Ghidra: Decompiler misinterprets `if (double_param != 0) {...}` as dead code

[YellowAfterlife]

Describe the bug
If a function takes a 64-bit float parameter and then proceeds to compare it to 0 (MOVSD->UCOMISD->LAHF->TEST->JNP), Ghidra considers the non-0 branch to be dead code and for parameter to be unused.

To Reproduce

1. Compile the following DLL from a fresh unmanaged C++ DLL project in VS2019 with platform tools v142 and targeting x86:

#include "pch.h"
extern "C" __declspec(dllexport) void __cdecl test(double enable, HWND hwnd) {
    if (enable != 0) {
        MessageBoxA(hwnd, "Not 0!", "", MB_OK);
    } else {
        MessageBoxA(hwnd, "0!", "", MB_OK|MB_ICONERROR);
    }
}
BOOL APIENTRY DllMain(HMODULE hModule, DWORD  ul_reason_for_call, LPVOID lpReserved) {
    return TRUE;
}

1. Create a Ghidra project, add the DLL to it, and run analyzer with default settings
2. Navigate to the test function in Exports and take a look at decompile view.

You get this:

void __cdecl _test(undefined4 param_1,undefined4 param_2,HWND param_3)

{
                    /* 0x1000  1  test
                       Symbol Ref: _test */
  if (false) {
    MessageBoxA(param_3,"Not 0!","",0);
    return;
  }
  MessageBoxA(param_3,"0!","",0x10);
  return;
}

(then-branch is considered to be dead code; 8-byte parameter is interpreted as two unknown4s)

Expected behavior
Recognizing MOSD/UCOMISD as a sign of the parameter being a 64-bit float, or, failing that, showing a comparison to literal's raw floating-point value.

Attachments
Pre-compiled version of the sample DLL

Environment (please complete the following information):

• OS: Win10, build 17134
• Java Version: 12.0.1
• Ghidra Version: 9.0.4

Additional context
The shared trait between this and my original code is use of MOVSD+UCOMISD for value != 0 comparisons, perhaps this is the source of trouble

        10001000 55              PUSH       EBP
        10001001 8b ec           MOV        EBP,ESP
        10001003 f2 0f 10        MOVSD      XMM0,qword ptr [EBP + param_1]
                 45 08
        10001008 66 0f 2e        UCOMISD    XMM0,qword ptr [__real@0000000000000000]
                 05 b0 20 
                 00 10
        10001010 9f              LAHF
        10001011 f6 c4 44        TEST       AH,0x44
        10001014 7b 17           JNP        LAB_1000102d
...

(from sample)

[mumbel]

assume you tested with a correct prototype, does that change anything?

[YellowAfterlife]

Hello, sorry, I forgot to include this in the issue text after moving things around - updating the function signature to match expected types doesn't seem to do anything for it,

void __cdecl _test(double val,HWND hwnd)

{
                    /* 0x1000  1  test
                       Symbol Ref: _test */
  if (false) {
    MessageBoxA(hwnd,"Not 0!","",0);
    return;
  }
  MessageBoxA(hwnd,"0!","",0x10);
  return;
}

[mumbel]

~Ah, looks that instruction has a todo~
wish github search was better, searching for UCOMISD goes to VUCOMISD, which does have todo, but UCOMISD looks implemented

[YellowAfterlife]

I think mine might be the same thing as this previously described issue https://github.com/NationalSecurityAgency/ghidra/issues/122

[caheckman]

This problem was caused by the parity flag not being modeled for x86. This is now fixed in master and the upcoming 9.2