvm::CallInst *callInst) {
         lCopyMetadata(lvalue, callInst);
         llvm::Instruction *store = 
             new llvm::StoreInst(rvalue, lvalue, false /* not volatile */,
-                                info->align);
+                                g->opt.forceAlignedMemory ? 0 : info->align);
         lCopyMetadata(store, callInst);
         llvm::ReplaceInstWithInst(callInst, store);
         return true;
@@ -2616,7 +2624,8 @@ lImproveMaskedLoad(llvm::CallInst *callInst,
                                     callInst);
         llvm::Instruction *load = 
             new llvm::LoadInst(ptr, callInst->getName(), false /* not volatile */,
-                               info->align, (llvm::Instruction *)NULL);
+                               g->opt.forceAlignedMemory ? 0 : info->align,
+                               (llvm::Instruction *)NULL);
         lCopyMetadata(load, callInst);
         llvm::ReplaceInstWithInst(callInst, load);
         return true;

test_static.cpp

@@ -99,15 +99,21 @@ void *ISPCAlloc(void **handle, int64_t size, int32_t alignment) {
 }
 
 
+#if defined(_WIN32) || defined(_WIN64)
+#define ALIGN
+#else
+#define ALIGN __attribute__((aligned(64)))
+#endif
 
 int main(int argc, char *argv[]) {
     int w = width();
     assert(w <= 64);
 
-    float returned_result[64];
-    float vfloat[64];
-    double vdouble[64];
-    int vint[64], vint2[64];
+    float returned_result[64] ALIGN;
+    float vfloat[64] ALIGN;
+    double vdouble[64] ALIGN;
+    int vint[64] ALIGN;
+    int vint2[64] ALIGN;
 
     for (int i = 0; i < 64; ++i) {
         returned_result[i] = -1e20;