Win64 Does not codegen correctly alloca align 64

[ZviRackover]

Bugzilla Link	8922
Resolution	FIXED
Resolved on	Mar 09, 2011 05:11
Version	trunk
OS	Windows XP
Blocks	llvm/llvm-bugzilla-archive#9100
Attachments	Fix candidate patch, Reproducer test
CC	@asl

Extended Description

The following test case produces the assertion below when running on the Win64 subtarget:

define noalias i32* @​factorial() nounwind readnone {
entry:
%a = alloca i32, align 64
ret i32* %a
}

Assertion failed: (-(Offset + StackSize)) % Align == 0, file ........\lib\Target\X86\X86FrameInfo.cpp, line 867

Explanation:

Win64 CodeGen adds an extra 32 bytes to frame size, to be used as a Red Zone for potential calls. This addition ruins the stack alignment in case the required alignment is greater than 32 bytes.

Proposed fix is to add max(32, MaxAlignment) bytes instead.
Attached are the testcase and patch.

[llvmbot]

I am planning to eliminate the block you attempt to patch.

See also;
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20101213/113852.html

Improved output is below;

factorial:
pushq %rbp
movq %rsp, %rbp
andq $-64, %rsp
subq $64, %rsp
leaq (%rsp), %rax
movq %rbp, %rsp
popq %rbp
ret

[ZviRackover]

I am planning to eliminate the block you attempt to patch.

See also;
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20101213/113852.html

Improved output is below;

factorial:
pushq %rbp
movq %rsp, %rbp
andq $-64, %rsp
subq $64, %rsp
leaq (%rsp), %rax
movq %rbp, %rsp
popq %rbp
ret

This looks right.
Please commit the attached test case as an indicator for this defect. It can remain as XFAIL until you commit your patch.

[llvmbot]

    andq    $-64, %rsp
    subq    $64, %rsp
    leaq    (%rsp), %rax

It would be suboptimal. We may do;

subq $4, %rsp ; or -8
andq $-64, %rsp
movq %rsp, %rax

It seems x86-codegen-generic issue. I will file one.

Anyway, I will add Zvi's testcase.

[llvmbot]

It has been fixed in r124949.
I have to file suboptimal alignment issue, thank you!

[llvmbot]

mentioned in issue llvm/llvm-bugzilla-archive#9100