53 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -O3 -mtriple=nvptx64 -enable-misched %s -o - | FileCheck %s3 4target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"5target triple = "nvptx64-nvidia-cuda"6 7define ptx_kernel void @wombat(i32 %arg, i32 %arg1, i32 %arg2) {8; CHECK-LABEL: wombat(9; CHECK: {10; CHECK-NEXT: .reg .b32 %r<8>;11; CHECK-NEXT: .reg .b64 %rd<5>;12; CHECK-EMPTY:13; CHECK-NEXT: // %bb.0: // %bb14; CHECK-NEXT: ld.param.b32 %r3, [wombat_param_2];15; CHECK-NEXT: ld.param.b32 %r2, [wombat_param_1];16; CHECK-NEXT: ld.param.b32 %r1, [wombat_param_0];17; CHECK-NEXT: mov.b32 %r7, 0;18; CHECK-NEXT: $L__BB0_1: // %bb319; CHECK-NEXT: // =>This Inner Loop Header: Depth=120; CHECK-NEXT: { // callseq 0, 021; CHECK-NEXT: .param .b64 param0;22; CHECK-NEXT: .param .b64 retval0;23; CHECK-NEXT: st.param.b64 [param0], 0;24; CHECK-NEXT: call.uni (retval0), quux, (param0);25; CHECK-NEXT: } // callseq 026; CHECK-NEXT: mul.lo.s32 %r4, %r7, %r2;27; CHECK-NEXT: or.b32 %r5, %r3, %r4;28; CHECK-NEXT: mul.lo.s32 %r6, %r1, %r5;29; CHECK-NEXT: cvt.rn.f64.s32 %rd1, %r6;30; CHECK-NEXT: cvt.rn.f64.u32 %rd2, %r7;31; CHECK-NEXT: add.rn.f64 %rd3, %rd2, %rd1;32; CHECK-NEXT: mov.b64 %rd4, 0;33; CHECK-NEXT: st.global.b64 [%rd4], %rd3;34; CHECK-NEXT: mov.b32 %r7, 1;35; CHECK-NEXT: bra.uni $L__BB0_1;36bb:37 br label %bb338 39bb3: ; preds = %bb3, %bb40 %phi = phi i32 [ 0, %bb ], [ 1, %bb3 ]41 %call = tail call double @quux(double 0.000000e+00)42 %mul = mul i32 %phi, %arg143 %or = or i32 %arg2, %mul44 %mul4 = mul i32 %arg, %or45 %sitofp = sitofp i32 %mul4 to double46 %uitofp = uitofp i32 %phi to double47 %fadd = fadd double %uitofp, %sitofp48 store double %fadd, ptr addrspace(1) null, align 849 br label %bb350}51 52declare double @quux(double)53