81 lines · plain
1; RUN: opt -passes='default<O1>,instnamer' -mtriple=amdgcn-- -S -o - %s | FileCheck -check-prefixes=GCN,O1 %s2; RUN: opt -passes='default<O2>,instnamer' -mtriple=amdgcn-- -S -o - %s | FileCheck -check-prefixes=GCN,O2 %s3; RUN: opt -passes='default<O3>,instnamer' -mtriple=amdgcn-- -S -o - %s | FileCheck -check-prefixes=GCN,O3 %s4 5; GCN-LABEL: t06; O1-NOT: alloca7; O2-NOT: alloca8; O3-NOT: alloca9; GCN-COUNT-27: = load10; GCN-COUNT-26: = add11define protected amdgpu_kernel void @t0(ptr addrspace(1) %p.coerce) #0 {12entry:13 %p = alloca ptr, align 8, addrspace(5)14 %p.ascast = addrspacecast ptr addrspace(5) %p to ptr15 %p.addr = alloca ptr, align 8, addrspace(5)16 %p.addr.ascast = addrspacecast ptr addrspace(5) %p.addr to ptr17 %t = alloca [27 x i32], align 16, addrspace(5)18 %t.ascast = addrspacecast ptr addrspace(5) %t to ptr19 %sum = alloca i32, align 4, addrspace(5)20 %sum.ascast = addrspacecast ptr addrspace(5) %sum to ptr21 %i = alloca i32, align 4, addrspace(5)22 %i.ascast = addrspacecast ptr addrspace(5) %i to ptr23 %cleanup.dest.slot = alloca i32, align 4, addrspace(5)24 %0 = addrspacecast ptr addrspace(1) %p.coerce to ptr25 store ptr %0, ptr %p.ascast, align 826 %p1 = load ptr, ptr %p.ascast, align 827 store ptr %p1, ptr %p.addr.ascast, align 828 call void @llvm.lifetime.start.p5(i64 48, ptr addrspace(5) %t)29 %1 = load ptr, ptr %p.addr.ascast, align 830 call void @copy(ptr %t.ascast, ptr %1, i32 27)31 call void @llvm.lifetime.start.p5(i64 4, ptr addrspace(5) %sum)32 store i32 0, ptr %sum.ascast, align 433 call void @llvm.lifetime.start.p5(i64 4, ptr addrspace(5) %i)34 store i32 0, ptr %i.ascast, align 435 br label %for.cond36 37for.cond: ; preds = %for.inc, %entry38 %2 = load i32, ptr %i.ascast, align 439 %cmp = icmp slt i32 %2, 2740 br i1 %cmp, label %for.body, label %for.cond.cleanup41 42for.cond.cleanup: ; preds = %for.cond43 call void @llvm.lifetime.end.p5(i64 4, ptr addrspace(5) %i)44 br label %for.end45 46for.body: ; preds = %for.cond47 %3 = load i32, ptr %i.ascast, align 448 %idxprom = sext i32 %3 to i6449 %arrayidx = getelementptr inbounds [27 x i32], ptr %t.ascast, i64 0, i64 %idxprom50 %4 = load i32, ptr %arrayidx, align 451 %5 = load i32, ptr %sum.ascast, align 452 %add = add nsw i32 %5, %453 store i32 %add, ptr %sum.ascast, align 454 br label %for.inc55 56for.inc: ; preds = %for.body57 %6 = load i32, ptr %i.ascast, align 458 %inc = add nsw i32 %6, 159 store i32 %inc, ptr %i.ascast, align 460 br label %for.cond61 62for.end: ; preds = %for.cond.cleanup63 %7 = load i32, ptr %sum.ascast, align 464 %8 = load ptr, ptr %p.addr.ascast, align 865 store i32 %7, ptr %8, align 466 call void @llvm.lifetime.end.p5(i64 4, ptr addrspace(5) %sum)67 call void @llvm.lifetime.end.p5(i64 48, ptr addrspace(5) %t)68 ret void69}70 71define internal void @copy(ptr %d, ptr %s, i32 %N) {72entry:73 %N8 = mul i32 %N, 474 tail call void @llvm.memcpy.p0.p0.i32(ptr %d, ptr %s, i32 %N8, i1 false)75 ret void76}77 78declare void @llvm.lifetime.start.p5(i64 immarg, ptr addrspace(5) nocapture)79declare void @llvm.lifetime.end.p5(i64 immarg, ptr addrspace(5) nocapture)80declare void @llvm.memcpy.p0.p0.i32(ptr nocapture, ptr nocapture, i32, i1)81