Scheduling new

backend/pre_optimizer/src/instruction_scheduling/instr_schedule.rs

@@ -513,7 +513,7 @@ pub fn instr_schedule_by_dag(
 				// calc my_reads
 				let my_reads = {
 					if state.instrs.last().unwrap().is_call() {
-						dag.call_reads.last().unwrap().clone()
+						dag.call_reads[state.call_ids.len() - 1].clone()
 					} else {
 						dag.nodes[*instr_idx].borrow().instr.get_riscv_read().clone()
 					}

backend/pre_optimizer/src/instruction_scheduling/instrdag.rs

@@ -25,6 +25,7 @@ pub struct InstrNode {
 	pub last_use: usize,
 	pub pred: Vec<Node>,
 	pub to_end: usize,
+	pub out_deg: usize,
 }
 impl InstrNode {
 	pub fn new(instr: &RiscvInstr, id: usize) -> Self {
@@ -36,6 +37,7 @@ impl InstrNode {
 			last_use: 0,
 			pred: Vec::new(),
 			to_end: 0,
+			out_deg: 0,
 		}
 	}
 }
@@ -225,6 +227,7 @@ impl InstrDag {
 		}
 		for (idx, instr) in processed_instrs.iter().rev().enumerate() {
 			let node = Rc::new(RefCell::new(InstrNode::new(instr, idx)));
+
 			let mut instr_node_succ = Vec::new();
 			let instructions_write = instr.get_riscv_write().clone();
 			if !instr.is_call() {
@@ -323,9 +326,11 @@ impl InstrDag {
 			nodes.push(node);
 		}
 		for node in nodes.iter() {
+			let len = node.borrow().succ.len();
 			for succ in node.borrow().succ.iter() {
 				succ.borrow_mut().in_deg += 1;
 			}
+			node.borrow_mut().out_deg += len;
 		}
 		for (index, instr) in nodes.iter_mut().enumerate().rev() {
 			instr.borrow_mut().last_use +=
@@ -347,10 +352,12 @@ impl InstrDag {
 		})
 	}
 	pub fn assign_nodes(&mut self) {
+		// 这个函数是软流水函数 实际不参与运算
 		// 先备份一遍所有 node 的 indegs
-		let indegs =
-			self.nodes.iter().map(|x| x.borrow().in_deg).collect::<Vec<_>>();
+		let out_degs =
+			self.nodes.iter().map(|x| x.borrow().out_deg).collect::<Vec<_>>();
 		// 开始遍历
+
 		let mut stack_ = Vec::new();
 		for i in self.nodes.iter() {
 			if i.borrow().succ.is_empty() {
@@ -367,15 +374,15 @@ impl InstrDag {
 					node.borrow().to_end + i.borrow().instr.get_rtn_array()[4] as usize,
 				);
 				i.borrow_mut().to_end = new_end;
-				i.borrow_mut().in_deg -= 1;
-				if i.borrow().in_deg == 0 {
+				i.borrow_mut().out_deg -= 1;
+				if i.borrow().out_deg == 0 {
 					stack_.push(i.clone());
 				}
 			}
 		}
 		// 对每个点恢复 in_deg
-		for (i, j) in self.nodes.iter().zip(indegs.iter()) {
-			i.borrow_mut().in_deg = *j;
+		for (i, j) in self.nodes.iter().zip(out_degs.iter()) {
+			i.borrow_mut().out_deg = *j;
 		}
 	}
 }

test

@@ -0,0 +1,85 @@
+  .file "./project-eval/testcases/performance/recursive_call_3.sy"
+  .option nopic
+  .attribute unaligned_access, 0
+  .attribute stack_align, 16
+  .text
+  .text
+  .global main
+  .align 1
+  .type func_calc_coef, @function
+func_calc_coef:
+    addi sp, sp, -16
+    sd s1, 0(sp)
+    sd ra, 8(sp)
+    mv s1, a0
+    li a0, 0
+    blt a1, a0, L_1
+    addiw a1, a1, -1
+    mv a0, s1
+    call func_calc_coef
+    flw fa3, 0(s1)
+    li a0, 4
+    add a1, s1, a0
+    flw fa2, 0(a1)
+    li a0, 1065353216
+    addi a0, a0, 0
+    fmv.w.x fa1, a0
+    fadd.s fa0, fa1, fa3
+    fmul.s fa1, fa3, fa0
+    fmul.s fa3, fa3, fa2
+    fadd.s fa3, fa2, fa3
+    fsub.s fa1, fa0, fa1
+    fsub.s fa2, fa2, fa3
+    fsw fa1, 0(s1)
+    fsw fa2, 0(a1)
+    j L_2
+  L_1:
+    li a1, 4
+    add a0, s1, a1
+    fmv.w.x fa0, x0
+    fsw fa0, 0(a0)
+    fsw fa0, 0(s1)
+  L_2:
+    ld s1, 0(sp)
+    ld ra, 8(sp)
+    addi sp, sp, 16
+    ret
+  .size func_calc_coef, .-func_calc_coef
+  .align 1
+  .type main, @function
+main:
+    addi sp, sp, -16
+    sd ra, 0(sp)
+    sd s1, 8(sp)
+    call getint
+    mv a1, a0
+    addi sp, sp, -16
+    mv s1, sp
+    mv a0, s1
+    call func_calc_coef
+    flw fa1, 0(s1)
+    li a1, 1065361408
+    addi a0, a1, 197
+    fmv.w.x fa2, a0
+    fmul.s fa1, fa1, fa2
+    li a1, 4
+    add a1, s1, a1
+    flw fa0, 0(a1)
+    fadd.s fa0, fa1, fa0
+    fsub.s fa1, fa0, fa2
+    fmv.w.x fa2, x0
+    feq.s s1, fa1, fa2
+    bne s1, x0, L_3
+    j L_4
+  L_3:
+    li a0, 112
+    call putch
+  L_4:
+    addi sp, sp, 16
+    mv a0, x0
+    ld ra, 0(sp)
+    ld s1, 8(sp)
+    addi sp, sp, 16
+    ret
+  .size main, .-main
+  .ident "SYSYC: (made by RRVM) 1.0.0"

utils/instruction/src/riscv/riscvinstr.rs

@@ -38,7 +38,6 @@ impl Clone for RiscvInstr {
 		self.clone_box()
 	}
 }
-
 pub trait RiscvInstrTrait:
 	Display + UseTemp<Temp> + CloneRiscvInstr + RTN
 {

utils/src/constants.rs

@@ -27,8 +27,8 @@ pub static VEC_EXTERN: [&str; 17] = [
 pub static VEC_MACRO: [&str; 2] = ["starttime", "stoptime"];
 pub const MAX_PHI_NUM: usize = 10;
 
-pub static EXTEND_TIMES: i32 = 4; // software pipelining 循环展开的次数
-pub static DEPENDENCY_EXPLORE_DEPTH: i32 = 10; //  software pipelining 过程中，对于数组的依赖，所枚举到的深度
+pub static EXTEND_TIMES: i32 = 0; // software pipelining 循环展开的次数
+pub static DEPENDENCY_EXPLORE_DEPTH: i32 = 0; //  software pipelining 过程中，对于数组的依赖，所枚举到的深度
 pub static BLOCKSIZE_THRESHOLD: usize = 100; // software pipelining 判断如果基本本块大小超了 BLOCKSIZE_THRESHOLD 后就不进行针对基本本块的优化
 pub static BFS_STATE_THRESHOLD: usize = 9; // 在 instr_scheduling 中，每轮 bfs 所保留的状态的阈值
 																					 // for instruction scheduling: register punishment