Split View: [컴파일러] 16. 명령어 수준 병렬성과 스케줄링

// High ILP code
a = x + y     // These three instructions are independent
b = p + q     // Can be executed simultaneously
c = r - s

// Low ILP code
a = x + y     // Sequential dependency
b = a + z     // Depends on a
c = b * 2     // Depends on b

a = b + c     // Write a
d = a + e     // Read a (must wait)

a = b + c     // Read b
b = d + e     // Write b (can be resolved by renaming)

a = b + c     // Write a
a = d + e     // Write a again (order must be maintained)

// 5-stage pipeline: IF(Fetch), ID(Decode), EX(Execute), MEM, WB(Writeback)

Time:    1    2    3    4    5    6    7
Inst 1:  IF   ID   EX   MEM  WB
Inst 2:       IF   ID   EX   MEM  WB
Inst 3:            IF   ID   EX   MEM  WB

// Ideally, one instruction completes every clock cycle

// Data hazard (RAW)
ADD R1, R2, R3    // Write result to R1 (WB stage)
SUB R4, R1, R5    // Read R1 (ID stage) - value not ready yet!

// Solution: forwarding or stall
ADD R1, R2, R3
NOP               // Bubble insertion (stall)
SUB R4, R1, R5

// Compiler inserts independent instructions to prevent stalls
ADD R1, R2, R3
MUL R6, R7, R8    // Fill empty slot with independent instruction
SUB R4, R1, R5

// 2-issue superscalar processor
Time:    1         2         3
Slot 1:  ADD       MUL       LD
Slot 2:  SUB       AND       ST

// Ideally, 2 instructions complete every clock cycle
// (IPC = Instructions Per Cycle = 2)

// VLIW instruction (4 operation slots)
// | ALU1     | ALU2     | MEM      | BRANCH   |
// | ADD R1.. | MUL R3.. | LD R5..  | NOP      |

// Each slot executes on an independent functional unit
// The compiler is responsible for parallel execution

// Original code:
// 1: LD  R1, a       // Load latency = 2
// 2: ADD R2, R1, R3  // Depends on R1 (RAW)
// 3: LD  R4, b       // Independent
// 4: MUL R5, R4, R6  // Depends on R4 (RAW)
// 5: ADD R7, R2, R5  // Depends on R2, R5

// Dependency DAG:
//   1 --2--> 2 --1--> 5
//   3 --2--> 4 --1--> 5
// (Edge weights = latencies)

Algorithm: List Scheduling

1. Build dependency DAG
2. Compute priority for each node (usually critical path length)
3. Add instructions with no dependencies to the ready list
4. while ready list is not empty:
     a. Select the instruction with highest priority
     b. Issue at current cycle
     c. Add successor instructions to ready list when dependencies are resolved

// Dependency DAG (latencies):
//   LD a(2) -> ADD(1) -> ST(1)
//   LD b(2) -> MUL(3) -> ST(1)

// Priorities (critical path length to root):
//   LD a: 2+1+1 = 4
//   LD b: 2+3+1 = 6  (higher priority)
//   ADD: 1+1 = 2
//   MUL: 3+1 = 4
//   ST(ADD result): 1
//   ST(MUL result): 1

// Scheduling:
// Cycle 1: LD b (priority 6, highest)
// Cycle 2: LD a (waiting for LD b result, issue something else)
// Cycle 3: MUL (LD b result ready, cycle 1+2=3)
// Cycle 4: ADD (LD a result ready, cycle 2+2=4)
// Cycle 5: ST (ADD result, cycle 4+1=5)
// Cycle 6: ST (MUL result, cycle 3+3=6)

// Original order would need 7 cycles
// After scheduling, reduced to 6 cycles

// Original loop (3 instructions per iteration, 1 cycle each)
// Iteration i:  A(i), B(i), C(i)
// Iteration i+1: A(i+1), B(i+1), C(i+1)

// Time:  1     2     3     4     5     6     7
// Orig:  A(0)  B(0)  C(0)  A(1)  B(1)  C(1)  A(2)...

// After software pipelining:
// Time:  1     2     3     4     5
// Orig:  A(0)  B(0)  C(0)
//              A(1)  B(1)  C(1)
//                    A(2)  B(2)  C(2)
//
// Steady state kernel:
// Time:  ...   k     k+1   k+2   ...
//              C(i)  C(i+1) ...
//              B(i+1) B(i+2) ...
//              A(i+2) A(i+3) ...

// 1. Prologue: Filling the pipeline
A(0)
A(1), B(0)

// 2. Kernel: Steady state iteration
loop:
    A(i+2), B(i+1), C(i)   // Three iterations overlap
    i = i + 1
    if i < n-2 goto loop

// 3. Epilogue: Draining the pipeline
B(n-1), C(n-2)
C(n-1)

// Code favorable for scheduling (uses many registers)
LD R1, a
LD R2, b        // Perform another load while waiting for R1
LD R3, c
ADD R4, R1, R2  // R1 is ready
MUL R5, R3, R4

// Code favorable for register allocation (uses few registers)
LD R1, a
ADD R1, R1, b   // Use R1 immediately (may stall!)
LD R1, c
MUL R1, R1, R1

1. Perform scheduling with virtual registers (infinite registers)
2. Perform register allocation on scheduled code
3. If spills occur, insert spill code and reschedule

1. Perform register allocation first
2. Schedule the allocated code
3. Scheduling freedom is limited (false dependencies from shared registers)

1. Consider register allocation and scheduling simultaneously
2. Can achieve better results but complexity is high

// Flow graph:
//   B1 (execution probability 100%)
//    |
//   B2 (if cond) -- 90% --> B3
//        |
//        10% --> B4

// Trace: B1 -> B2 -> B3 (most frequent path)
// Schedule this path as one long block

// Instructions can be moved from B2 to B3
// or from B1 to B2
// (compensation code needed for the less frequent path B4)

// Superblock: one entry point, multiple exit points
//   entry -> A -> B -> C -> exit1
//                 |
//                 v
//               exit2
//
// No incoming edges in the middle, so code motion is freer

// VLIW instruction packet (3 slots: ALU, MEM, BR)
// Place independent operations in each slot

// Cycle 1: | ADD R1,R2,R3 | LD R4, mem1 | NOP        |
// Cycle 2: | MUL R5,R4,R6 | LD R7, mem2 | NOP        |
// Cycle 3: | SUB R8,R5,R7 | ST mem3, R1 | BEQ R8, L1 |

// Code with low ILP:
// | ADD R1,R2,R3 | NOP         | NOP |
// | MUL R4,R1,R5 | NOP         | NOP |  // All operations are dependent
// | SUB R6,R4,R7 | NOP         | NOP |

// Solution: Increase ILP through loop unrolling + software pipelining