Function calling convention: Unnecessary MOVSP instruction (LLVM-65)

[inetbowser]

When using Zig (https://github.com/ziglang/zig) with the Xtensa LLVM backend, I noticed that all functions use the calling convention below. The code works flawlessly but I think the MOVSP instruction can be omitted because the register window of the function guarantees the stack pointer will stay the same.

Code (compiled with Zig and Xtensa LLVM backend):

40080984 <testfn>:
40080984:	007136        	entry	a1, 56
40080987:	017d      	mov.n	a7, a1
40080989:	...    // some code
400809ce:	001710        	movsp	a1, a7
400809d1:	f01d      	retw.n

When compiling some function with the standard GCC C compiler a function is structured something like this.

Code (compiled with xtensa-esp32-elf-gcc):

004001a4 <testfn>:
  4001a4:	006136        	entry	a1, 48
  4001a7:	017d      	mov.n	a7, a1
  4001a9:	...    // some code
  4001d2:	f01d      	retw.n

As you can see there is no MOVSP instruction.

Although this is not a critical bug (it doesn't crash the program nor introduce undefined behaviour), this can decrease performance of a program significantly because the MOVSP instruction can cause an Alloca exception (which could hit performance quite hard).

If I'm mistaken at some point, feel free to let me know. I'm quite a newbie with the Xtensa architecture. 😄

[andreisfr]

Hi, @inetbowser ! Thank you for your comment. Current Xtensa backend can generate some redundant instructions, including movsp. This is not a critical bug, as you mentioned, but a reason to improve code generation) We plan to work on code generation quality (both size and speed) in nearest future. So, any observations or some tricks, which could help to generate more effective code are very helpful.