$Id$ Partial x86 code optimisation guide =================================== Priority should be given to P6 and P4, then K7, then P5, and last to K6. Rules that are blatantly obvious or irrelevant for HiPE are generally not listed. These includes things like alignment of basic data types, store-forwarding rules when alignment or sizes don't match, and partial register stalls. Intel P4 -------- The P6 4-1-1 insn decode template no longer applies. Simple insns (add/sub/cmp/test/and/or/xor/neg/not/mov/sahf) are twice as fast as in P6. Shifts are "movsx" (sign-extend) are slower than in P6. Always avoid "inc" and "dec", use "add" and "sub" instead, due to condition codes dependencies overhead. "fxch" is slightly more expensive than in P6, where it was free. Use "setcc" or "cmov" to eliminate unpredictable branches. For hot code executing out of the trace cache, alignment of branch targets is less of an issue compared to P6. Do use "fxch" to simulate a flat FP register file, but only for that purpose, not for manual scheduling for parallelism. Using "lea" is highly recommended. Eliminate redundant loads. Use regs as much as possible. Left shifts up to 3 have longer latencies than the equivalent sequence of adds. Do utilise the addressing modes, to save registers and trace cache bandwidth. "xor reg,reg" or "sub reg,reg" preferred over moving zero to reg. "test reg,reg" preferred over "cmp" with zero or "and". Avoid explicit cmp/test;jcc if the preceeding insn (alu, but not mov or lea) set the condition codes. Load-execute alu insns (mem src) are Ok. Add-reg-to-mem slightly better than add-mem-to-reg. Add-reg-to-mem is better than load;add;store. Intel P6 -------- 4-1-1 instruction decoding template: can decode one semi-complex (max 4 uops) and two simple (1 uop) insns per clock; follow a complex insn by two simple ones, otherwise the decoders will stall. Load-execute (mem src) alu insns are 2 uops. Read-modify-write (mem dst) alu insns are 4 uops. Insns longer than 7 bytes block parallel decoding. Avoid insns longer than 7 bytes. Lea is useful. "movzx" is preferred for zero-extension; the xor;mov alternative causes a partial register stall. Use "test" instead of "cmp" with zero. Pull address calculations into load and store insn addressing modes. Clear a reg with "xor", not by moving zero to it. Many alu insns set the condition codes. Replace "alu;cmp;jcc" with "alu;jcc". This is not applicable for "mov" or "lea". For FP code, simulate a flat register file on the x87 stack by using fxch to reorder it. AMD K7 ------ Select DirectPath insns. Avoid VectorPath insns due to slower decode. Alu insns with mem src are very efficient. Alu insns with mem dst are very efficient. Fetches from I-cache are 16-byte aligned. Align functions and frequently used labels at or near the start of 16-byte aligned blocks. "movzx" preferred over "xor;mov" for zero-extension. "push mem" preferred over "load;push reg". "xor reg,reg" preferred over moving zero to the reg. "test" preferred over "cmp". "pop" insns are VectorPath. "pop mem" has latency 3, "pop reg" has latency 4. "push reg" and "push imm" are DirectPath, "push mem" is VectorPath. The latency is 3 clocks. Intel P5 -------- If a loop header is less than 8 bytes away from a 16-byte boundary, align it to the 16-byte boundary. If a return address is less than 8 bytes away from a 16-byte boundary, align it to the 16-byte boundary. Align function entry points to 16-byte boundaries. Ensure that doubles are 64-bit aligned. Data cache line size is 32 bytes. The whole line is brought in on a read miss. "push mem" is not pairable; loading a temp reg and pushing the reg pairs better -- this is also faster on the 486. No conditional move instruction. Insns longer than 7 bytes can't go down the V-pipe or share the insn FIFO with other insns. Avoid insns longer than 7 bytes. Lea is useful when it replaces several other add/shift insns. Lea is not a good replacement for a single shl since a scaled index requires a disp32 (or base), making the insn longer. "movzx" is worse than the xor;mov alternative -- the opcode prefix causes a slowdown and it is not pariable. Use "test" instead of "cmp" with zero. "test eax,imm" and "test reg,reg" are pairable, other forms are not. Pull address calculations into load and store insn addressing modes. Clear a reg with "xor", not by moving zero to it. Many alu insns set the condition codes. Replace "alu;cmp;jcc" with "alu;jcc". This is not applicable for "mov" or "lea". For FP code, simulate a flat register file on the x87 stack by using fxch to reorder it. "neg" and "not" are not pairable. "test imm,reg" and "test imm,mem" are not pairable. Shifts by "cl" are not pairable. Shifts by "1" or "imm" are pairable but only execute in the U-pipe. AMD K6 ------ The insn size predecoder has a 3-byte window. Insns with both prefix and SIB bytes cannot be short-decoded. Use short and simple insns, including mem src alu insns. Avoid insns longer than 7 bytes. They cannot be short-decoded. Short-decode: max 7 bytes, max 2 uops. Long-decode: max 11 bytes, max 4 uops. Vector-decode: longer than 11 bytes or more than 4 uops. Prefer read-modify-write alu insns (mem dst) over "load;op;store" sequences, for code density and register pressure reasons. Avoid the "(esi)" addressing mode: it forces the insn to be vector-decoded. Use a different reg or add an explicit zero displacement. "add reg,reg" preferred over a shl by 1, it parallelises better. "movzx" preferred over "xor;mov" for zero-extension. Moving zero to a reg preferred over "xor reg,reg" due to dependencies and condition codes overhead. "push mem" preferred over "load;push reg" due to code density and register pressure. (Page 64.) Explicit moves preferred when pushing args for fn calls, due to %esp dependencies and random access possibility. (Page 58.) [hmm, these two are in conflict] There is no penalty for seg reg prefix unless there are multiple prefixes. Align function entries and frequent branch targets to 16-byte boundaries. Shifts by imm only go down one of the pipes. "test reg,reg" preferred over "cmp" with zero. "test reg,imm" is a long-decode insn. No conditional move insn.