525 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -verify-machineinstrs --show-mc-encoding -mtriple=x86_64-unknown-unknown -mattr=+avx512fp16,avx512vl | FileCheck %s --check-prefixes=CHECK,HasVL3; RUN: llc < %s -verify-machineinstrs --show-mc-encoding -mtriple=x86_64-unknown-unknown -mattr=+avx512fp16 | FileCheck %s --check-prefixes=CHECK,NOVL4 5declare half @llvm.maxnum.f16(half, half)6declare <2 x half> @llvm.maxnum.v2f16(<2 x half>, <2 x half>)7declare <4 x half> @llvm.maxnum.v4f16(<4 x half>, <4 x half>)8declare <8 x half> @llvm.maxnum.v8f16(<8 x half>, <8 x half>)9declare <16 x half> @llvm.maxnum.v16f16(<16 x half>, <16 x half>)10declare <32 x half> @llvm.maxnum.v32f16(<32 x half>, <32 x half>)11 12define half @test_intrinsic_fmaxh(half %x, half %y) {13; HasVL-LABEL: test_intrinsic_fmaxh:14; HasVL: # %bb.0:15; HasVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm2 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xd0]16; HasVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]17; HasVL-NEXT: vmovsh %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xd1]18; HasVL-NEXT: vmovaps %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc2]19; HasVL-NEXT: retq # encoding: [0xc3]20;21; NOVL-LABEL: test_intrinsic_fmaxh:22; NOVL: # %bb.0:23; NOVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm2 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xd0]24; NOVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]25; NOVL-NEXT: vmovsh %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xd1]26; NOVL-NEXT: vmovaps %xmm2, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc2]27; NOVL-NEXT: retq # encoding: [0xc3]28 %z = call half @llvm.maxnum.f16(half %x, half %y) readnone29 ret half %z30}31 32define <2 x half> @test_intrinsic_fmax_v2f16(<2 x half> %x, <2 x half> %y) {33; HasVL-LABEL: test_intrinsic_fmax_v2f16:34; HasVL: # %bb.0:35; HasVL-NEXT: vmaxph %xmm0, %xmm1, %xmm2 # encoding: [0x62,0xf5,0x74,0x08,0x5f,0xd0]36; HasVL-NEXT: vcmpunordph %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7c,0x08,0xc2,0xc8,0x03]37; HasVL-NEXT: vmovdqu16 %xmm1, %xmm2 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x6f,0xd1]38; HasVL-NEXT: vmovdqa %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]39; HasVL-NEXT: retq # encoding: [0xc3]40;41; NOVL-LABEL: test_intrinsic_fmax_v2f16:42; NOVL: # %bb.0:43; NOVL-NEXT: vpsrldq $14, %xmm0, %xmm2 # encoding: [0xc5,0xe9,0x73,0xd8,0x0e]44; NOVL-NEXT: # xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero45; NOVL-NEXT: vpsrldq $14, %xmm1, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd9,0x0e]46; NOVL-NEXT: # xmm3 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero47; NOVL-NEXT: vmaxsh %xmm2, %xmm3, %xmm4 # encoding: [0x62,0xf5,0x66,0x08,0x5f,0xe2]48; NOVL-NEXT: vcmpunordsh %xmm2, %xmm2, %k1 # encoding: [0x62,0xf3,0x6e,0x08,0xc2,0xca,0x03]49; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe3]50; NOVL-NEXT: vshufps $255, %xmm0, %xmm0, %xmm2 # encoding: [0xc5,0xf8,0xc6,0xd0,0xff]51; NOVL-NEXT: # xmm2 = xmm0[3,3,3,3]52; NOVL-NEXT: vpshufd $255, %xmm1, %xmm3 # encoding: [0xc5,0xf9,0x70,0xd9,0xff]53; NOVL-NEXT: # xmm3 = xmm1[3,3,3,3]54; NOVL-NEXT: vmaxsh %xmm2, %xmm3, %xmm5 # encoding: [0x62,0xf5,0x66,0x08,0x5f,0xea]55; NOVL-NEXT: vcmpunordsh %xmm2, %xmm2, %k1 # encoding: [0x62,0xf3,0x6e,0x08,0xc2,0xca,0x03]56; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xeb]57; NOVL-NEXT: vpunpcklwd %xmm4, %xmm5, %xmm2 # encoding: [0xc5,0xd1,0x61,0xd4]58; NOVL-NEXT: # xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]59; NOVL-NEXT: vpsrldq $10, %xmm0, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd8,0x0a]60; NOVL-NEXT: # xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero61; NOVL-NEXT: vpsrldq $10, %xmm1, %xmm4 # encoding: [0xc5,0xd9,0x73,0xd9,0x0a]62; NOVL-NEXT: # xmm4 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero63; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm5 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xeb]64; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]65; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xec]66; NOVL-NEXT: vshufpd $1, %xmm0, %xmm0, %xmm3 # encoding: [0xc5,0xf9,0xc6,0xd8,0x01]67; NOVL-NEXT: # xmm3 = xmm0[1,0]68; NOVL-NEXT: vshufpd $1, %xmm1, %xmm1, %xmm4 # encoding: [0xc5,0xf1,0xc6,0xe1,0x01]69; NOVL-NEXT: # xmm4 = xmm1[1,0]70; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm6 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xf3]71; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]72; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf4]73; NOVL-NEXT: vpunpcklwd %xmm5, %xmm6, %xmm3 # encoding: [0xc5,0xc9,0x61,0xdd]74; NOVL-NEXT: # xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]75; NOVL-NEXT: vpunpckldq %xmm2, %xmm3, %xmm2 # encoding: [0xc5,0xe1,0x62,0xd2]76; NOVL-NEXT: # xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]77; NOVL-NEXT: vpsrlq $48, %xmm0, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd0,0x30]78; NOVL-NEXT: vpsrlq $48, %xmm1, %xmm4 # encoding: [0xc5,0xd9,0x73,0xd1,0x30]79; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm5 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xeb]80; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]81; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xec]82; NOVL-NEXT: vmovshdup %xmm0, %xmm3 # encoding: [0xc5,0xfa,0x16,0xd8]83; NOVL-NEXT: # xmm3 = xmm0[1,1,3,3]84; NOVL-NEXT: vmovshdup %xmm1, %xmm4 # encoding: [0xc5,0xfa,0x16,0xe1]85; NOVL-NEXT: # xmm4 = xmm1[1,1,3,3]86; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm6 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xf3]87; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]88; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf4]89; NOVL-NEXT: vpunpcklwd %xmm5, %xmm6, %xmm3 # encoding: [0xc5,0xc9,0x61,0xdd]90; NOVL-NEXT: # xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]91; NOVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm4 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xe0]92; NOVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]93; NOVL-NEXT: vmovsh %xmm1, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe1]94; NOVL-NEXT: vpsrld $16, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x72,0xd0,0x10]95; NOVL-NEXT: vpsrld $16, %xmm1, %xmm1 # encoding: [0xc5,0xf1,0x72,0xd1,0x10]96; NOVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm5 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xe8]97; NOVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]98; NOVL-NEXT: vmovsh %xmm1, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe9]99; NOVL-NEXT: vpunpcklwd %xmm5, %xmm4, %xmm0 # encoding: [0xc5,0xd9,0x61,0xc5]100; NOVL-NEXT: # xmm0 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]101; NOVL-NEXT: vpunpckldq %xmm3, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x62,0xc3]102; NOVL-NEXT: # xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]103; NOVL-NEXT: vpunpcklqdq %xmm2, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x6c,0xc2]104; NOVL-NEXT: # xmm0 = xmm0[0],xmm2[0]105; NOVL-NEXT: retq # encoding: [0xc3]106 %z = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %x, <2 x half> %y) readnone107 ret <2 x half> %z108}109 110define <4 x half> @test_intrinsic_fmax_v4f16(<4 x half> %x, <4 x half> %y) {111; HasVL-LABEL: test_intrinsic_fmax_v4f16:112; HasVL: # %bb.0:113; HasVL-NEXT: vmaxph %xmm0, %xmm1, %xmm2 # encoding: [0x62,0xf5,0x74,0x08,0x5f,0xd0]114; HasVL-NEXT: vcmpunordph %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7c,0x08,0xc2,0xc8,0x03]115; HasVL-NEXT: vmovdqu16 %xmm1, %xmm2 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x6f,0xd1]116; HasVL-NEXT: vmovdqa %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]117; HasVL-NEXT: retq # encoding: [0xc3]118;119; NOVL-LABEL: test_intrinsic_fmax_v4f16:120; NOVL: # %bb.0:121; NOVL-NEXT: vpsrldq $14, %xmm0, %xmm2 # encoding: [0xc5,0xe9,0x73,0xd8,0x0e]122; NOVL-NEXT: # xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero123; NOVL-NEXT: vpsrldq $14, %xmm1, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd9,0x0e]124; NOVL-NEXT: # xmm3 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero125; NOVL-NEXT: vmaxsh %xmm2, %xmm3, %xmm4 # encoding: [0x62,0xf5,0x66,0x08,0x5f,0xe2]126; NOVL-NEXT: vcmpunordsh %xmm2, %xmm2, %k1 # encoding: [0x62,0xf3,0x6e,0x08,0xc2,0xca,0x03]127; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe3]128; NOVL-NEXT: vshufps $255, %xmm0, %xmm0, %xmm2 # encoding: [0xc5,0xf8,0xc6,0xd0,0xff]129; NOVL-NEXT: # xmm2 = xmm0[3,3,3,3]130; NOVL-NEXT: vpshufd $255, %xmm1, %xmm3 # encoding: [0xc5,0xf9,0x70,0xd9,0xff]131; NOVL-NEXT: # xmm3 = xmm1[3,3,3,3]132; NOVL-NEXT: vmaxsh %xmm2, %xmm3, %xmm5 # encoding: [0x62,0xf5,0x66,0x08,0x5f,0xea]133; NOVL-NEXT: vcmpunordsh %xmm2, %xmm2, %k1 # encoding: [0x62,0xf3,0x6e,0x08,0xc2,0xca,0x03]134; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xeb]135; NOVL-NEXT: vpunpcklwd %xmm4, %xmm5, %xmm2 # encoding: [0xc5,0xd1,0x61,0xd4]136; NOVL-NEXT: # xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]137; NOVL-NEXT: vpsrldq $10, %xmm0, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd8,0x0a]138; NOVL-NEXT: # xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero139; NOVL-NEXT: vpsrldq $10, %xmm1, %xmm4 # encoding: [0xc5,0xd9,0x73,0xd9,0x0a]140; NOVL-NEXT: # xmm4 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero141; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm5 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xeb]142; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]143; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xec]144; NOVL-NEXT: vshufpd $1, %xmm0, %xmm0, %xmm3 # encoding: [0xc5,0xf9,0xc6,0xd8,0x01]145; NOVL-NEXT: # xmm3 = xmm0[1,0]146; NOVL-NEXT: vshufpd $1, %xmm1, %xmm1, %xmm4 # encoding: [0xc5,0xf1,0xc6,0xe1,0x01]147; NOVL-NEXT: # xmm4 = xmm1[1,0]148; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm6 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xf3]149; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]150; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf4]151; NOVL-NEXT: vpunpcklwd %xmm5, %xmm6, %xmm3 # encoding: [0xc5,0xc9,0x61,0xdd]152; NOVL-NEXT: # xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]153; NOVL-NEXT: vpunpckldq %xmm2, %xmm3, %xmm2 # encoding: [0xc5,0xe1,0x62,0xd2]154; NOVL-NEXT: # xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]155; NOVL-NEXT: vpsrlq $48, %xmm0, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd0,0x30]156; NOVL-NEXT: vpsrlq $48, %xmm1, %xmm4 # encoding: [0xc5,0xd9,0x73,0xd1,0x30]157; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm5 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xeb]158; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]159; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xec]160; NOVL-NEXT: vmovshdup %xmm0, %xmm3 # encoding: [0xc5,0xfa,0x16,0xd8]161; NOVL-NEXT: # xmm3 = xmm0[1,1,3,3]162; NOVL-NEXT: vmovshdup %xmm1, %xmm4 # encoding: [0xc5,0xfa,0x16,0xe1]163; NOVL-NEXT: # xmm4 = xmm1[1,1,3,3]164; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm6 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xf3]165; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]166; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf4]167; NOVL-NEXT: vpunpcklwd %xmm5, %xmm6, %xmm3 # encoding: [0xc5,0xc9,0x61,0xdd]168; NOVL-NEXT: # xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]169; NOVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm4 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xe0]170; NOVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]171; NOVL-NEXT: vmovsh %xmm1, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe1]172; NOVL-NEXT: vpsrld $16, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x72,0xd0,0x10]173; NOVL-NEXT: vpsrld $16, %xmm1, %xmm1 # encoding: [0xc5,0xf1,0x72,0xd1,0x10]174; NOVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm5 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xe8]175; NOVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]176; NOVL-NEXT: vmovsh %xmm1, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe9]177; NOVL-NEXT: vpunpcklwd %xmm5, %xmm4, %xmm0 # encoding: [0xc5,0xd9,0x61,0xc5]178; NOVL-NEXT: # xmm0 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]179; NOVL-NEXT: vpunpckldq %xmm3, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x62,0xc3]180; NOVL-NEXT: # xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]181; NOVL-NEXT: vpunpcklqdq %xmm2, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x6c,0xc2]182; NOVL-NEXT: # xmm0 = xmm0[0],xmm2[0]183; NOVL-NEXT: retq # encoding: [0xc3]184 %z = call <4 x half> @llvm.maxnum.v4f16(<4 x half> %x, <4 x half> %y) readnone185 ret <4 x half> %z186}187 188define <8 x half> @test_intrinsic_fmax_v8f16(<8 x half> %x, <8 x half> %y) {189; HasVL-LABEL: test_intrinsic_fmax_v8f16:190; HasVL: # %bb.0:191; HasVL-NEXT: vmaxph %xmm0, %xmm1, %xmm2 # encoding: [0x62,0xf5,0x74,0x08,0x5f,0xd0]192; HasVL-NEXT: vcmpunordph %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7c,0x08,0xc2,0xc8,0x03]193; HasVL-NEXT: vmovdqu16 %xmm1, %xmm2 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x6f,0xd1]194; HasVL-NEXT: vmovdqa %xmm2, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0xc2]195; HasVL-NEXT: retq # encoding: [0xc3]196;197; NOVL-LABEL: test_intrinsic_fmax_v8f16:198; NOVL: # %bb.0:199; NOVL-NEXT: vpsrldq $14, %xmm0, %xmm2 # encoding: [0xc5,0xe9,0x73,0xd8,0x0e]200; NOVL-NEXT: # xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero201; NOVL-NEXT: vpsrldq $14, %xmm1, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd9,0x0e]202; NOVL-NEXT: # xmm3 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero203; NOVL-NEXT: vmaxsh %xmm2, %xmm3, %xmm4 # encoding: [0x62,0xf5,0x66,0x08,0x5f,0xe2]204; NOVL-NEXT: vcmpunordsh %xmm2, %xmm2, %k1 # encoding: [0x62,0xf3,0x6e,0x08,0xc2,0xca,0x03]205; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe3]206; NOVL-NEXT: vshufps $255, %xmm0, %xmm0, %xmm2 # encoding: [0xc5,0xf8,0xc6,0xd0,0xff]207; NOVL-NEXT: # xmm2 = xmm0[3,3,3,3]208; NOVL-NEXT: vpshufd $255, %xmm1, %xmm3 # encoding: [0xc5,0xf9,0x70,0xd9,0xff]209; NOVL-NEXT: # xmm3 = xmm1[3,3,3,3]210; NOVL-NEXT: vmaxsh %xmm2, %xmm3, %xmm5 # encoding: [0x62,0xf5,0x66,0x08,0x5f,0xea]211; NOVL-NEXT: vcmpunordsh %xmm2, %xmm2, %k1 # encoding: [0x62,0xf3,0x6e,0x08,0xc2,0xca,0x03]212; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xeb]213; NOVL-NEXT: vpunpcklwd %xmm4, %xmm5, %xmm2 # encoding: [0xc5,0xd1,0x61,0xd4]214; NOVL-NEXT: # xmm2 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]215; NOVL-NEXT: vpsrldq $10, %xmm0, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd8,0x0a]216; NOVL-NEXT: # xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero217; NOVL-NEXT: vpsrldq $10, %xmm1, %xmm4 # encoding: [0xc5,0xd9,0x73,0xd9,0x0a]218; NOVL-NEXT: # xmm4 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero219; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm5 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xeb]220; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]221; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xec]222; NOVL-NEXT: vshufpd $1, %xmm0, %xmm0, %xmm3 # encoding: [0xc5,0xf9,0xc6,0xd8,0x01]223; NOVL-NEXT: # xmm3 = xmm0[1,0]224; NOVL-NEXT: vshufpd $1, %xmm1, %xmm1, %xmm4 # encoding: [0xc5,0xf1,0xc6,0xe1,0x01]225; NOVL-NEXT: # xmm4 = xmm1[1,0]226; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm6 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xf3]227; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]228; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf4]229; NOVL-NEXT: vpunpcklwd %xmm5, %xmm6, %xmm3 # encoding: [0xc5,0xc9,0x61,0xdd]230; NOVL-NEXT: # xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]231; NOVL-NEXT: vpunpckldq %xmm2, %xmm3, %xmm2 # encoding: [0xc5,0xe1,0x62,0xd2]232; NOVL-NEXT: # xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]233; NOVL-NEXT: vpsrlq $48, %xmm0, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd0,0x30]234; NOVL-NEXT: vpsrlq $48, %xmm1, %xmm4 # encoding: [0xc5,0xd9,0x73,0xd1,0x30]235; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm5 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xeb]236; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]237; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xec]238; NOVL-NEXT: vmovshdup %xmm0, %xmm3 # encoding: [0xc5,0xfa,0x16,0xd8]239; NOVL-NEXT: # xmm3 = xmm0[1,1,3,3]240; NOVL-NEXT: vmovshdup %xmm1, %xmm4 # encoding: [0xc5,0xfa,0x16,0xe1]241; NOVL-NEXT: # xmm4 = xmm1[1,1,3,3]242; NOVL-NEXT: vmaxsh %xmm3, %xmm4, %xmm6 # encoding: [0x62,0xf5,0x5e,0x08,0x5f,0xf3]243; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]244; NOVL-NEXT: vmovsh %xmm4, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf4]245; NOVL-NEXT: vpunpcklwd %xmm5, %xmm6, %xmm3 # encoding: [0xc5,0xc9,0x61,0xdd]246; NOVL-NEXT: # xmm3 = xmm6[0],xmm5[0],xmm6[1],xmm5[1],xmm6[2],xmm5[2],xmm6[3],xmm5[3]247; NOVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm4 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xe0]248; NOVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]249; NOVL-NEXT: vmovsh %xmm1, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe1]250; NOVL-NEXT: vpsrld $16, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x72,0xd0,0x10]251; NOVL-NEXT: vpsrld $16, %xmm1, %xmm1 # encoding: [0xc5,0xf1,0x72,0xd1,0x10]252; NOVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm5 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xe8]253; NOVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]254; NOVL-NEXT: vmovsh %xmm1, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe9]255; NOVL-NEXT: vpunpcklwd %xmm5, %xmm4, %xmm0 # encoding: [0xc5,0xd9,0x61,0xc5]256; NOVL-NEXT: # xmm0 = xmm4[0],xmm5[0],xmm4[1],xmm5[1],xmm4[2],xmm5[2],xmm4[3],xmm5[3]257; NOVL-NEXT: vpunpckldq %xmm3, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x62,0xc3]258; NOVL-NEXT: # xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]259; NOVL-NEXT: vpunpcklqdq %xmm2, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x6c,0xc2]260; NOVL-NEXT: # xmm0 = xmm0[0],xmm2[0]261; NOVL-NEXT: retq # encoding: [0xc3]262 %z = call <8 x half> @llvm.maxnum.v8f16(<8 x half> %x, <8 x half> %y) readnone263 ret <8 x half> %z264}265 266define <16 x half> @test_intrinsic_fmax_v16f16(<16 x half> %x, <16 x half> %y) {267; HasVL-LABEL: test_intrinsic_fmax_v16f16:268; HasVL: # %bb.0:269; HasVL-NEXT: vmaxph %ymm0, %ymm1, %ymm2 # encoding: [0x62,0xf5,0x74,0x28,0x5f,0xd0]270; HasVL-NEXT: vcmpunordph %ymm0, %ymm0, %k1 # encoding: [0x62,0xf3,0x7c,0x28,0xc2,0xc8,0x03]271; HasVL-NEXT: vmovdqu16 %ymm1, %ymm2 {%k1} # encoding: [0x62,0xf1,0xff,0x29,0x6f,0xd1]272; HasVL-NEXT: vmovdqa %ymm2, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0x6f,0xc2]273; HasVL-NEXT: retq # encoding: [0xc3]274;275; NOVL-LABEL: test_intrinsic_fmax_v16f16:276; NOVL: # %bb.0:277; NOVL-NEXT: vextracti128 $1, %ymm0, %xmm2 # encoding: [0xc4,0xe3,0x7d,0x39,0xc2,0x01]278; NOVL-NEXT: vpsrldq $14, %xmm2, %xmm4 # encoding: [0xc5,0xd9,0x73,0xda,0x0e]279; NOVL-NEXT: # xmm4 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero280; NOVL-NEXT: vextracti128 $1, %ymm1, %xmm3 # encoding: [0xc4,0xe3,0x7d,0x39,0xcb,0x01]281; NOVL-NEXT: vpsrldq $14, %xmm3, %xmm5 # encoding: [0xc5,0xd1,0x73,0xdb,0x0e]282; NOVL-NEXT: # xmm5 = xmm3[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero283; NOVL-NEXT: vmaxsh %xmm4, %xmm5, %xmm6 # encoding: [0x62,0xf5,0x56,0x08,0x5f,0xf4]284; NOVL-NEXT: vcmpunordsh %xmm4, %xmm4, %k1 # encoding: [0x62,0xf3,0x5e,0x08,0xc2,0xcc,0x03]285; NOVL-NEXT: vmovsh %xmm5, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf5]286; NOVL-NEXT: vpshufd $255, %xmm2, %xmm4 # encoding: [0xc5,0xf9,0x70,0xe2,0xff]287; NOVL-NEXT: # xmm4 = xmm2[3,3,3,3]288; NOVL-NEXT: vpshufd $255, %xmm3, %xmm5 # encoding: [0xc5,0xf9,0x70,0xeb,0xff]289; NOVL-NEXT: # xmm5 = xmm3[3,3,3,3]290; NOVL-NEXT: vmaxsh %xmm4, %xmm5, %xmm7 # encoding: [0x62,0xf5,0x56,0x08,0x5f,0xfc]291; NOVL-NEXT: vcmpunordsh %xmm4, %xmm4, %k1 # encoding: [0x62,0xf3,0x5e,0x08,0xc2,0xcc,0x03]292; NOVL-NEXT: vmovsh %xmm5, %xmm0, %xmm7 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xfd]293; NOVL-NEXT: vpunpcklwd %xmm6, %xmm7, %xmm4 # encoding: [0xc5,0xc1,0x61,0xe6]294; NOVL-NEXT: # xmm4 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]295; NOVL-NEXT: vpsrldq $10, %xmm2, %xmm5 # encoding: [0xc5,0xd1,0x73,0xda,0x0a]296; NOVL-NEXT: # xmm5 = xmm2[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero297; NOVL-NEXT: vpsrldq $10, %xmm3, %xmm6 # encoding: [0xc5,0xc9,0x73,0xdb,0x0a]298; NOVL-NEXT: # xmm6 = xmm3[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero299; NOVL-NEXT: vmaxsh %xmm5, %xmm6, %xmm7 # encoding: [0x62,0xf5,0x4e,0x08,0x5f,0xfd]300; NOVL-NEXT: vcmpunordsh %xmm5, %xmm5, %k1 # encoding: [0x62,0xf3,0x56,0x08,0xc2,0xcd,0x03]301; NOVL-NEXT: vmovsh %xmm6, %xmm0, %xmm7 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xfe]302; NOVL-NEXT: vshufpd $1, %xmm2, %xmm2, %xmm5 # encoding: [0xc5,0xe9,0xc6,0xea,0x01]303; NOVL-NEXT: # xmm5 = xmm2[1,0]304; NOVL-NEXT: vshufpd $1, %xmm3, %xmm3, %xmm6 # encoding: [0xc5,0xe1,0xc6,0xf3,0x01]305; NOVL-NEXT: # xmm6 = xmm3[1,0]306; NOVL-NEXT: vmaxsh %xmm5, %xmm6, %xmm8 # encoding: [0x62,0x75,0x4e,0x08,0x5f,0xc5]307; NOVL-NEXT: vcmpunordsh %xmm5, %xmm5, %k1 # encoding: [0x62,0xf3,0x56,0x08,0xc2,0xcd,0x03]308; NOVL-NEXT: vmovsh %xmm6, %xmm0, %xmm8 {%k1} # encoding: [0x62,0x75,0x7e,0x09,0x10,0xc6]309; NOVL-NEXT: vpunpcklwd %xmm7, %xmm8, %xmm5 # encoding: [0xc5,0xb9,0x61,0xef]310; NOVL-NEXT: # xmm5 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]311; NOVL-NEXT: vpunpckldq %xmm4, %xmm5, %xmm4 # encoding: [0xc5,0xd1,0x62,0xe4]312; NOVL-NEXT: # xmm4 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]313; NOVL-NEXT: vpsrldq $14, %xmm0, %xmm5 # encoding: [0xc5,0xd1,0x73,0xd8,0x0e]314; NOVL-NEXT: # xmm5 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero315; NOVL-NEXT: vpsrldq $14, %xmm1, %xmm6 # encoding: [0xc5,0xc9,0x73,0xd9,0x0e]316; NOVL-NEXT: # xmm6 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero317; NOVL-NEXT: vmaxsh %xmm5, %xmm6, %xmm7 # encoding: [0x62,0xf5,0x4e,0x08,0x5f,0xfd]318; NOVL-NEXT: vcmpunordsh %xmm5, %xmm5, %k1 # encoding: [0x62,0xf3,0x56,0x08,0xc2,0xcd,0x03]319; NOVL-NEXT: vmovsh %xmm6, %xmm0, %xmm7 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xfe]320; NOVL-NEXT: vshufps $255, %xmm0, %xmm0, %xmm5 # encoding: [0xc5,0xf8,0xc6,0xe8,0xff]321; NOVL-NEXT: # xmm5 = xmm0[3,3,3,3]322; NOVL-NEXT: vpshufd $255, %xmm1, %xmm6 # encoding: [0xc5,0xf9,0x70,0xf1,0xff]323; NOVL-NEXT: # xmm6 = xmm1[3,3,3,3]324; NOVL-NEXT: vmaxsh %xmm5, %xmm6, %xmm8 # encoding: [0x62,0x75,0x4e,0x08,0x5f,0xc5]325; NOVL-NEXT: vcmpunordsh %xmm5, %xmm5, %k1 # encoding: [0x62,0xf3,0x56,0x08,0xc2,0xcd,0x03]326; NOVL-NEXT: vmovsh %xmm6, %xmm0, %xmm8 {%k1} # encoding: [0x62,0x75,0x7e,0x09,0x10,0xc6]327; NOVL-NEXT: vpunpcklwd %xmm7, %xmm8, %xmm5 # encoding: [0xc5,0xb9,0x61,0xef]328; NOVL-NEXT: # xmm5 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]329; NOVL-NEXT: vpsrldq $10, %xmm0, %xmm6 # encoding: [0xc5,0xc9,0x73,0xd8,0x0a]330; NOVL-NEXT: # xmm6 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero331; NOVL-NEXT: vpsrldq $10, %xmm1, %xmm7 # encoding: [0xc5,0xc1,0x73,0xd9,0x0a]332; NOVL-NEXT: # xmm7 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero333; NOVL-NEXT: vmaxsh %xmm6, %xmm7, %xmm8 # encoding: [0x62,0x75,0x46,0x08,0x5f,0xc6]334; NOVL-NEXT: vcmpunordsh %xmm6, %xmm6, %k1 # encoding: [0x62,0xf3,0x4e,0x08,0xc2,0xce,0x03]335; NOVL-NEXT: vmovsh %xmm7, %xmm0, %xmm8 {%k1} # encoding: [0x62,0x75,0x7e,0x09,0x10,0xc7]336; NOVL-NEXT: vshufpd $1, %xmm0, %xmm0, %xmm6 # encoding: [0xc5,0xf9,0xc6,0xf0,0x01]337; NOVL-NEXT: # xmm6 = xmm0[1,0]338; NOVL-NEXT: vshufpd $1, %xmm1, %xmm1, %xmm7 # encoding: [0xc5,0xf1,0xc6,0xf9,0x01]339; NOVL-NEXT: # xmm7 = xmm1[1,0]340; NOVL-NEXT: vmaxsh %xmm6, %xmm7, %xmm9 # encoding: [0x62,0x75,0x46,0x08,0x5f,0xce]341; NOVL-NEXT: vcmpunordsh %xmm6, %xmm6, %k1 # encoding: [0x62,0xf3,0x4e,0x08,0xc2,0xce,0x03]342; NOVL-NEXT: vmovsh %xmm7, %xmm0, %xmm9 {%k1} # encoding: [0x62,0x75,0x7e,0x09,0x10,0xcf]343; NOVL-NEXT: vpunpcklwd %xmm8, %xmm9, %xmm6 # encoding: [0xc4,0xc1,0x31,0x61,0xf0]344; NOVL-NEXT: # xmm6 = xmm9[0],xmm8[0],xmm9[1],xmm8[1],xmm9[2],xmm8[2],xmm9[3],xmm8[3]345; NOVL-NEXT: vpunpckldq %xmm5, %xmm6, %xmm5 # encoding: [0xc5,0xc9,0x62,0xed]346; NOVL-NEXT: # xmm5 = xmm6[0],xmm5[0],xmm6[1],xmm5[1]347; NOVL-NEXT: vinserti128 $1, %xmm4, %ymm5, %ymm4 # encoding: [0xc4,0xe3,0x55,0x38,0xe4,0x01]348; NOVL-NEXT: vpsrlq $48, %xmm2, %xmm5 # encoding: [0xc5,0xd1,0x73,0xd2,0x30]349; NOVL-NEXT: vpsrlq $48, %xmm3, %xmm6 # encoding: [0xc5,0xc9,0x73,0xd3,0x30]350; NOVL-NEXT: vmaxsh %xmm5, %xmm6, %xmm7 # encoding: [0x62,0xf5,0x4e,0x08,0x5f,0xfd]351; NOVL-NEXT: vcmpunordsh %xmm5, %xmm5, %k1 # encoding: [0x62,0xf3,0x56,0x08,0xc2,0xcd,0x03]352; NOVL-NEXT: vmovsh %xmm6, %xmm0, %xmm7 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xfe]353; NOVL-NEXT: vmovshdup %xmm2, %xmm5 # encoding: [0xc5,0xfa,0x16,0xea]354; NOVL-NEXT: # xmm5 = xmm2[1,1,3,3]355; NOVL-NEXT: vmovshdup %xmm3, %xmm6 # encoding: [0xc5,0xfa,0x16,0xf3]356; NOVL-NEXT: # xmm6 = xmm3[1,1,3,3]357; NOVL-NEXT: vmaxsh %xmm5, %xmm6, %xmm8 # encoding: [0x62,0x75,0x4e,0x08,0x5f,0xc5]358; NOVL-NEXT: vcmpunordsh %xmm5, %xmm5, %k1 # encoding: [0x62,0xf3,0x56,0x08,0xc2,0xcd,0x03]359; NOVL-NEXT: vmovsh %xmm6, %xmm0, %xmm8 {%k1} # encoding: [0x62,0x75,0x7e,0x09,0x10,0xc6]360; NOVL-NEXT: vpunpcklwd %xmm7, %xmm8, %xmm5 # encoding: [0xc5,0xb9,0x61,0xef]361; NOVL-NEXT: # xmm5 = xmm8[0],xmm7[0],xmm8[1],xmm7[1],xmm8[2],xmm7[2],xmm8[3],xmm7[3]362; NOVL-NEXT: vmaxsh %xmm2, %xmm3, %xmm6 # encoding: [0x62,0xf5,0x66,0x08,0x5f,0xf2]363; NOVL-NEXT: vcmpunordsh %xmm2, %xmm2, %k1 # encoding: [0x62,0xf3,0x6e,0x08,0xc2,0xca,0x03]364; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf3]365; NOVL-NEXT: vpsrld $16, %xmm2, %xmm2 # encoding: [0xc5,0xe9,0x72,0xd2,0x10]366; NOVL-NEXT: vpsrld $16, %xmm3, %xmm3 # encoding: [0xc5,0xe1,0x72,0xd3,0x10]367; NOVL-NEXT: vmaxsh %xmm2, %xmm3, %xmm7 # encoding: [0x62,0xf5,0x66,0x08,0x5f,0xfa]368; NOVL-NEXT: vcmpunordsh %xmm2, %xmm2, %k1 # encoding: [0x62,0xf3,0x6e,0x08,0xc2,0xca,0x03]369; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm7 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xfb]370; NOVL-NEXT: vpunpcklwd %xmm7, %xmm6, %xmm2 # encoding: [0xc5,0xc9,0x61,0xd7]371; NOVL-NEXT: # xmm2 = xmm6[0],xmm7[0],xmm6[1],xmm7[1],xmm6[2],xmm7[2],xmm6[3],xmm7[3]372; NOVL-NEXT: vpunpckldq %xmm5, %xmm2, %xmm2 # encoding: [0xc5,0xe9,0x62,0xd5]373; NOVL-NEXT: # xmm2 = xmm2[0],xmm5[0],xmm2[1],xmm5[1]374; NOVL-NEXT: vpsrlq $48, %xmm0, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd0,0x30]375; NOVL-NEXT: vpsrlq $48, %xmm1, %xmm5 # encoding: [0xc5,0xd1,0x73,0xd1,0x30]376; NOVL-NEXT: vmaxsh %xmm3, %xmm5, %xmm6 # encoding: [0x62,0xf5,0x56,0x08,0x5f,0xf3]377; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]378; NOVL-NEXT: vmovsh %xmm5, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf5]379; NOVL-NEXT: vmovshdup %xmm0, %xmm3 # encoding: [0xc5,0xfa,0x16,0xd8]380; NOVL-NEXT: # xmm3 = xmm0[1,1,3,3]381; NOVL-NEXT: vmovshdup %xmm1, %xmm5 # encoding: [0xc5,0xfa,0x16,0xe9]382; NOVL-NEXT: # xmm5 = xmm1[1,1,3,3]383; NOVL-NEXT: vmaxsh %xmm3, %xmm5, %xmm7 # encoding: [0x62,0xf5,0x56,0x08,0x5f,0xfb]384; NOVL-NEXT: vcmpunordsh %xmm3, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xcb,0x03]385; NOVL-NEXT: vmovsh %xmm5, %xmm0, %xmm7 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xfd]386; NOVL-NEXT: vpunpcklwd %xmm6, %xmm7, %xmm3 # encoding: [0xc5,0xc1,0x61,0xde]387; NOVL-NEXT: # xmm3 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]388; NOVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm5 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xe8]389; NOVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]390; NOVL-NEXT: vmovsh %xmm1, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe9]391; NOVL-NEXT: vpsrld $16, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x72,0xd0,0x10]392; NOVL-NEXT: vpsrld $16, %xmm1, %xmm1 # encoding: [0xc5,0xf1,0x72,0xd1,0x10]393; NOVL-NEXT: vmaxsh %xmm0, %xmm1, %xmm6 # encoding: [0x62,0xf5,0x76,0x08,0x5f,0xf0]394; NOVL-NEXT: vcmpunordsh %xmm0, %xmm0, %k1 # encoding: [0x62,0xf3,0x7e,0x08,0xc2,0xc8,0x03]395; NOVL-NEXT: vmovsh %xmm1, %xmm0, %xmm6 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xf1]396; NOVL-NEXT: vpunpcklwd %xmm6, %xmm5, %xmm0 # encoding: [0xc5,0xd1,0x61,0xc6]397; NOVL-NEXT: # xmm0 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3]398; NOVL-NEXT: vpunpckldq %xmm3, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x62,0xc3]399; NOVL-NEXT: # xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]400; NOVL-NEXT: vinserti128 $1, %xmm2, %ymm0, %ymm0 # encoding: [0xc4,0xe3,0x7d,0x38,0xc2,0x01]401; NOVL-NEXT: vpunpcklqdq %ymm4, %ymm0, %ymm0 # encoding: [0xc5,0xfd,0x6c,0xc4]402; NOVL-NEXT: # ymm0 = ymm0[0],ymm4[0],ymm0[2],ymm4[2]403; NOVL-NEXT: retq # encoding: [0xc3]404 %z = call <16 x half> @llvm.maxnum.v16f16(<16 x half> %x, <16 x half> %y) readnone405 ret <16 x half> %z406}407 408define <32 x half> @test_intrinsic_fmax_v32f16(<32 x half> %x, <32 x half> %y) {409; CHECK-LABEL: test_intrinsic_fmax_v32f16:410; CHECK: # %bb.0:411; CHECK-NEXT: vmaxph %zmm0, %zmm1, %zmm2 # encoding: [0x62,0xf5,0x74,0x48,0x5f,0xd0]412; CHECK-NEXT: vcmpunordph %zmm0, %zmm0, %k1 # encoding: [0x62,0xf3,0x7c,0x48,0xc2,0xc8,0x03]413; CHECK-NEXT: vmovdqu16 %zmm1, %zmm2 {%k1} # encoding: [0x62,0xf1,0xff,0x49,0x6f,0xd1]414; CHECK-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]415; CHECK-NEXT: retq # encoding: [0xc3]416 %z = call <32 x half> @llvm.maxnum.v32f16(<32 x half> %x, <32 x half> %y) readnone417 ret <32 x half> %z418}419 420define <4 x half> @maxnum_intrinsic_nnan_fmf_f432(<4 x half> %a, <4 x half> %b) {421; HasVL-LABEL: maxnum_intrinsic_nnan_fmf_f432:422; HasVL: # %bb.0:423; HasVL-NEXT: vmaxph %xmm1, %xmm0, %xmm0 # encoding: [0x62,0xf5,0x7c,0x08,0x5f,0xc1]424; HasVL-NEXT: retq # encoding: [0xc3]425;426; NOVL-LABEL: maxnum_intrinsic_nnan_fmf_f432:427; NOVL: # %bb.0:428; NOVL-NEXT: vpsrldq $14, %xmm0, %xmm2 # encoding: [0xc5,0xe9,0x73,0xd8,0x0e]429; NOVL-NEXT: # xmm2 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero430; NOVL-NEXT: vpsrldq $14, %xmm1, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd9,0x0e]431; NOVL-NEXT: # xmm3 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero432; NOVL-NEXT: vcmpltsh %xmm2, %xmm3, %k1 # encoding: [0x62,0xf3,0x66,0x08,0xc2,0xca,0x01]433; NOVL-NEXT: vmovsh %xmm2, %xmm0, %xmm3 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xda]434; NOVL-NEXT: vshufps $255, %xmm0, %xmm0, %xmm2 # encoding: [0xc5,0xf8,0xc6,0xd0,0xff]435; NOVL-NEXT: # xmm2 = xmm0[3,3,3,3]436; NOVL-NEXT: vpshufd $255, %xmm1, %xmm4 # encoding: [0xc5,0xf9,0x70,0xe1,0xff]437; NOVL-NEXT: # xmm4 = xmm1[3,3,3,3]438; NOVL-NEXT: vcmpltsh %xmm2, %xmm4, %k1 # encoding: [0x62,0xf3,0x5e,0x08,0xc2,0xca,0x01]439; NOVL-NEXT: vmovsh %xmm2, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe2]440; NOVL-NEXT: vpunpcklwd %xmm3, %xmm4, %xmm2 # encoding: [0xc5,0xd9,0x61,0xd3]441; NOVL-NEXT: # xmm2 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]442; NOVL-NEXT: vpsrldq $10, %xmm0, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd8,0x0a]443; NOVL-NEXT: # xmm3 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero444; NOVL-NEXT: vpsrldq $10, %xmm1, %xmm4 # encoding: [0xc5,0xd9,0x73,0xd9,0x0a]445; NOVL-NEXT: # xmm4 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero446; NOVL-NEXT: vcmpltsh %xmm3, %xmm4, %k1 # encoding: [0x62,0xf3,0x5e,0x08,0xc2,0xcb,0x01]447; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe3]448; NOVL-NEXT: vshufpd $1, %xmm0, %xmm0, %xmm3 # encoding: [0xc5,0xf9,0xc6,0xd8,0x01]449; NOVL-NEXT: # xmm3 = xmm0[1,0]450; NOVL-NEXT: vshufpd $1, %xmm1, %xmm1, %xmm5 # encoding: [0xc5,0xf1,0xc6,0xe9,0x01]451; NOVL-NEXT: # xmm5 = xmm1[1,0]452; NOVL-NEXT: vcmpltsh %xmm3, %xmm5, %k1 # encoding: [0x62,0xf3,0x56,0x08,0xc2,0xcb,0x01]453; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xeb]454; NOVL-NEXT: vpunpcklwd %xmm4, %xmm5, %xmm3 # encoding: [0xc5,0xd1,0x61,0xdc]455; NOVL-NEXT: # xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]456; NOVL-NEXT: vpunpckldq %xmm2, %xmm3, %xmm2 # encoding: [0xc5,0xe1,0x62,0xd2]457; NOVL-NEXT: # xmm2 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]458; NOVL-NEXT: vpsrlq $48, %xmm0, %xmm3 # encoding: [0xc5,0xe1,0x73,0xd0,0x30]459; NOVL-NEXT: vpsrlq $48, %xmm1, %xmm4 # encoding: [0xc5,0xd9,0x73,0xd1,0x30]460; NOVL-NEXT: vcmpltsh %xmm3, %xmm4, %k1 # encoding: [0x62,0xf3,0x5e,0x08,0xc2,0xcb,0x01]461; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe3]462; NOVL-NEXT: vmovshdup %xmm0, %xmm3 # encoding: [0xc5,0xfa,0x16,0xd8]463; NOVL-NEXT: # xmm3 = xmm0[1,1,3,3]464; NOVL-NEXT: vmovshdup %xmm1, %xmm5 # encoding: [0xc5,0xfa,0x16,0xe9]465; NOVL-NEXT: # xmm5 = xmm1[1,1,3,3]466; NOVL-NEXT: vcmpltsh %xmm3, %xmm5, %k1 # encoding: [0x62,0xf3,0x56,0x08,0xc2,0xcb,0x01]467; NOVL-NEXT: vmovsh %xmm3, %xmm0, %xmm5 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xeb]468; NOVL-NEXT: vpunpcklwd %xmm4, %xmm5, %xmm3 # encoding: [0xc5,0xd1,0x61,0xdc]469; NOVL-NEXT: # xmm3 = xmm5[0],xmm4[0],xmm5[1],xmm4[1],xmm5[2],xmm4[2],xmm5[3],xmm4[3]470; NOVL-NEXT: vcmpltsh %xmm0, %xmm1, %k1 # encoding: [0x62,0xf3,0x76,0x08,0xc2,0xc8,0x01]471; NOVL-NEXT: vpsrld $16, %xmm1, %xmm4 # encoding: [0xc5,0xd9,0x72,0xd1,0x10]472; NOVL-NEXT: vmovsh %xmm0, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xc8]473; NOVL-NEXT: vpsrld $16, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x72,0xd0,0x10]474; NOVL-NEXT: vcmpltsh %xmm0, %xmm4, %k1 # encoding: [0x62,0xf3,0x5e,0x08,0xc2,0xc8,0x01]475; NOVL-NEXT: vmovsh %xmm0, %xmm0, %xmm4 {%k1} # encoding: [0x62,0xf5,0x7e,0x09,0x10,0xe0]476; NOVL-NEXT: vpunpcklwd %xmm4, %xmm1, %xmm0 # encoding: [0xc5,0xf1,0x61,0xc4]477; NOVL-NEXT: # xmm0 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]478; NOVL-NEXT: vpunpckldq %xmm3, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x62,0xc3]479; NOVL-NEXT: # xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]480; NOVL-NEXT: vpunpcklqdq %xmm2, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x6c,0xc2]481; NOVL-NEXT: # xmm0 = xmm0[0],xmm2[0]482; NOVL-NEXT: retq # encoding: [0xc3]483 %r = tail call nnan <4 x half> @llvm.maxnum.v4f16(<4 x half> %a, <4 x half> %b)484 ret <4 x half> %r485}486 487define half @maxnum_intrinsic_nnan_attr_f16(half %a, half %b) #0 {488; CHECK-LABEL: maxnum_intrinsic_nnan_attr_f16:489; CHECK: # %bb.0:490; CHECK-NEXT: vmaxsh %xmm1, %xmm0, %xmm0 # encoding: [0x62,0xf5,0x7e,0x08,0x5f,0xc1]491; CHECK-NEXT: retq # encoding: [0xc3]492 %r = tail call half @llvm.maxnum.f16(half %a, half %b)493 ret half %r494}495 496define half @test_maxnum_const_op1(half %x) {497; CHECK-LABEL: test_maxnum_const_op1:498; CHECK: # %bb.0:499; CHECK-NEXT: vmaxsh {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0 # encoding: [0x62,0xf5,0x7e,0x08,0x5f,0x05,A,A,A,A]500; CHECK-NEXT: # fixup A - offset: 6, value: {{\.?LCPI[0-9]+_[0-9]+}}, kind: reloc_riprel_4byte501; CHECK-NEXT: retq # encoding: [0xc3]502 %r = call half @llvm.maxnum.f16(half 1.0, half %x)503 ret half %r504}505 506define half @test_maxnum_const_op2(half %x) {507; CHECK-LABEL: test_maxnum_const_op2:508; CHECK: # %bb.0:509; CHECK-NEXT: vmaxsh {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0 # encoding: [0x62,0xf5,0x7e,0x08,0x5f,0x05,A,A,A,A]510; CHECK-NEXT: # fixup A - offset: 6, value: {{\.?LCPI[0-9]+_[0-9]+}}, kind: reloc_riprel_4byte511; CHECK-NEXT: retq # encoding: [0xc3]512 %r = call half @llvm.maxnum.f16(half %x, half 1.0)513 ret half %r514}515 516define half @test_maxnum_const_nan(half %x) {517; CHECK-LABEL: test_maxnum_const_nan:518; CHECK: # %bb.0:519; CHECK-NEXT: retq # encoding: [0xc3]520 %r = call half @llvm.maxnum.f16(half %x, half 0x7fff000000000000)521 ret half %r522}523 524attributes #0 = { "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true"}525