Debian Codegraph @

llvm/test/Transforms/LoopUnroll/AMDGPU/unroll-runtime.ll

raw
1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt -mtriple=amdgcn-amd-amdhsa -passes=loop-unroll -S %s -o - | FileCheck %s
3
4 %struct.wombat = type { %struct.zot, i32, [16 x i32], [16 x i32], i32, i32, [16 x i32], i32 }
5 %struct.zot = type { i32, i32, [1024 x i32] }
6
7 @global = external addrspace(3) global %struct.wombat
8
9 ; Ensure that a cascaded GEP for local address space does not inhibit unrolling
10 ;
11 define amdgpu_kernel void @unroll_when_cascaded_gep(i32 %arg) {
12 ; CHECK-LABEL: @unroll_when_cascaded_gep(
13 ; CHECK-NEXT: bb:
14 ; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[ARG:%.*]], 1
15 ; CHECK-NEXT: [[XTRAITER:%.*]] = and i32 [[TMP0]], 7
16 ; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i32 [[ARG]], 7
17 ; CHECK-NEXT: br i1 [[TMP1]], label [[BB1_EPIL_PREHEADER:%.*]], label [[BB_NEW:%.*]]
18 ; CHECK: bb.new:
19 ; CHECK-NEXT: [[UNROLL_ITER:%.*]] = sub i32 [[TMP0]], [[XTRAITER]]
20 ; CHECK-NEXT: br label [[BB1:%.*]]
21 ; CHECK: bb1:
22 ; CHECK-NEXT: [[PHI:%.*]] = phi i32 [ 0, [[BB_NEW]] ], [ [[ADD_7:%.*]], [[BB1]] ]
23 ; CHECK-NEXT: [[NITER:%.*]] = phi i32 [ 0, [[BB_NEW]] ], [ [[NITER_NEXT_7:%.*]], [[BB1]] ]
24 ; CHECK-NEXT: [[ADD_7]] = add i32 [[PHI]], 8
25 ; CHECK-NEXT: [[NITER_NEXT_7]] = add i32 [[NITER]], 8
26 ; CHECK-NEXT: [[NITER_NCMP_7:%.*]] = icmp eq i32 [[NITER_NEXT_7]], [[UNROLL_ITER]]
27 ; CHECK-NEXT: br i1 [[NITER_NCMP_7]], label [[BB2_UNR_LCSSA:%.*]], label [[BB1]]
28 ; CHECK: bb2.unr-lcssa:
29 ; CHECK-NEXT: [[PHI_UNR:%.*]] = phi i32 [ [[ADD_7]], [[BB1]] ]
30 ; CHECK-NEXT: [[LCMP_MOD:%.*]] = icmp ne i32 [[XTRAITER]], 0
31 ; CHECK-NEXT: br i1 [[LCMP_MOD]], label [[BB1_EPIL_PREHEADER]], label [[BB2:%.*]]
32 ; CHECK: bb1.epil.preheader:
33 ; CHECK-NEXT: [[PHI_EPIL_INIT:%.*]] = phi i32 [ 0, [[BB:%.*]] ], [ [[PHI_UNR]], [[BB2_UNR_LCSSA]] ]
34 ; CHECK-NEXT: [[LCMP_MOD1:%.*]] = icmp ne i32 [[XTRAITER]], 0
35 ; CHECK-NEXT: call void @llvm.assume(i1 [[LCMP_MOD1]])
36 ; CHECK-NEXT: br label [[BB1_EPIL:%.*]]
37 ; CHECK: bb1.epil:
38 ; CHECK-NEXT: [[PHI_EPIL:%.*]] = phi i32 [ [[PHI_EPIL_INIT]], [[BB1_EPIL_PREHEADER]] ], [ [[ADD_EPIL:%.*]], [[BB1_EPIL]] ]
39 ; CHECK-NEXT: [[EPIL_ITER:%.*]] = phi i32 [ 0, [[BB1_EPIL_PREHEADER]] ], [ [[EPIL_ITER_NEXT:%.*]], [[BB1_EPIL]] ]
40 ; CHECK-NEXT: [[GETELEMENTPTR_EPIL:%.*]] = getelementptr [1024 x i32], ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) @global, i32 8), i32 0, i32 0
41 ; CHECK-NEXT: [[ADD_EPIL]] = add i32 [[PHI_EPIL]], 1
42 ; CHECK-NEXT: [[ICMP_EPIL:%.*]] = icmp eq i32 [[PHI_EPIL]], [[ARG]]
43 ; CHECK-NEXT: [[EPIL_ITER_NEXT]] = add i32 [[EPIL_ITER]], 1
44 ; CHECK-NEXT: [[EPIL_ITER_CMP:%.*]] = icmp ne i32 [[EPIL_ITER_NEXT]], [[XTRAITER]]
45 ; CHECK-NEXT: br i1 [[EPIL_ITER_CMP]], label [[BB1_EPIL]], label [[BB2_EPILOG_LCSSA:%.*]], !llvm.loop [[LOOP0:![0-9]+]]
46 ; CHECK: bb2.epilog-lcssa:
47 ; CHECK-NEXT: br label [[BB2]]
48 ; CHECK: bb2:
49 ; CHECK-NEXT: ret void
50 ;
51 bb:
52 br label %bb1
53
54 bb1: ; preds = %bb1, %bb
55 %phi = phi i32 [ 0, %bb ], [ %add, %bb1 ]
56 %getelementptr = getelementptr [1024 x i32], ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) @global, i32 8), i32 0, i32 0
57 %add = add i32 %phi, 1
58 %icmp = icmp eq i32 %phi, %arg
59 br i1 %icmp, label %bb2, label %bb1
60
61 bb2: ; preds = %bb1
62 ret void
63 }
64