217 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 62; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 -verify-machineinstrs | FileCheck %s --check-prefixes=SSE3; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx -verify-machineinstrs | FileCheck %s --check-prefixes=AVX,AVX14; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 -verify-machineinstrs | FileCheck %s --check-prefixes=AVX,AVX25; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512vl -verify-machineinstrs | FileCheck %s --check-prefixes=AVX512,AVX512VL6; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512vl,+avx512bw -verify-machineinstrs | FileCheck %s --check-prefixes=AVX512,AVX512BW7 8define void @zero_xmm(<4 x i32> %arg) #0 {9; SSE-LABEL: zero_xmm:10; SSE: # %bb.0:11; SSE-NEXT: movaps %xmm0, 012; SSE-NEXT: xorps %xmm0, %xmm013; SSE-NEXT: retq14;15; AVX-LABEL: zero_xmm:16; AVX: # %bb.0:17; AVX-NEXT: vmovaps %xmm0, 018; AVX-NEXT: vxorps %xmm0, %xmm0, %xmm019; AVX-NEXT: retq20;21; AVX512-LABEL: zero_xmm:22; AVX512: # %bb.0:23; AVX512-NEXT: vmovaps %xmm0, 024; AVX512-NEXT: vxorps %xmm0, %xmm0, %xmm025; AVX512-NEXT: retq26 store <4 x i32> %arg, ptr null, align 3227 ret void28}29 30define void @zero_ymm(<8 x i32> %arg) #0 {31; SSE-LABEL: zero_ymm:32; SSE: # %bb.0:33; SSE-NEXT: movaps %xmm1, 1634; SSE-NEXT: movaps %xmm0, 035; SSE-NEXT: xorps %xmm0, %xmm036; SSE-NEXT: xorps %xmm1, %xmm137; SSE-NEXT: retq38;39; AVX-LABEL: zero_ymm:40; AVX: # %bb.0:41; AVX-NEXT: vmovaps %ymm0, 042; AVX-NEXT: vxorps %xmm0, %xmm0, %xmm043; AVX-NEXT: vzeroupper44; AVX-NEXT: retq45;46; AVX512-LABEL: zero_ymm:47; AVX512: # %bb.0:48; AVX512-NEXT: vmovaps %ymm0, 049; AVX512-NEXT: vxorps %xmm0, %xmm0, %xmm050; AVX512-NEXT: vzeroupper51; AVX512-NEXT: retq52 store <8 x i32> %arg, ptr null, align 3253 ret void54}55 56define void @zero_zmm(<16 x i32> %arg) #0 {57; SSE-LABEL: zero_zmm:58; SSE: # %bb.0:59; SSE-NEXT: movaps %xmm3, 4860; SSE-NEXT: movaps %xmm2, 3261; SSE-NEXT: movaps %xmm1, 1662; SSE-NEXT: movaps %xmm0, 063; SSE-NEXT: xorps %xmm0, %xmm064; SSE-NEXT: xorps %xmm1, %xmm165; SSE-NEXT: xorps %xmm2, %xmm266; SSE-NEXT: xorps %xmm3, %xmm367; SSE-NEXT: retq68;69; AVX-LABEL: zero_zmm:70; AVX: # %bb.0:71; AVX-NEXT: vmovaps %ymm1, 3272; AVX-NEXT: vmovaps %ymm0, 073; AVX-NEXT: vxorps %xmm0, %xmm0, %xmm074; AVX-NEXT: vxorps %xmm1, %xmm1, %xmm175; AVX-NEXT: vzeroupper76; AVX-NEXT: retq77;78; AVX512-LABEL: zero_zmm:79; AVX512: # %bb.0:80; AVX512-NEXT: vmovups %zmm0, 081; AVX512-NEXT: vxorps %xmm0, %xmm0, %xmm082; AVX512-NEXT: vzeroupper83; AVX512-NEXT: retq84 store <16 x i32> %arg, ptr null, align 3285 ret void86}87 88define void @zero_k(<8 x i32> %arg, <8 x i1> %mask) #0 {89; SSE-LABEL: zero_k:90; SSE: # %bb.0:91; SSE-NEXT: psllw $15, %xmm292; SSE-NEXT: packsswb %xmm2, %xmm293; SSE-NEXT: pmovmskb %xmm2, %eax94; SSE-NEXT: testb $1, %al95; SSE-NEXT: jne .LBB3_196; SSE-NEXT: # %bb.2: # %else97; SSE-NEXT: testb $2, %al98; SSE-NEXT: jne .LBB3_399; SSE-NEXT: .LBB3_4: # %else2100; SSE-NEXT: testb $4, %al101; SSE-NEXT: jne .LBB3_5102; SSE-NEXT: .LBB3_6: # %else4103; SSE-NEXT: testb $8, %al104; SSE-NEXT: jne .LBB3_7105; SSE-NEXT: .LBB3_8: # %else6106; SSE-NEXT: testb $16, %al107; SSE-NEXT: jne .LBB3_9108; SSE-NEXT: .LBB3_10: # %else8109; SSE-NEXT: testb $32, %al110; SSE-NEXT: jne .LBB3_11111; SSE-NEXT: .LBB3_12: # %else10112; SSE-NEXT: testb $64, %al113; SSE-NEXT: jne .LBB3_13114; SSE-NEXT: .LBB3_14: # %else12115; SSE-NEXT: testb $-128, %al116; SSE-NEXT: je .LBB3_16117; SSE-NEXT: .LBB3_15: # %cond.store13118; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]119; SSE-NEXT: movd %xmm0, 28120; SSE-NEXT: .LBB3_16: # %else14121; SSE-NEXT: xorl %eax, %eax122; SSE-NEXT: pxor %xmm0, %xmm0123; SSE-NEXT: pxor %xmm1, %xmm1124; SSE-NEXT: pxor %xmm2, %xmm2125; SSE-NEXT: retq126; SSE-NEXT: .LBB3_1: # %cond.store127; SSE-NEXT: movd %xmm0, 0128; SSE-NEXT: testb $2, %al129; SSE-NEXT: je .LBB3_4130; SSE-NEXT: .LBB3_3: # %cond.store1131; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]132; SSE-NEXT: movd %xmm2, 4133; SSE-NEXT: testb $4, %al134; SSE-NEXT: je .LBB3_6135; SSE-NEXT: .LBB3_5: # %cond.store3136; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]137; SSE-NEXT: movd %xmm2, 8138; SSE-NEXT: testb $8, %al139; SSE-NEXT: je .LBB3_8140; SSE-NEXT: .LBB3_7: # %cond.store5141; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]142; SSE-NEXT: movd %xmm0, 12143; SSE-NEXT: testb $16, %al144; SSE-NEXT: je .LBB3_10145; SSE-NEXT: .LBB3_9: # %cond.store7146; SSE-NEXT: movd %xmm1, 16147; SSE-NEXT: testb $32, %al148; SSE-NEXT: je .LBB3_12149; SSE-NEXT: .LBB3_11: # %cond.store9150; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]151; SSE-NEXT: movd %xmm0, 20152; SSE-NEXT: testb $64, %al153; SSE-NEXT: je .LBB3_14154; SSE-NEXT: .LBB3_13: # %cond.store11155; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]156; SSE-NEXT: movd %xmm0, 24157; SSE-NEXT: testb $-128, %al158; SSE-NEXT: jne .LBB3_15159; SSE-NEXT: jmp .LBB3_16160;161; AVX1-LABEL: zero_k:162; AVX1: # %bb.0:163; AVX1-NEXT: vpmovzxwd {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero164; AVX1-NEXT: vpslld $31, %xmm2, %xmm2165; AVX1-NEXT: vpunpckhwd {{.*#+}} xmm1 = xmm1[4,4,5,5,6,6,7,7]166; AVX1-NEXT: vpslld $31, %xmm1, %xmm1167; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1168; AVX1-NEXT: vmaskmovps %ymm0, %ymm1, 0169; AVX1-NEXT: vxorps %xmm0, %xmm0, %xmm0170; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm1171; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2172; AVX1-NEXT: vxorps %xmm0, %xmm0, %xmm0173; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm1174; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2175; AVX1-NEXT: vzeroupper176; AVX1-NEXT: retq177;178; AVX2-LABEL: zero_k:179; AVX2: # %bb.0:180; AVX2-NEXT: vpmovzxwd {{.*#+}} ymm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero181; AVX2-NEXT: vpslld $31, %ymm1, %ymm1182; AVX2-NEXT: vpmaskmovd %ymm0, %ymm1, 0183; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1184; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0185; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1186; AVX2-NEXT: vzeroupper187; AVX2-NEXT: retq188;189; AVX512VL-LABEL: zero_k:190; AVX512VL: # %bb.0:191; AVX512VL-NEXT: vpmovsxwd %xmm1, %ymm1192; AVX512VL-NEXT: vpslld $31, %ymm1, %ymm1193; AVX512VL-NEXT: vptestmd %ymm1, %ymm1, %k1194; AVX512VL-NEXT: vmovdqa32 %ymm0, 0 {%k1}195; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1196; AVX512VL-NEXT: vpxor %xmm0, %xmm0, %xmm0197; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1198; AVX512VL-NEXT: kxorw %k0, %k0, %k1199; AVX512VL-NEXT: vzeroupper200; AVX512VL-NEXT: retq201;202; AVX512BW-LABEL: zero_k:203; AVX512BW: # %bb.0:204; AVX512BW-NEXT: vpsllw $15, %xmm1, %xmm1205; AVX512BW-NEXT: vpmovw2m %xmm1, %k1206; AVX512BW-NEXT: vmovdqa32 %ymm0, 0 {%k1}207; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1208; AVX512BW-NEXT: vpxor %xmm0, %xmm0, %xmm0209; AVX512BW-NEXT: kxorq %k0, %k0, %k1210; AVX512BW-NEXT: vzeroupper211; AVX512BW-NEXT: retq212 tail call void @llvm.masked.store.v8i32.p0(<8 x i32> %arg, ptr null, i32 32, <8 x i1> %mask)213 ret void214}215 216attributes #0 = { "zero-call-used-regs"="used" }217