Floating Point

Floating Point in CupidOS

CupidOS supports hardware x87 + SSE/SSE2 floating point end-to-end: kernel, language, assembler, libm, printf.

Overview

• Build: -mfpmath=sse -msse -msse2 -mstackrealign in CFLAGS.
• Init: fpu_init() is the first call in kmain. Sets CR0.EM=0/MP=1/NE=1/TS=0 and CR4.OSFXSR=1/OSXMMEXCPT=1, runs FNINIT, loads MXCSR=0x1F80.
• Context switch: context_switch.asm FXSAVEs the outgoing PCB's fp_state[512] and FXRSTORs the incoming task's. Eager: no CR0.TS games.
• PCB: process_t.fp_state is 512 bytes at 16-byte alignment. Offset 80.
• Exceptions: IDT vectors 7 (#NM), 16 (#MF), 19 (#XF) route to fpu_nm_handler/fpu_mf_handler/fpu_xf_handler, which call panic_fpu with a serial dump of FSW/FCW/MXCSR. Should never fire under masked MXCSR + eager switch.

CupidC FP types

• float - 4 bytes, SSE scalar.
• double - 8 bytes, SSE scalar.
• float4 - 16 bytes, SSE packed (4 floats).
• double2 - 16 bytes, SSE packed (2 doubles).

Arithmetic

Binary operators +, -, *, / work on scalar float/double via SSE scalar opcodes (ADDSS/ADDSD/etc.). Implicit promotion: int + float -> float, float + double -> double. SIMD types need matching types for +/-/*//; mixing scalar and SIMD is a compile error.

Casts

(int)3.7 -> 3 (truncating). (float)5 -> 5.0. (double)1.5f widens. Casts lower to CVTSI2SS/CVTTSS2SI/CVTSS2SD/etc.

Element access

v.x / v.y / v.z / v.w on float4 extracts a scalar float via SHUFPS. v.x / v.y on double2 via SHUFPD.

Intrinsics

_mm_*_ps (17 variants) and _mm_*_pd (11 variants) are recognized by name and inlined as SSE opcodes. See kernel/simd_intrin.h for the full list.

libm

25 operations (50 symbols with f-variants). Hardware fast-paths for sqrt, sin, cos, tan, atan, atan2, fabs, floor, ceil, round, trunc, fmod, exp2, log2. Composite for exp, log, pow, asin, acos, sinh, cosh, tanh, cbrt, hypot, nextafter.

Important: libm functions use a CupidC-internal ABI (result returned in XMM0, not ST(0) as System V would expect). Calling them from plain kernel C will NOT work with default GCC code. They are intended to be called from CupidC JIT'd code via the BIND_T table.

CupidASM opcodes

Phase B adds ~80 new opcodes: FPU state control (FXSAVE, FXRSTOR, FINIT, FNINIT, FWAIT, LDMXCSR, STMXCSR), SSE scalar (23: MOVSS/SD, ADDSS/SD, etc.), SSE packed (24: MOVAPS/UPS, ADDPS/PD, SHUFPS, CMPPS, etc.), x87 (25: FLD, FSIN, FPATAN, F2XM1, FYL2X, etc.). XMM0-7 and ST0-7 register tokens.

Testing

• bin/feature12_float.cc - scalar float arithmetic, casts, element access.
• bin/feature13_double.cc - double + transcendentals.
• bin/feature14_simd.cc - float4/double2 + intrinsics.
• bin/feature15_libm.cc - cycle 8 functions x 7 inputs vs glibc reference.
• bin/feature16_asm_fpu.cc - CupidC inline asm using SSE + x87.
• bin/fp_drill.cc - manual #XF provocation (panics kernel).
• demos/fpu_kernel.asm - CupidASM FPU state + scalar + packed + x87.
• demos/simd_blur.asm - SIMD box blur via MOVUPS/ADDPS/MULPS.

Interactive: boot QEMU graphical, type the command in the shell.

Stress test

fpu_context_stress() in kernel/fpu.c spawns 8 threads each running 100k sin() loops, compares against a serial reference. Gated by -DFPU_STRESS. To run: add -DFPU_STRESS to CFLAGS, rebuild, boot. Panics on corruption.

Exception drill

fp_drill shell command unmasks DE in MXCSR and divides by zero, verifying the #XF handler panics with MXCSR dump. Kernel reboot required afterward.