Skip to content

Commit ed4f439

Browse files
wangpc-ppwangpc-pp
committed
Fix comments, rework test and add test for PHI
1 parent 415c30e commit ed4f439

File tree

2 files changed

+72
-41
lines changed

2 files changed

+72
-41
lines changed

‎llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4227,7 +4227,7 @@ class RISCVPipelinerLoopInfo : public TargetInstrInfo::PipelinerLoopInfo {
42274227

42284228
bool shouldIgnoreForPipelining(const MachineInstr *MI) const override {
42294229
// Make the instructions for loop control be placed in stage 0.
4230-
// The predecessors of PredBranch are considered by the caller.
4230+
// The predecessors of LHS/RHS are considered by the caller.
42314231
if (LHS && MI == LHS)
42324232
return true;
42334233
if (RHS && MI == RHS)

‎llvm/test/CodeGen/RISCV/machine-pipeliner.ll

Lines changed: 71 additions & 40 deletions
Original file line numberDiff line numberDiff line change
@@ -1,78 +1,109 @@
11
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
22
; RUN: llc -mtriple=riscv64 -mcpu=sifive-p670 -O3 -verify-machineinstrs -riscv-enable-pipeliner=false < %s \
3-
; RUN: | FileCheck %s --check-prefixes=CHECK-NOT-PIPELINED
3+
; RUN: | FileCheck %s --check-prefixes=CHECK,CHECK-NOT-PIPELINED
44
; RUN: llc -mtriple=riscv64 -mcpu=sifive-p670 -O3 -verify-machineinstrs -riscv-enable-pipeliner=true < %s \
5-
; RUN: | FileCheck %s --check-prefixes=CHECK-PIPELINED
5+
; RUN: | FileCheck %s --check-prefixes=CHECK,CHECK-PIPELINED
66

7-
define void @test_1(ptr noalias %in, ptr noalias %out, i32 signext %cnt) "no-builtins" {
8-
; CHECK-NOT-PIPELINED-LABEL: test_1:
7+
; We shouldn't pipeline this loop as one operand of branch is a PHI.
8+
define i32 @test_phi() {
9+
; CHECK-LABEL: test_phi:
10+
; CHECK: # %bb.0: # %entry
11+
; CHECK-NEXT: li a0, 0
12+
; CHECK-NEXT: .LBB0_1: # %for.body
13+
; CHECK-NEXT: # =>This Inner Loop Header: Depth=1
14+
; CHECK-NEXT: mv a1, a0
15+
; CHECK-NEXT: li a0, 1
16+
; CHECK-NEXT: sh a0, 0(zero)
17+
; CHECK-NEXT: bnez a1, .LBB0_1
18+
; CHECK-NEXT: # %bb.2: # %for.cond.cleanup
19+
; CHECK-NEXT: li a0, 0
20+
; CHECK-NEXT: ret
21+
entry:
22+
br label %for.body
23+
24+
for.cond.cleanup: ; preds = %for.body
25+
ret i32 0
26+
27+
for.body: ; preds = %for.body, %entry
28+
%indvars.iv1 = phi i64 [ 0, %entry ], [ 1, %for.body ]
29+
store i16 1, ptr null, align 4
30+
%exitcond.not.31 = icmp eq i64 %indvars.iv1, 0
31+
br i1 %exitcond.not.31, label %for.cond.cleanup, label %for.body
32+
}
33+
34+
define void @test_pipelined_1(ptr noalias %in, ptr noalias %out, i32 signext %cnt) {
35+
; CHECK-NOT-PIPELINED-LABEL: test_pipelined_1:
936
; CHECK-NOT-PIPELINED: # %bb.0: # %entry
10-
; CHECK-NOT-PIPELINED-NEXT: blez a2, .LBB0_3
37+
; CHECK-NOT-PIPELINED-NEXT: blez a2, .LBB1_3
1138
; CHECK-NOT-PIPELINED-NEXT: # %bb.1: # %for.body.preheader
1239
; CHECK-NOT-PIPELINED-NEXT: addi a2, a2, -1
1340
; CHECK-NOT-PIPELINED-NEXT: sh2add.uw a2, a2, a1
1441
; CHECK-NOT-PIPELINED-NEXT: addi a2, a2, 4
15-
; CHECK-NOT-PIPELINED-NEXT: .LBB0_2: # %for.body
42+
; CHECK-NOT-PIPELINED-NEXT: .LBB1_2: # %for.body
1643
; CHECK-NOT-PIPELINED-NEXT: # =>This Inner Loop Header: Depth=1
1744
; CHECK-NOT-PIPELINED-NEXT: lw a3, 0(a1)
1845
; CHECK-NOT-PIPELINED-NEXT: addi a1, a1, 4
46+
; CHECK-NOT-PIPELINED-NEXT: addi a3, a3, 1
1947
; CHECK-NOT-PIPELINED-NEXT: sw a3, 0(a0)
2048
; CHECK-NOT-PIPELINED-NEXT: addi a0, a0, 4
21-
; CHECK-NOT-PIPELINED-NEXT: bne a1, a2, .LBB0_2
22-
; CHECK-NOT-PIPELINED-NEXT: .LBB0_3: # %for.end
49+
; CHECK-NOT-PIPELINED-NEXT: bne a1, a2, .LBB1_2
50+
; CHECK-NOT-PIPELINED-NEXT: .LBB1_3: # %for.end
2351
; CHECK-NOT-PIPELINED-NEXT: ret
2452
;
25-
; CHECK-PIPELINED-LABEL: test_1:
53+
; CHECK-PIPELINED-LABEL: test_pipelined_1:
2654
; CHECK-PIPELINED: # %bb.0: # %entry
27-
; CHECK-PIPELINED-NEXT: blez a2, .LBB0_6
55+
; CHECK-PIPELINED-NEXT: blez a2, .LBB1_6
2856
; CHECK-PIPELINED-NEXT: # %bb.1: # %for.body.preheader
29-
; CHECK-PIPELINED-NEXT: lw a3, 0(a1)
57+
; CHECK-PIPELINED-NEXT: lw a4, 0(a1)
3058
; CHECK-PIPELINED-NEXT: addi a2, a2, -1
31-
; CHECK-PIPELINED-NEXT: addi a4, a0, 4
3259
; CHECK-PIPELINED-NEXT: sh2add.uw a6, a2, a1
60+
; CHECK-PIPELINED-NEXT: addi a2, a0, 4
3361
; CHECK-PIPELINED-NEXT: addi a1, a1, 4
3462
; CHECK-PIPELINED-NEXT: addi a6, a6, 4
35-
; CHECK-PIPELINED-NEXT: beq a1, a6, .LBB0_5
63+
; CHECK-PIPELINED-NEXT: beq a1, a6, .LBB1_5
3664
; CHECK-PIPELINED-NEXT: # %bb.2: # %for.body
3765
; CHECK-PIPELINED-NEXT: lw a5, 0(a1)
38-
; CHECK-PIPELINED-NEXT: addi a2, a4, 4
66+
; CHECK-PIPELINED-NEXT: addi a3, a2, 4
67+
; CHECK-PIPELINED-NEXT: addi a4, a4, 1
3968
; CHECK-PIPELINED-NEXT: addi a1, a1, 4
40-
; CHECK-PIPELINED-NEXT: beq a1, a6, .LBB0_4
41-
; CHECK-PIPELINED-NEXT: .LBB0_3: # %for.body
69+
; CHECK-PIPELINED-NEXT: beq a1, a6, .LBB1_4
70+
; CHECK-PIPELINED-NEXT: .LBB1_3: # %for.body
4271
; CHECK-PIPELINED-NEXT: # =>This Inner Loop Header: Depth=1
43-
; CHECK-PIPELINED-NEXT: sw a3, 0(a0)
44-
; CHECK-PIPELINED-NEXT: mv a3, a5
72+
; CHECK-PIPELINED-NEXT: sw a4, 0(a0)
73+
; CHECK-PIPELINED-NEXT: mv a4, a5
4574
; CHECK-PIPELINED-NEXT: lw a5, 0(a1)
46-
; CHECK-PIPELINED-NEXT: mv a0, a4
47-
; CHECK-PIPELINED-NEXT: mv a4, a2
48-
; CHECK-PIPELINED-NEXT: addi a2, a2, 4
75+
; CHECK-PIPELINED-NEXT: mv a0, a2
76+
; CHECK-PIPELINED-NEXT: mv a2, a3
77+
; CHECK-PIPELINED-NEXT: addi a3, a3, 4
78+
; CHECK-PIPELINED-NEXT: addi a4, a4, 1
4979
; CHECK-PIPELINED-NEXT: addi a1, a1, 4
50-
; CHECK-PIPELINED-NEXT: bne a1, a6, .LBB0_3
51-
; CHECK-PIPELINED-NEXT: .LBB0_4:
52-
; CHECK-PIPELINED-NEXT: sw a3, 0(a0)
53-
; CHECK-PIPELINED-NEXT: mv a0, a4
54-
; CHECK-PIPELINED-NEXT: mv a3, a5
55-
; CHECK-PIPELINED-NEXT: .LBB0_5:
56-
; CHECK-PIPELINED-NEXT: sw a3, 0(a0)
57-
; CHECK-PIPELINED-NEXT: .LBB0_6: # %for.end
80+
; CHECK-PIPELINED-NEXT: bne a1, a6, .LBB1_3
81+
; CHECK-PIPELINED-NEXT: .LBB1_4:
82+
; CHECK-PIPELINED-NEXT: sw a4, 0(a0)
83+
; CHECK-PIPELINED-NEXT: mv a0, a2
84+
; CHECK-PIPELINED-NEXT: mv a4, a5
85+
; CHECK-PIPELINED-NEXT: .LBB1_5:
86+
; CHECK-PIPELINED-NEXT: addi a4, a4, 1
87+
; CHECK-PIPELINED-NEXT: sw a4, 0(a0)
88+
; CHECK-PIPELINED-NEXT: .LBB1_6: # %for.end
5889
; CHECK-PIPELINED-NEXT: ret
5990
entry:
60-
%cmp5 = icmp sgt i32 %cnt, 0
61-
br i1 %cmp5, label %for.body, label %for.end
91+
%cmp = icmp sgt i32 %cnt, 0
92+
br i1 %cmp, label %for.body, label %for.end
6293

6394
for.body: ; preds = %entry, %for.body
64-
%i.08 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
65-
%in.addr.07 = phi ptr [ %incdec.ptr, %for.body ], [ %in, %entry ]
66-
%out.addr.06 = phi ptr [ %incdec.ptr1, %for.body ], [ %out, %entry ]
67-
%0 = load i32, ptr %out.addr.06, align 4
68-
store i32 %0, ptr %in.addr.07, align 4
69-
%incdec.ptr = getelementptr inbounds i8, ptr %in.addr.07, i64 4
70-
%incdec.ptr1 = getelementptr inbounds i8, ptr %out.addr.06, i64 4
71-
%inc = add nuw nsw i32 %i.08, 1
95+
%inc.next = phi i32 [ %inc, %for.body ], [ 0, %entry ]
96+
%in.addr.next = phi ptr [ %incdec.in, %for.body ], [ %in, %entry ]
97+
%out.addr.next = phi ptr [ %incdec.out, %for.body ], [ %out, %entry ]
98+
%0 = load i32, ptr %out.addr.next, align 4
99+
%1 = add i32 %0, 1
100+
store i32 %1, ptr %in.addr.next, align 4
101+
%incdec.in = getelementptr inbounds i8, ptr %in.addr.next, i64 4
102+
%incdec.out = getelementptr inbounds i8, ptr %out.addr.next, i64 4
103+
%inc = add nuw nsw i32 %inc.next, 1
72104
%exitcond.not = icmp eq i32 %inc, %cnt
73105
br i1 %exitcond.not, label %for.end, label %for.body
74106

75107
for.end: ; preds = %for.body, %entry
76108
ret void
77109
}
78-

0 commit comments

Comments
 (0)