| ; RUN: llc -mtriple=mips64el-unknown-linux-gnu -mcpu=mips64r2 -target-abi=n32 < %s -o - \ | |
| ; RUN: | FileCheck -check-prefixes=N32 %s | |
| ; RUN: llc -mtriple=mips64el-unknown-linux-gnu -mcpu=mips64r2 < %s -o - \ | |
| ; RUN: | FileCheck -check-prefixes=N64 %s | |
| ; Test that when a user declares $28 as a global register, assignments to it are | |
| ; not optimized away. | |
| ; Clang previously generated no instruction, which GCC generated `move $gp, $4`. | |
| ; This test ensures Clang matches GCC behavior. | |
| define void @setglobal(i64 %x) nounwind { | |
| ; N32-LABEL: setglobal: | |
| ; N32: # %bb.0: # %entry | |
| ; N32-NEXT: jr $ra | |
| ; N32-NEXT: move $gp, $4 | |
| ; N64-LABEL: setglobal: | |
| ; N64: # %bb.0: # %entry | |
| ; N64-NEXT: jr $ra | |
| ; N64-NEXT: move $gp, $4 | |
| entry: | |
| tail call void @llvm.write_register.i64(metadata !0, i64 %x) | |
| ret void | |
| } | |
| declare void @llvm.write_register.i64(metadata, i64) nounwind | |
| !llvm.named.register.$28 = !{!0} | |
| !0 = !{!"$28"} | |