clang generates incorrect code when storing to memory location 0x1 for target avr-none

[tomtor]

sample code:

#define PA      0x1 // Bad
//#define PA    0x2 // Good

volatile char *const PORT = (char *const) PA;

extern void nil(short);

void bug()
{
  short s= 0;
 
  while (1) {
    s++;
    *PORT = *PORT | 0x80;
    nil((s & 0xF) + 1);
  }
}

compiled with clang build from git yesterday:

clang -O2 -target avr-none -mmcu=attiny402 -S bug.c

generates an incorrect load (from 0x0 via X+) and store (uses expression based on s which changes in each loop iteration instead of constant 0x1):

bug:                                    ; @bug
; %bb.0:                                ; %entry
        push    r16
        push    r17
        ldi     r26, 0
        ldi     r27, 0
.LBB0_1:                                ; %while.cond
                                        ; =>This Inner Loop Header: Depth=1
        ld      r24, X+
        ori     r24, -128
        movw    r16, r26
        andi    r26, 15
        andi    r27, 0
        st      X+, r24
        movw    r24, r26
        call    nil
        movw    r26, r16
        rjmp    .LBB0_1

using #define PA 0x2

generates correct code with sbi instruction:

bug:                                    ; @bug
; %bb.0:                                ; %entry
        push    r16
        push    r17
        ldi     r24, 0
        ldi     r25, 0
.LBB0_1:                                ; %while.cond
                                        ; =>This Inner Loop Header: Depth=1
        sbi     2, 7
        adiw    r24, 1
        movw    r16, r24
        andi    r24, 15
        andi    r25, 0
        adiw    r24, 1
        call    nil
        movw    r24, r16
        rjmp    .LBB0_1

[tomtor]

@benshi001 The same approach as for #59914 works for this bug, simply disabling

AVRTargetLowering::getPostIndexedAddressParts

Before searching for a more elegant fix,
I am searching for code fragments where AVRTargetLowering::getPostIndexedAddressParts really produces better code when enabled. Have you any suggestions?

[tomtor]

Ik wrote the following script to compare 2 compiler versions, with and without the post index code.

Very few effective optimizations.

OPT="/home/tom/src/rust/llvm-project/build/bin/clang-opt"
NOPT="/home/tom/src/rust/llvm-project/build/bin/clang-noopt"
O="-Os"
A="-target avr-none"
M="-mmcu=avr128db28"

for f in gcc/gcc/testsuite/gcc.c-torture/execute/*.c
do
  echo "$(wc -l $f)" | tee -a torture.log
  base="$(basename $f .c)"
  asm="$base.s"
  if $OPT $O $A $M -S $f;
  then
    wc -l $asm | tee -a torture.log
    $NOPT $O $A $M -S $f
    wc -l $asm | tee -a torture.log
    rm $asm
  else
    rm -f /tmp/$base.c /tmp/$base.sh
  fi
  echo "" >> torture.log
  sleep 1
done

[benshi001]

@benshi001 The same approach as for #59914 works for this bug, simply disabling

AVRTargetLowering::getPostIndexedAddressParts

Before searching for a more elegant fix, I am searching for code fragments where AVRTargetLowering::getPostIndexedAddressParts really produces better code when enabled. Have you any suggestions?

Currently I have not a better suggestion, since I have not considered its impact to real world AVR programs.

I would like to find the underlying root cause for this issue, since it may affects other programs, not only your simple example.