SelectionDAG: 32 byte aligned store to 16 byte aligned stack generated for no-realign-stack functions

[Nirhar]

Example IR:

target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128-ni:1-p2:32:8:8:32-ni:2"
target triple = "x86_64-unknown-linux-gnu"

define i32 @foo(i32 %arg1) #0 {
entry:
  %a = extractelement <32 x i8> zeroinitializer, i32 %arg1
  %b = zext i8 %a to i32
  ret i32 %b
}

attributes #0 = { "no-realign-stack" "target-cpu"="skylake-avx512" }

And here is the corresponding Assembly :

foo:                                    # @foo
        vxorps  %xmm0, %xmm0, %xmm0
        vmovaps %ymm0, -40(%rsp)
        andl    $31, %edi
        movzbl  -40(%rsp,%rdi), %eax
        vzeroupper
        retq

with the stack slot generated:

Frame Objects:
  fi#0: size=32, align=16, at location [SP+8]

Here is the link to the same example on godbolt: Link

The vmovaps instruction can generate a fault for this example.

[Nirhar]

Perhaps replacing vmovaps with vmovups is a viable solution.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Manish Kausik H (Nirhar)

Example IR: ``` target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128-ni:1-p2:32:8:8:32-ni:2" target triple = "x86_64-unknown-linux-gnu"

define i32 @foo(i32 %arg1) #0 {
bci_0:
%a = extractelement <32 x i8> zeroinitializer, i32 %arg1
%b = zext i8 %a to i32
ret i32 %b
}

attributes #0 = { "no-realign-stack" "target-cpu"="skylake-avx512" }

And here is the corresponding Assembly :

foo: # @foo
vxorps %xmm0, %xmm0, %xmm0
vmovaps %ymm0, -40(%rsp)
andl $31, %edi
movzbl -40(%rsp,%rdi), %eax
vzeroupper
retq

with the stack slot generated:

Frame Objects:
fi#0: size=32, align=16, at location [SP+8]

Here is the link to the same example on godbolt: [Link](https://godbolt.org/z/cYqcPMrYT)

The `vmovaps` instruction can generate a fault for this example.
</details>

[RKSimon]

SelectionDAGLegalize::ExpandExtractFromVectorThroughStack doesn't look to be setting a suitable alignment, so gets the default for the type

[KanRobert]

cc @HaohaiWen

[danilaml]

@RKSimon probably needs a fix similar to 511be2a