Prototype code for quickly reading the values of all CPU Performance Monitoring Counters from Lua.
Uses dynasm to generate the routine at runtime based on the capabilities of the running CPU.
dynasm: https://github.com/luapower/dynasm/blob/master/dynasm.md
| -- rdpmc_all(uint64_t[nfixed+ngeneral] *dst) | |
| -- | |
| -- Generate code to read performance monitoring counter values into dst. | |
| -- | |
| -- The generated code is specialized based on the number of fixed- and | |
| -- general-purpose performance counters that are available in | |
| -- hardware. These numbers are determined separately via the CPUID instruction. | |
| -- | |
| -- In practice on Sandy Bridge - Skylake CPUs the number of | |
| -- fixed-purpose registers is three and the number of general-purpose | |
| -- registers is either four (hyperthreads enabled) or eight | |
| -- (hyperthreads disabled). | |
| function rdpmc_multi (Dst, nfixed, ngeneral) | |
| local offset = 0 | |
| -- Read a PMC register value into the next slot of the destination buffer | |
| local function rdpmc (isfixed, index) | |
| local arg = (isfixed and 0x40000000 or 0) + index | |
| | mov ecx, arg | |
| | rdpmc | |
| | mov [edi+offset], eax | |
| | mov [edi+offset+4], edx | |
| offset = offset + 8 | |
| end | |
| for i = 0, nfixed-1 do rdpmc(true, i) end | |
| for i = 0, ngeneral-1 do rdpmc(false, i) end | |
| | ret | |
| end |
| ;; Code generated by: rdpmc_multi(Dst, 3, 4) | |
| 7fa487095000 B900000040 mov ecx, 0x40000000 | |
| 7fa487095005 0F33 rdpmc | |
| 7fa487095007 8907 mov [rdi], eax | |
| 7fa487095009 895704 mov [rdi+0x4], edx | |
| 7fa48709500c B901000040 mov ecx, 0x40000001 | |
| 7fa487095011 0F33 rdpmc | |
| 7fa487095013 894708 mov [rdi+0x8], eax | |
| 7fa487095016 89570C mov [rdi+0xc], edx | |
| 7fa487095019 B902000040 mov ecx, 0x40000002 | |
| 7fa48709501e 0F33 rdpmc | |
| 7fa487095020 894710 mov [rdi+0x10], eax | |
| 7fa487095023 895714 mov [rdi+0x14], edx | |
| 7fa487095026 B900000000 mov ecx, 0x0 | |
| 7fa48709502b 0F33 rdpmc | |
| 7fa48709502d 894718 mov [rdi+0x18], eax | |
| 7fa487095030 89571C mov [rdi+0x1c], edx | |
| 7fa487095033 B901000000 mov ecx, 0x1 | |
| 7fa487095038 0F33 rdpmc | |
| 7fa48709503a 894720 mov [rdi+0x20], eax | |
| 7fa48709503d 895724 mov [rdi+0x24], edx | |
| 7fa487095040 B902000000 mov ecx, 0x2 | |
| 7fa487095045 0F33 rdpmc | |
| 7fa487095047 894728 mov [rdi+0x28], eax | |
| 7fa48709504a 89572C mov [rdi+0x2c], edx | |
| 7fa48709504d B903000000 mov ecx, 0x3 | |
| 7fa487095052 0F33 rdpmc | |
| 7fa487095054 894730 mov [rdi+0x30], eax | |
| 7fa487095057 895734 mov [rdi+0x34], edx | |
| 7fa48709505a C3 ret |