PowerPC comparisons returned or stored in variables decompile poorly

[Pokechu22]

Describe the bug
PowerPC comparisons returned or stored in variables decompile poorly, as they use countLeadingZeros(v) >> 5 as a way to check if v is 0 (as if so, v has 32 bits set to 0, and 32 >> 5 == 1).

To Reproduce
Decompile the attached binaries, using PowerPC:BE:32:default:default as the language. For instance, equals3_32 in cntlz_O1.o looks like this:

uint equals3_32(uint param_1)

{
  uint uVar1;
  
  uVar1 = countLeadingZeros(param_1 ^ 3);
  return uVar1 >> 5;
}

The 64-bit version looks sillier, producing this (note that I had to manually set the data type and set custom storage to r3 and r4 for this; by default ghidra detects it as 2 uint parameters and with just longlong it tries to place it on the stack):

uint equals3_64(longlong param_1)

{
  uint uVar1;
  
  uVar1 = countLeadingZeros((uint)((ulonglong)param_1 >> 0x20) | (uint)param_1 ^ 3);
  return uVar1 >> 5;
}

The unoptimized versions have various sign-extensions and masks that don't show up when optimisations are turned on:

uint equals3_8(char param_1)

{
  uint uVar1;
  
  uVar1 = countLeadingZeros((int)param_1 ^ 3);
  return uVar1 >> 5 & 0xff;
}

Expected behavior
The decompiled code should look closer to the original code, ideally generating param_1 == 3 in all cases (but something like param_1 ^ 3 == 0 would still be a major improvement).

Attachments
cntlz.zip — contains .o files at -O0 through -O3 with and without -g (though -O1, -O2, and -O3 are all the same without -g) compiled using powerpc-eabi-gcc provided via devkitPPC. (I release this code under CC0.)

#include <stdbool.h>
#include <stdint.h>

bool equals3_8 (int8_t  param) { return param == 3; }
bool equals3_16(int16_t param) { return param == 3; }
bool equals3_32(int32_t param) { return param == 3; }
bool equals3_64(int64_t param) { return param == 3; }

Environment (please complete the following information):

• OS: Windows 10, insider build 19042
• Java Version: 11.0.3
• Ghidra Version: 9.2.2
• Ghidra Origin: official ghidra-sre.org distro

Additional context
I have also seen variable - 3 instead of variable ^ 3, but this doesn't seem to happen with GCC.

This relates to #2121 (though I don't think adding a PcodeOp for countLeadingZeros would fix it on its own, nor am I sure that a fix would actually require adding a PcodeOp).

[ryanmkurtz]

Fixed by 9cf60fa