Increase __cilkrts_stack_frame alignment to ABI alignment of stack po…

…inter.

All the __cilkrts_stack_frame objects are aligned on the stack.
There is no cost to increasing their alignment and the compiler
might generate better code.  This alignment is only used within
a module.  The runtime will use whatever value is declared in C code.

llvm/include/llvm/Transforms/Tapir/OpenCilkABI.h

@@ -60,7 +60,12 @@ class OpenCilkABI final : public TapirTarget {
   FunctionCallee CilkRTSCilkForGrainsize32 = nullptr;
   FunctionCallee CilkRTSCilkForGrainsize64 = nullptr;
 
-  MaybeAlign StackFrameAlign{8};
+  // The alignment of __cilkrts_stack_frame.  This value will be increased
+  // to the ABI-prescribed stack pointer alignment of the module.  This has
+  // no runtime cost.  The bitcode file may request even greater alignment
+  // by defining a variable __cilkrts_stack_frame_align.  This may have a
+  // runtime cost by forcing overalignment of stack pointers.
+  MaybeAlign StackFrameAlign;
 
   // Accessors for CilkRTS ABI functions. When a bitcode file is loaded, these
   // functions should return the function defined in the bitcode file.

llvm/lib/Transforms/Tapir/OpenCilkABI.cpp

@@ -309,6 +309,7 @@ void OpenCilkABI::prepareModule() {
     }
     if (GlobalVariable *AlignVar =
         M.getGlobalVariable("__cilkrts_stack_frame_align", true)) {
+      // StackFrameAlign is undefined here.
       StackFrameAlign = AlignVar->getAlign();
       // Mark this variable with private linkage, to avoid linker failures when
       // compiling with no optimizations.
@@ -328,6 +329,12 @@ void OpenCilkABI::prepareModule() {
   if (StackFrameTy->isOpaque()) {
     // Create a dummy __cilkrts_stack_frame structure
     StackFrameTy->setBody(Int64Ty);
+  } else {
+    // Promote the stack frame structure alignment to the largest convenient
+    // value given the ABI.
+    Align ABIStackAlign = M.getDataLayout().getStackAlignment();
+    if (ABIStackAlign > StackFrameAlign.valueOrOne())
+      StackFrameAlign = ABIStackAlign;
   }
   // Create declarations of all CilkRTS functions, and add basic attributes to
   // those declarations.
@@ -875,9 +882,9 @@ void OpenCilkABI::processSubTaskCall(TaskOutlineInfo &TOI, DominatorTree &DT) {
          ConstantPointerNull::get(PointerType::getUnqual(C)));
   ReplCall->setOperand(ParentSFArgNum, SF);
   Argument *ParentSFArg = TOI.Outline->getArg(ParentSFArgNum);
-  ParentSFArg->addAttr(
-      Attribute::getWithAlignment(C, StackFrameAlign.valueOrOne()));
-
+  if (StackFrameAlign)
+    ParentSFArg->addAttr(
+      Attribute::getWithAlignment(C, StackFrameAlign.value()));
   // Split the basic block containing the detach replacement just before the
   // start of the detach-replacement instructions.
   BasicBlock *DetBlock = ReplStart->getParent();