30 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck %s3 4; Make sure that the 16-bit constant 0x3c00 isn't folded as 0 into5; v_bfi_b32.6define i32 @mov16_bfi_fold_regression(half %arg, i32 %arg1) {7; CHECK-LABEL: bfi_fold_regression:8; CHECK: ; %bb.0: ; %bb9; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)10; CHECK-NEXT: v_mov_b16_e32 v2.l, 0x3c0011; CHECK-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v112; CHECK-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)13; CHECK-NEXT: v_bfi_b32 v0, 0x7fff, v2, v014; CHECK-NEXT: v_cndmask_b16 v0.l, 0x3c00, v0.l, vcc_lo15; CHECK-NEXT: s_delay_alu instid0(VALU_DEP_1)16; CHECK-NEXT: v_pack_b32_f16 v0, v0.l, 017; CHECK-NEXT: s_setpc_b64 s[30:31]18bb:19 %cmp = icmp eq i32 %arg1, 020 %call = call half @llvm.copysign.f16(half 0xH3C00, half %arg)21 %select = select i1 %cmp, half 0xH3C00, half %call22 %insertelement = insertelement <2 x half> zeroinitializer, half %select, i64 023 %bitcast = bitcast <2 x half> %insertelement to i3224 ret i32 %bitcast25}26 27declare half @llvm.copysign.f16(half, half) #028 29attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }30