[PATCH] D80934: [AIX] Update data directives for AIX assembly

[Xiangling Liao via Phabricator via llvm-commits]

Xiangling_L added inline comments.


================
Comment at: llvm/lib/MC/MCAsmInfoXCOFF.cpp:33
+  // The standard AIX assembly directives auto-align, so they are not usable.
+  Data16bitsDirective = "\t.vbyte\t2, ";
+  Data32bitsDirective = "\t.vbyte\t4, ";
----------------
Just a question here.
I checked the XL behavior, in the following case:

```
  struct {
    int i;
    double n;
  } a[] = {
           {9, 1.5},
  };

```

test.s:

```
        .csect  a{RW}, 3
        .long   0x00000009              # "\0\0\0\t"
        .long   0x3ff80000              # "?\370\0\0"
        .long   0x00000000              # "\0\0\0\0"
```

Under 32-bit mode, XL split a double to two `.long`. So I am wondering can we do the same in LLVM?


================
Comment at: llvm/test/CodeGen/PowerPC/aix-xcoff-data.ll:29
+%struct.anon = type <{ i32, double }>
+ at astruct = global [1 x %struct.anon] [%struct.anon <{ i32 1, double 7.000000e+00 }>], align 1
+
----------------
I am wondering do you make it align to 1 due to any special reason? If not, according to AIX power alignment rule, this struct is supposed to be aligned to 4.

And also, we may want to add a more interesting case where the double is the first member of the struct, which is 8-byte aligned.
e.g.

```
%struct.anon = type <{ double, i32 }>
@astruct = global [1 x %struct.anon] [%struct.anon <{ double 7.000000e+00 , i32 1}>], align 8
```

By doing that, we can see how the padding is added using `vbyte`.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D80934/new/

https://reviews.llvm.org/D80934