49 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -S -passes='require<profile-summary>,function(codegenprepare)' -mtriple=amdgcn--amdhsa < %s | FileCheck %s3 4define amdgpu_kernel void @test_sink_as999_small_max_mubuf_offset(ptr addrspace(999) %out, ptr addrspace(999) %in) {5; CHECK-LABEL: @test_sink_as999_small_max_mubuf_offset(6; CHECK-NEXT: entry:7; CHECK-NEXT: [[OUT_GEP:%.*]] = getelementptr i32, ptr addrspace(999) [[OUT:%.*]], i32 10248; CHECK-NEXT: [[IN_GEP:%.*]] = getelementptr i8, ptr addrspace(999) [[IN:%.*]], i64 40959; CHECK-NEXT: [[TID:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #110; CHECK-NEXT: [[TMP0:%.*]] = icmp eq i32 [[TID]], 011; CHECK-NEXT: br i1 [[TMP0]], label [[ENDIF:%.*]], label [[IF:%.*]]12; CHECK: if:13; CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr addrspace(999) [[IN_GEP]], align 114; CHECK-NEXT: [[TMP2:%.*]] = sext i8 [[TMP1]] to i3215; CHECK-NEXT: br label [[ENDIF]]16; CHECK: endif:17; CHECK-NEXT: [[X:%.*]] = phi i32 [ [[TMP2]], [[IF]] ], [ 0, [[ENTRY:%.*]] ]18; CHECK-NEXT: store i32 [[X]], ptr addrspace(999) [[OUT_GEP]], align 419; CHECK-NEXT: br label [[DONE:%.*]]20; CHECK: done:21; CHECK-NEXT: ret void22;23entry:24 %out.gep = getelementptr i32, ptr addrspace(999) %out, i32 102425 %in.gep = getelementptr i8, ptr addrspace(999) %in, i64 409526 %tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #027 %tmp0 = icmp eq i32 %tid, 028 br i1 %tmp0, label %endif, label %if29 30if:31 %tmp1 = load i8, ptr addrspace(999) %in.gep32 %tmp2 = sext i8 %tmp1 to i3233 br label %endif34 35endif:36 %x = phi i32 [ %tmp2, %if ], [ 0, %entry ]37 store i32 %x, ptr addrspace(999) %out.gep38 br label %done39 40done:41 ret void42}43 44declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #045 46attributes #0 = { nounwind readnone }47attributes #1 = { nounwind }48attributes #2 = { nounwind argmemonly }49