X86: Minor regression from llvm 14 with single array access and lea

[jaskarth]

Description

When arrays are accessed once in a function, llvm14 produces a mov instruction that references the array in memory. However, trunk produces a lea that moves the array into a register, then performs the mov. This change makes sense when the array is accessed multiple times as it removes the extraneous memory reference, but it adds an additional unneeded instruction when there is only a single access of the array.

Code

Compiling with: -O3
Godbolt: https://godbolt.org/z/W4saM9fha

constexpr const static int array[5] = {1, 2, 3, 4, 5};

int get(int a) {
    return array[a];
}

int get2(int a, int b) {
    return array[a] + array[b];
}

trunk codegen

get(int):                                # @get(int)
        movsxd  rax, edi
        lea     rcx, [rip + array]
        mov     eax, dword ptr [rcx + 4*rax]
        ret
get2(int, int):                              # @get2(int, int)
        movsxd  rcx, edi
        lea     rdx, [rip + array]
        movsxd  rax, esi
        mov     eax, dword ptr [rdx + 4*rax]
        add     eax, dword ptr [rdx + 4*rcx]
        ret

llvm 14 codegen

get(int):
        movsxd  rax, edi
        mov     eax, dword ptr [4*rax + array]
        ret
get2(int, int):
        movsxd  rcx, edi
        movsxd  rax, esi
        mov     eax, dword ptr [4*rax + array]
        add     eax, dword ptr [4*rcx + array]
        ret

[topperc]

The use of rip in the address indicates that we've changed how the address of the array is being computed. Trunk is using a position independent address, but llvm-14 isn't. I'm guessing this is related to -fpie being made the default by https://reviews.llvm.org/D120305

Can you try with -fno-pie on the clang command line?

[llvmbot]

@llvm/issue-subscribers-backend-x86

[jaskarth]

Oh yep, that does seems to change it to what llvm 14 produces! Would this still qualify as a bug, then?

[topperc]

Adding @MaskRay

[MaskRay]

This is not a regression. It is a -fno-pic vs -fpie code generation difference. See #57449 (comment) and compiler-explorer/gcc-cross-builder#33