Properly provide alignment of 'byval' arguments down to llvm.

[jyknight]

Summary

This is important in the case that the LLVM-inferred llvm-struct
alignment is not the same as the clang-known C-struct alignment.

[majnemer]

LGTM

cfe/trunk/lib/CodeGen/CGCall.cpp

case ABIArgInfo::Extend:
       // FALL THROUGH
     case ABIArgInfo::Direct:
       if (ArgNo == 0 && FI.isChainCall())
         Attrs.addAttribute(llvm::Attribute::Nest);
       else if (AI.getInReg())
         Attrs.addAttribute(llvm::Attribute::InReg);
       break;
 
-    case ABIArgInfo::Indirect:
+    case ABIArgInfo::Indirect: {
       if (AI.getInReg())
         Attrs.addAttribute(llvm::Attribute::InReg);
 
       if (AI.getIndirectByVal())
         Attrs.addAttribute(llvm::Attribute::ByVal);
 
-      Attrs.addAlignmentAttr(AI.getIndirectAlign());
+      unsigned Align = AI.getIndirectAlign();
+
+      // In a byval argument, it is important that the required
+      // alignment of the type is honored, as LLVM might be creating a
+      // *new* stack object, and needs to know what alignment to give
+      // it. (Sometimes it can deduce a sensible alignment on its own,
+      // but not if clang decides it must emit a packed struct, or the
+      // user specifies increased alignment requirements.)
+      //
+      // This is different from indirect *not* byval, where the object
+      // exists already, and the align attribute is purely
+      // informative.
+      if (Align == 0 && AI.getIndirectByVal())
+        Align = getContext().getTypeAlignInChars(ParamType).getQuantity();
+
+      Attrs.addAlignmentAttr(Align);
 
       // byval disables readnone and readonly.
       FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
         .removeAttribute(llvm::Attribute::ReadNone);
       break;
-
+    }
     case ABIArgInfo::Ignore:
     case ABIArgInfo::Expand:
       continue;
 
     case ABIArgInfo::InAlloca:
       // inalloca disables readnone and readonly.
       FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
           .removeAttribute(llvm::Attribute::ReadNone);

cfe/trunk/test/CodeGen/le32-arguments.c

 // RUN: %clang_cc1 -triple le32-unknown-nacl %s -emit-llvm -o - | FileCheck %s
 
 // Basic argument/attribute tests for le32/PNaCl
 
 // CHECK-LABEL: define void @f0(i32 %i, i32 %j, double %k)
 void f0(int i, long j, double k) {}
 
 typedef struct {
   int aa;
   int bb;
 } s1;
 // Structs should be passed byval and not split up
-// CHECK-LABEL: define void @f1(%struct.s1* byval %i)
+// CHECK-LABEL: define void @f1(%struct.s1* byval align 4 %i)
 void f1(s1 i) {}
 
 typedef struct {
   int cc;
 } s2;
 // Structs should be returned sret and not simplified by the frontend
 // CHECK-LABEL: define void @f2(%struct.s2* noalias sret %agg.result)
 s2 f2() {

 // CHECK-LABEL: define void @f6(i32 %a)
 void f6(enum my_enum a) {}
 
 union simple_union {
   int a;
   char b;
 };
 // Unions should be passed as byval structs
-// CHECK-LABEL: define void @f7(%union.simple_union* byval %s)
+// CHECK-LABEL: define void @f7(%union.simple_union* byval align 4 %s)
 void f7(union simple_union s) {}
 
 typedef struct {
   int b4 : 4;
   int b3 : 3;
   int b8 : 8;
 } bitfield1;
 // Bitfields should be passed as byval structs
-// CHECK-LABEL: define void @f8(%struct.bitfield1* byval %bf1)
+// CHECK-LABEL: define void @f8(%struct.bitfield1* byval align 4 %bf1)
 void f8(bitfield1 bf1) {}

cfe/trunk/test/CodeGen/nvptx-abi.c

 // CHECK-LABEL: @bar
 // CHECK: call %struct.float4_s @my_function
   ret = my_function();
   return ret.x;
 }
 
 void foo(float4_t x) {
 // CHECK-LABEL: @foo
-// CHECK: %struct.float4_s* byval %x
+// CHECK: %struct.float4_s* byval align 4 %x
 }
 
 void fooN(float4_t x, float4_t y, float4_t z) {
 // CHECK-LABEL: @fooN
-// CHECK: %struct.float4_s* byval %x
-// CHECK: %struct.float4_s* byval %y
-// CHECK: %struct.float4_s* byval %z
+// CHECK: %struct.float4_s* byval align 4 %x
+// CHECK: %struct.float4_s* byval align 4 %y
+// CHECK: %struct.float4_s* byval align 4 %z
 }
 
 typedef struct nested_s {
   unsigned long long x;
   float z[64];
   float4_t t;
 } nested_t;
 
 void baz(nested_t x) {
 // CHECK-LABEL: @baz
-// CHECK: %struct.nested_s* byval %x)
+// CHECK: %struct.nested_s* byval align 8 %x)
 }

cfe/trunk/test/CodeGen/sparc-arguments.c

+// RUN: %clang_cc1 -triple sparc-unknown-unknown -emit-llvm -o - %s | FileCheck %s
+
+// Ensure that we pass proper alignment to llvm in the call
+// instruction. The proper alignment for the type is sometimes known
+// only by clang, and is not manifest in the LLVM-type. So, it must be
+// explicitly passed through. (Besides the case of the user specifying
+// alignment, as here, this situation also occurrs for non-POD C++
+// structs with tail-padding: clang emits these as packed llvm-structs
+// for ABI reasons.)
+
+struct s1 {
+  int x;
+} __attribute__((aligned(8)));
+
+struct s1 x1;
+
+
+// Ensure the align 8 is passed through:
+// CHECK-LABEL: define void @f1()
+// CHECK: call void @f1_helper(%struct.s1* byval align 8 @x1)
+// Also ensure the declaration of f1_helper includes it
+// CHECK: declare void @f1_helper(%struct.s1* byval align 8)
+
+void f1_helper(struct s1);
+void f1() {
+  f1_helper(x1);
+}