AArch64_be generate diff code for int8_t with AArch64_le?

Code Generation AArch64
,
1

I test follow code

#include <stdint.h>
int foo(int8_t var1)
{
    return var1 + var1;
}

which in godblot aarch64_be clang 15.0.0 generate follow assemble

foo:                                    // @foo
        sub     sp, sp, #16
        strb    w0, [sp, #12]
        ldrsb   w8, [sp, #12]
        add     w0, w8, w8
        add     sp, sp, #16
        ret

But, in the godblot aarch64 le clang 15.0.0 generate follow

foo:                                    // @foo
        sub     sp, sp, #16
        strb    w0, [sp, #15]
        ldrsb   w8, [sp, #15]
        ldrsb   w9, [sp, #15]
        add     w0, w8, w9
        add     sp, sp, #16
        ret

The diff is the instruction strb store param var1 to 1(16-15) byte or 4(16-12) bytes which the type of var1 is int8_t. So, I think the code generated by aarch64_le is more reasonable than aarch64_be.

I want to know how to fix it, or which part of code impact this behavior.

2

This is somewhere in GlobalISel-specific code; compiling both with -fno-global-isel yields identical code for both.

3

Hi, I found in the clang trunk still generate 15 even set the option -fno-global-isel. I wonder whether have some new changes impact this, and is which commit?

4

So does aarch64-linux-gnu. The point is that big-endian and little-endian only diverge when GlobalISel is enabled.

5

BigEndian is not a thing:

It will fall back to SDAG.

6

But in the source code, the aarch64 not impl the function enableBigEndian, so the enableBigEndian default return false. Is somewhere else code impact this behavior?

7

You can try:

cc -Rpass-missed='gisel*' -fglobal-isel .... foo.c

It will give you remarks about what happened.

8

Hi, I try it. And report

clang-15: warning: -fglobal-isel support for the 'aarch64_be' architecture is incomplete [-Wglobal-isel]
foo.c:28:1: remark: unable to translate in big endian mode [-Rpass-missed=gisel-irtranslator]
{
^
warning: Instruction selection used fallback path for foo [-Wbackend-plugin]
1 warning generated.
9

New progress：in clang aarch64_be trunk the instruction strb store param var1 to 1(16-15) byte, and in clang aarch64_be 15.0.0 is 4(16-12) bytes.
I guess have a commit fixed this problem, but I don’t know which. Maybe I can learn some thing from it.

10

It would be unsurprising if GlobalISel has big endian bugs

11

+1.I remember that somebody did BE fixes in GlobalISel. I forgot who and didn’t found it anymore.

12

Yeap， I see this commit, and also override the method for aarch64 like:

bool enableBigEndian() const { return true; }

A lot of problems is coming. :joy:

Why big endian is so unwelcome?

13

Because so far all of GlobalISel’s supported architectures have been little-endian, and so there are likely places where GlobalISel accidentally assumes the architecture is little-endian rather than checking and doing something different for big-endian.