[llvm] [CodeGen] Avoid aligning alloca size. (PR #132064)

[via llvm-commits]

================
@@ -114,7 +114,7 @@ define void @f5() {
 ; CHECK-NEXT:    lgr %r11, %r15
 ; CHECK-NEXT:    .cfi_def_cfa_register %r11
 ; CHECK-NEXT:    lgr %r1, %r15
-; CHECK-NEXT:    aghi %r1, -128
+; CHECK-NEXT:    aghi %r1, -124
 ; CHECK-NEXT:    la %r2, 280(%r1)
 ; CHECK-NEXT:    nill %r2, 65408
 ; CHECK-NEXT:    lgr %r15, %r1
----------------
jcogan-nv wrote:

I see that `SystemZTargetLowering::lowerDYNAMIC_STACKALLOC_ELF` computes the amount subtracted from the stack pointer, called `NeededSpace` like this:

`NeededSpace` = `AllocaSize` + (`RequiredAlign` - `StackAlign`)

With my change, we get 4 + (128 - 8) which is 124. It makes sense that we would want `AllocaSize` to be a multiple of `StackAlign` to ensure that `NeededSpace` is also a multiple of `StackAlign`.

Do you think it is reasonable to add the code rounding the alloca size up to the nearest `StackAlign` to `lowerDYNAMIC_STACKALLOC_ELF`? Another option is to disable my change for targets that don't have a realignable stack.

I think the SPARC backend has the same behavior.

https://github.com/llvm/llvm-project/pull/132064