A game that loads its own x87 control word, such as 0x0c7f (single precision, truncation)

What to know

Many games of the 1990s set the x87 FPU's control word at start-up. Direct3D did it too: it set single precision unless told not to. The control word decides the precision of every result (bits 8 and 9: 24, 53 or 64 bits) and the rounding mode (bits 10 and 11: nearest, down, up, toward zero). Hardwar's maths start-up loads 0x0c7f with fldcw: 24-bit precision, round toward zero, all exceptions masked (observed, FUN_004af9a0). Host floats round to nearest in 53 or 64 bits, so a port's float maths will differ in the last bits from such a game, and the differences grow over a simulation.

• Find every fldcw and what it loads before you trust any float result: ./gameframe run insns GAME.EXE '^fldcw' --functions analysis/export/GAME.EXE/functions.tsv. Read the word each one loads. A helper may change the mode for one conversion and put it back (a rounding helper).
• Decode the word: 0x037f is the default (64-bit, nearest); 0x027f is 53-bit, nearest; 0x007f is 24-bit, nearest; 0x0c7f is 24-bit, toward zero. 0x0f7f is 64-bit, toward zero.
• Measure under the emulator, with the game's own control word set first. GameFrame's emulators run the x87 code itself, so the results are the original's (a mode measured in POD).
• In the port, compute those routines with the same precision and rounding: a soft-float routine that rounds each result to 24 bits toward zero, or the fixed-point form the routine really computes. Mark any other choice as a Deviation:.
• A game can change the word later: a renderer's start-up, a DLL, a timer callback. Look for fldcw in every binary the game loads, and read the word at the points the port measures.