Reverse-Engineering the Intel 8087 FPU Trigonometric Functions
- Reverse-engineering efforts focusing on Intel’s historic 8087 floating-point unit (FPU) have detailed the precise hybrid arithmetic algorithms used to compute trigonometric functions like FPTAN.
- Simple microcontrollers or microprocessors operating without dedicated hardware floating-point support, such as the 6502 or Z80, typically utilize algorithms like CORDIC (Coordinate Rotation Digital Computer).
- The complete microcode listing with detailed comments reveals exactly how clock cycles were distributed during the operation of trigonometric instructions on the silicon.
Reverse-engineering efforts focusing on Intel’s historic 8087 floating-point unit (FPU) have detailed the precise hybrid arithmetic algorithms used to compute trigonometric functions like FPTAN. The analysis, conducted by Ken Shirriff and collaborators, unpacks the microcode and mathematical techniques Intel engineers relied on to achieve 64 bits of accuracy before the era of modern scaling and instruction set extensions.
Hybrid Arithmetic on the Intel 8087
Simple microcontrollers or microprocessors operating without dedicated hardware floating-point support, such as the 6502 or Z80, typically utilize algorithms like CORDIC (Coordinate Rotation Digital Computer). CORDIC relies on basic hardware operations including addition, subtraction, bitshifts, and look-up tables. By contrast, the Intel 8087 employed a hybrid approach designed to balance processing speed and mathematical accuracy. According to Shirriff’s analysis, the FPTAN implementation starts by calculating the first 16 bits using CORDIC before transitioning to the Padé approximant technique, which involves the ratio of two polynomials. Because CORDIC handles the brunt of the calculation first, the remaining value is relatively small, allowing the polynomial approximation step to execute both rapidly and accurately. This design avoided the massive look-up table sizes and extended processing time that would have been required to resolve all 64 bits through CORDIC alone.
Microcode Execution and Processing Ratios
The complete microcode listing with detailed comments reveals exactly how clock cycles were distributed during the operation of trigonometric instructions on the silicon. Calculations for a representative value show that FPTAN spent approximately 33 percent of its time on CORDIC pseudo-division, 47 percent on CORDIC pseudo-multiplication, and just 15 percent on the polynomial approximation, alongside roughly 5 percent general overhead. This structural breakdown demonstrates why the hybrid technique succeeded in maximizing performance under the hardware constraints of the early 1980s.
The Shift Away from CORDIC in Later Architectures
With the introduction of the Pentium series of CPUs, Intel moved away from CORDIC entirely. While the algorithm offers high precision, it does not scale efficiently to a significant number of bits without introducing severe time penalties. Subsequent hardware developments, including the advent of Single Instruction, Multiple Data (SIMD) extensions, further reduced the functional reliance of modern systems on the legacy x87 instruction set architecture. Despite these sweeping shifts in chip design, examining the microcode of the 8087 highlights the engineering ingenuity required to deliver mainframe-class mathematical precision on early microprocessors.