Emit .align 1 by default if unspecified

Change to explicitly emit .align 1 for both CL and CUDA if alignment
is unspecified. In most cases, the frontend should already specify
alignment, so this should ideally not change much in practice.

Author: LewisCrawford

llvm/lib/Target/NVPTX/NVPTXAsmPrinter.cpp

@@ -1600,24 +1600,27 @@ void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
 
       if (isKernelFunc) {
         if (PTy) {
-          O << "\t.param .u" << PTySizeInBits << " .ptr ";
-
-          const unsigned AddrSpace = PTy->getAddressSpace();
-          if (AddrSpace != ADDRESS_SPACE_GENERIC) {
-            O << ".";
-            emitPTXAddressSpace(AddrSpace, O);
-            O << " ";
+          O << "\t.param .u" << PTySizeInBits << " .ptr";
+
+          switch (PTy->getAddressSpace()) {
+          default:
+            break;
+          case ADDRESS_SPACE_GLOBAL:
+            O << " .global";
+            break;
+          case ADDRESS_SPACE_SHARED:
+            O << " .shared";
+            break;
+          case ADDRESS_SPACE_CONST:
+            O << " .const";
+            break;
+          case ADDRESS_SPACE_LOCAL:
+            O << " .local";
+            break;
           }
 
-          const bool IsCUDA =
-              static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() ==
-              NVPTX::CUDA;
-
-          MaybeAlign ParamAlign = I->getParamAlign();
-          if (ParamAlign.has_value() || !IsCUDA)
-            O << ".align " << ParamAlign.valueOrOne().value() << " ";
-
-          O << TLI->getParamName(F, paramIndex);
+          O << " .align " << I->getParamAlign().valueOrOne().value();
+          O << " " << TLI->getParamName(F, paramIndex);
           continue;
         }
 

llvm/test/CodeGen/NVPTX/i1-param.ll

@@ -8,7 +8,7 @@ target triple = "nvptx-nvidia-cuda"
 
 ; CHECK: .entry foo
 ; CHECK:   .param .u8 foo_param_0
-; CHECK:   .param .u64 .ptr foo_param_1
+; CHECK:   .param .u64 .ptr .align 1 foo_param_1
 define void @foo(i1 %p, ptr %out) {
   %val = zext i1 %p to i32
   store i32 %val, ptr %out

llvm/test/CodeGen/NVPTX/kernel-param-align.ll

@@ -21,12 +21,12 @@ entry:
 }
 
 ; CHECK-LABEL: .entry func_noalign(
-; CHECK: .param .u64 .ptr         func_noalign_param_0
-; CHECK: .param .u64 .ptr         func_noalign_param_1
-; CHECK: .param .u64 .ptr .global func_noalign_param_2
-; CHECK: .param .u64 .ptr .shared func_noalign_param_3
-; CHECK: .param .u64 .ptr .const  func_noalign_param_4
-; CHECK: .param .u64 .ptr .local  func_noalign_param_5
+; CHECK: .param .u64 .ptr         .align 1 func_noalign_param_0
+; CHECK: .param .u64 .ptr         .align 1 func_noalign_param_1
+; CHECK: .param .u64 .ptr .global .align 1 func_noalign_param_2
+; CHECK: .param .u64 .ptr .shared .align 1 func_noalign_param_3
+; CHECK: .param .u64 .ptr .const  .align 1 func_noalign_param_4
+; CHECK: .param .u64 .ptr .local  .align 1 func_noalign_param_5
 define void @func_noalign(ptr nocapture readonly %input,
                           ptr nocapture %out,
                           ptr addrspace(1) %global,