Fix SQRT.S operand selection in the FPU translator

[yanm1103]

Problem

FpuTranslator reads the wrong operand for SQRT.S:

case COP1_S_SQRT: return fmt::format("ctx->f[{}] = FPU_SQRT_S(ctx->f[{}]);", fd, fs);

On the R5900, SQRT.S fd, ft takes its source from the ft field; the fs field is 0 in the encoding. For example sqrt.s $f2, $f2 is the word 0x46020084 (fmt=S, ft=2, fs=0, fd=2, funct=4), and it is translated to ctx->f[2] = FPU_SQRT_S(ctx->f[0]);, the square root of the wrong register.

Fix

Read ft. FPU_SQRT_S (ps2_fsqrt, from #272) already takes the magnitude of its input, so nothing else changes.

Scope

This PR originally also fixed RSQRT.S (1/sqrt(fs) instead of fs/sqrt(ft)). #272 fixed that part in the meantime, so it has been rebased onto current main and reduced to SQRT.S only. #218 proposed the same SQRT.S change earlier.

Verification

• Added a ps2x_tests case in code_generator_tests.cpp for sqrt.s $f2, $f2 (0x46020084). Full suite on Linux (GCC, Release): 446/446 pass with the fix; without it, this new case is the only failure.
• Found while recompiling a retail PS2 game whose ELF has sqrt.s in 108 functions. An original function and an independent reimplementation, both recompiled with PS2Recomp, disagreed in 300/300 random states without the fix and matched 300/300 with it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HhodgessnwK5k3pipLwFp3