cmd/internal/obj: wrong code generated by the arm assembler

[benshi001]

The MULA is wrongly assembled by the ARM assembler, even the newest go.

For example, MULA r1, r2, r3, r4 is assembled to e0 23 41 92, which should be e0 24 31 92.

The addend is r3 and the result is r4, the assembler confuses them.

[benshi001]

While MULAWT is right,

MULAWT -> e1 24 31 c2

[benshi001]

I found this issue when tried to add test cases for my CL https://go-review.googlesource.com/#/c/35565/

[benshi001]

I have fixed this issue in patch set 15 of https://go-review.googlesource.com/#/c/35565/

The compiler also needs modification, since it followed the assembler's error in MULA.

[gopherbot]

CL https://golang.org/cl/37021 mentions this issue.

[gopherbot]

CL https://golang.org/cl/35565 mentions this issue.

[cherrymui]

The addend is r3 and the result is r4, the assembler confuses them.

It would be good and consistent that the last register is the destination. But I don't think this is documented and required. MULA has been there for a long time and there are probably user assembly code that rely on this encoding. Maybe just leave it as is?

[benshi001]

But MULAWB MULAWT use the last register as destination. That would confuse the user.

Maybe we would add other MULA serial instructions, like

Multiply and Subtract MLS 32 = 32 – 32 × 32
Signed Multiply Accumulate (halfwords) SMLABB 32 = 32 + 16 × 16
Signed Most Significant Word Multiply Accumulate SMMLA 32 = 32 + 32 × 32

How about these ones? Use the first as destination like MULA or use the last like MULAWB / MULAWT?

[benshi001]

Now I also agree with Cherry that we keep current circumstance unchanged.

Since a change in the assembler will cause
either changes in the compiler
or changes in the disassembler (otherwise test of MULA in cmd/asm/internal/asm/testdata/arm.s will fail)

For others MULS, MMULA, MMULS, MULABB, MULAWT, MULAWB, they can follow MULA, use the third register as destination and the fourth as the addend.

[cherrymui]

No, I didn't mean that we should change everything to the "wrong" MULA encoding.

MULAWT, MULAWB are in the "correct" encoding, which uses last register as destination. They should keep unchanged. The new instructions should follow the correct encoding.

I think we'd want MULA fixed. But it needs to be done more carefully, since there might be user code that uses it.

[benshi001]

If we want everything be in the right order, the disassembler / objdump also need change. Otherwise tests in arm.s will fail, since it check both the output's 32-bit binary code and the disassembled text.

[benshi001]

Now CL https://go-review.googlesource.com/#/c/35565/ has nothing to do with this issue. Since I
removed all changes to MULA/MULAWT/MULAWB, and only make the new MULA like instructions follow MULAWT in the right form.

And I will create a new CL to fix this issue.