270 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefixes=SSE3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1OR25; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1OR26; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=AVX,AVX512F7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512VL8 9define void @widen_fdiv_v2f32_v4f32(ptr %a0, ptr %b0, ptr %c0) {10; SSE-LABEL: widen_fdiv_v2f32_v4f32:11; SSE: # %bb.0:12; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero13; SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero14; SSE-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero15; SSE-NEXT: divps %xmm2, %xmm016; SSE-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero17; SSE-NEXT: divps %xmm2, %xmm118; SSE-NEXT: movlps %xmm0, (%rdx)19; SSE-NEXT: movlps %xmm1, 8(%rdx)20; SSE-NEXT: retq21;22; AVX-LABEL: widen_fdiv_v2f32_v4f32:23; AVX: # %bb.0:24; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero25; AVX-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero26; AVX-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero27; AVX-NEXT: vdivps %xmm2, %xmm0, %xmm028; AVX-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero29; AVX-NEXT: vdivps %xmm2, %xmm1, %xmm130; AVX-NEXT: vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]31; AVX-NEXT: vmovups %xmm0, (%rdx)32; AVX-NEXT: retq33 %a2 = getelementptr inbounds i8, ptr %a0, i64 834 %b2 = getelementptr inbounds i8, ptr %b0, i64 835 %c2 = getelementptr inbounds i8, ptr %c0, i64 836 %va0 = load <2 x float>, ptr %a0, align 437 %vb0 = load <2 x float>, ptr %b0, align 438 %va2 = load <2 x float>, ptr %a2, align 439 %vb2 = load <2 x float>, ptr %b2, align 440 %vc0 = fdiv <2 x float> %va0, %vb041 %vc2 = fdiv <2 x float> %va2, %vb242 store <2 x float> %vc0, ptr %c0, align 443 store <2 x float> %vc2, ptr %c2, align 444 ret void45}46 47define void @widen_fdiv_v2f32_v8f32(ptr %a0, ptr %b0, ptr %c0) {48; SSE-LABEL: widen_fdiv_v2f32_v8f32:49; SSE: # %bb.0:50; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero51; SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero52; SSE-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero53; SSE-NEXT: movsd {{.*#+}} xmm3 = mem[0],zero54; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero55; SSE-NEXT: divps %xmm4, %xmm056; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero57; SSE-NEXT: divps %xmm4, %xmm158; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero59; SSE-NEXT: divps %xmm4, %xmm260; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero61; SSE-NEXT: divps %xmm4, %xmm362; SSE-NEXT: movlps %xmm0, (%rdx)63; SSE-NEXT: movlps %xmm1, 8(%rdx)64; SSE-NEXT: movlps %xmm2, 16(%rdx)65; SSE-NEXT: movlps %xmm3, 24(%rdx)66; SSE-NEXT: retq67;68; AVX-LABEL: widen_fdiv_v2f32_v8f32:69; AVX: # %bb.0:70; AVX-NEXT: vmovups (%rdi), %ymm071; AVX-NEXT: vdivps (%rsi), %ymm0, %ymm072; AVX-NEXT: vmovups %ymm0, (%rdx)73; AVX-NEXT: vzeroupper74; AVX-NEXT: retq75 %a2 = getelementptr inbounds i8, ptr %a0, i64 876 %b2 = getelementptr inbounds i8, ptr %b0, i64 877 %c2 = getelementptr inbounds i8, ptr %c0, i64 878 %a4 = getelementptr inbounds i8, ptr %a0, i64 1679 %b4 = getelementptr inbounds i8, ptr %b0, i64 1680 %c4 = getelementptr inbounds i8, ptr %c0, i64 1681 %a6 = getelementptr inbounds i8, ptr %a0, i64 2482 %b6 = getelementptr inbounds i8, ptr %b0, i64 2483 %c6 = getelementptr inbounds i8, ptr %c0, i64 2484 %va0 = load <2 x float>, ptr %a0, align 485 %vb0 = load <2 x float>, ptr %b0, align 486 %va2 = load <2 x float>, ptr %a2, align 487 %vb2 = load <2 x float>, ptr %b2, align 488 %va4 = load <2 x float>, ptr %a4, align 489 %vb4 = load <2 x float>, ptr %b4, align 490 %va6 = load <2 x float>, ptr %a6, align 491 %vb6 = load <2 x float>, ptr %b6, align 492 %vc0 = fdiv <2 x float> %va0, %vb093 %vc2 = fdiv <2 x float> %va2, %vb294 %vc4 = fdiv <2 x float> %va4, %vb495 %vc6 = fdiv <2 x float> %va6, %vb696 store <2 x float> %vc0, ptr %c0, align 497 store <2 x float> %vc2, ptr %c2, align 498 store <2 x float> %vc4, ptr %c4, align 499 store <2 x float> %vc6, ptr %c6, align 4100 ret void101}102 103define void @widen_fdiv_v2f32_v16f32(ptr %a0, ptr %b0, ptr %c0) {104; SSE-LABEL: widen_fdiv_v2f32_v16f32:105; SSE: # %bb.0:106; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero107; SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero108; SSE-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero109; SSE-NEXT: movsd {{.*#+}} xmm3 = mem[0],zero110; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero111; SSE-NEXT: divps %xmm4, %xmm0112; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero113; SSE-NEXT: divps %xmm4, %xmm1114; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero115; SSE-NEXT: divps %xmm4, %xmm2116; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero117; SSE-NEXT: divps %xmm4, %xmm3118; SSE-NEXT: movsd {{.*#+}} xmm4 = mem[0],zero119; SSE-NEXT: movsd {{.*#+}} xmm5 = mem[0],zero120; SSE-NEXT: divps %xmm5, %xmm4121; SSE-NEXT: movsd {{.*#+}} xmm5 = mem[0],zero122; SSE-NEXT: movsd {{.*#+}} xmm6 = mem[0],zero123; SSE-NEXT: divps %xmm6, %xmm5124; SSE-NEXT: movsd {{.*#+}} xmm6 = mem[0],zero125; SSE-NEXT: movsd {{.*#+}} xmm7 = mem[0],zero126; SSE-NEXT: divps %xmm7, %xmm6127; SSE-NEXT: movsd {{.*#+}} xmm7 = mem[0],zero128; SSE-NEXT: movsd {{.*#+}} xmm8 = mem[0],zero129; SSE-NEXT: divps %xmm8, %xmm7130; SSE-NEXT: movlps %xmm0, (%rdx)131; SSE-NEXT: movlps %xmm1, 8(%rdx)132; SSE-NEXT: movlps %xmm2, 16(%rdx)133; SSE-NEXT: movlps %xmm3, 24(%rdx)134; SSE-NEXT: movlps %xmm4, 32(%rdx)135; SSE-NEXT: movlps %xmm5, 40(%rdx)136; SSE-NEXT: movlps %xmm6, 48(%rdx)137; SSE-NEXT: movlps %xmm7, 56(%rdx)138; SSE-NEXT: retq139;140; AVX1OR2-LABEL: widen_fdiv_v2f32_v16f32:141; AVX1OR2: # %bb.0:142; AVX1OR2-NEXT: vmovups (%rdi), %ymm0143; AVX1OR2-NEXT: vmovups 32(%rdi), %ymm1144; AVX1OR2-NEXT: vdivps (%rsi), %ymm0, %ymm0145; AVX1OR2-NEXT: vdivps 32(%rsi), %ymm1, %ymm1146; AVX1OR2-NEXT: vmovups %ymm0, (%rdx)147; AVX1OR2-NEXT: vmovups %ymm1, 32(%rdx)148; AVX1OR2-NEXT: vzeroupper149; AVX1OR2-NEXT: retq150;151; AVX512F-LABEL: widen_fdiv_v2f32_v16f32:152; AVX512F: # %bb.0:153; AVX512F-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero154; AVX512F-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero155; AVX512F-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero156; AVX512F-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero157; AVX512F-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero158; AVX512F-NEXT: vdivps %xmm4, %xmm0, %xmm0159; AVX512F-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero160; AVX512F-NEXT: vdivps %xmm4, %xmm1, %xmm1161; AVX512F-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero162; AVX512F-NEXT: vdivps %xmm4, %xmm2, %xmm2163; AVX512F-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero164; AVX512F-NEXT: vdivps %xmm4, %xmm3, %xmm3165; AVX512F-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero166; AVX512F-NEXT: vmovsd {{.*#+}} xmm5 = mem[0],zero167; AVX512F-NEXT: vdivps %xmm5, %xmm4, %xmm4168; AVX512F-NEXT: vmovsd {{.*#+}} xmm5 = mem[0],zero169; AVX512F-NEXT: vmovsd {{.*#+}} xmm6 = mem[0],zero170; AVX512F-NEXT: vdivps %xmm6, %xmm5, %xmm5171; AVX512F-NEXT: vmovsd {{.*#+}} xmm6 = mem[0],zero172; AVX512F-NEXT: vmovsd {{.*#+}} xmm7 = mem[0],zero173; AVX512F-NEXT: vdivps %xmm7, %xmm6, %xmm6174; AVX512F-NEXT: vmovsd {{.*#+}} xmm7 = mem[0],zero175; AVX512F-NEXT: vmovsd {{.*#+}} xmm8 = mem[0],zero176; AVX512F-NEXT: vdivps %xmm8, %xmm7, %xmm7177; AVX512F-NEXT: vinsertf32x4 $1, %xmm7, %zmm6, %zmm6178; AVX512F-NEXT: vinsertf32x4 $1, %xmm5, %zmm4, %zmm4179; AVX512F-NEXT: vbroadcasti64x4 {{.*#+}} zmm5 = [0,2,8,10,0,2,8,10]180; AVX512F-NEXT: # zmm5 = mem[0,1,2,3,0,1,2,3]181; AVX512F-NEXT: vpermt2pd %zmm6, %zmm5, %zmm4182; AVX512F-NEXT: vinsertf32x4 $1, %xmm3, %zmm2, %zmm2183; AVX512F-NEXT: vinsertf32x4 $1, %xmm1, %zmm0, %zmm0184; AVX512F-NEXT: vpermt2pd %zmm2, %zmm5, %zmm0185; AVX512F-NEXT: vshuff64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]186; AVX512F-NEXT: vmovupd %zmm0, (%rdx)187; AVX512F-NEXT: vzeroupper188; AVX512F-NEXT: retq189;190; AVX512VL-LABEL: widen_fdiv_v2f32_v16f32:191; AVX512VL: # %bb.0:192; AVX512VL-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero193; AVX512VL-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero194; AVX512VL-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero195; AVX512VL-NEXT: vmovsd {{.*#+}} xmm3 = mem[0],zero196; AVX512VL-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero197; AVX512VL-NEXT: vdivps %xmm4, %xmm0, %xmm0198; AVX512VL-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero199; AVX512VL-NEXT: vdivps %xmm4, %xmm1, %xmm1200; AVX512VL-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero201; AVX512VL-NEXT: vdivps %xmm4, %xmm2, %xmm2202; AVX512VL-NEXT: vmovsd {{.*#+}} xmm4 = mem[0],zero203; AVX512VL-NEXT: vdivps %xmm4, %xmm3, %xmm3204; AVX512VL-NEXT: vinsertf32x4 $1, %xmm3, %zmm2, %zmm2205; AVX512VL-NEXT: vinsertf32x4 $1, %xmm1, %zmm0, %zmm0206; AVX512VL-NEXT: vbroadcasti64x4 {{.*#+}} zmm1 = [0,2,8,10,0,2,8,10]207; AVX512VL-NEXT: # zmm1 = mem[0,1,2,3,0,1,2,3]208; AVX512VL-NEXT: vpermi2pd %zmm2, %zmm0, %zmm1209; AVX512VL-NEXT: vmovups (%rdi), %ymm0210; AVX512VL-NEXT: vdivps (%rsi), %ymm0, %ymm0211; AVX512VL-NEXT: vinsertf64x4 $0, %ymm0, %zmm1, %zmm0212; AVX512VL-NEXT: vmovupd %zmm0, (%rdx)213; AVX512VL-NEXT: vzeroupper214; AVX512VL-NEXT: retq215 %a2 = getelementptr inbounds i8, ptr %a0, i64 8216 %b2 = getelementptr inbounds i8, ptr %b0, i64 8217 %c2 = getelementptr inbounds i8, ptr %c0, i64 8218 %a4 = getelementptr inbounds i8, ptr %a0, i64 16219 %b4 = getelementptr inbounds i8, ptr %b0, i64 16220 %c4 = getelementptr inbounds i8, ptr %c0, i64 16221 %a6 = getelementptr inbounds i8, ptr %a0, i64 24222 %b6 = getelementptr inbounds i8, ptr %b0, i64 24223 %c6 = getelementptr inbounds i8, ptr %c0, i64 24224 %a8 = getelementptr inbounds i8, ptr %a0, i64 32225 %b8 = getelementptr inbounds i8, ptr %b0, i64 32226 %c8 = getelementptr inbounds i8, ptr %c0, i64 32227 %a10 = getelementptr inbounds i8, ptr %a0, i64 40228 %b10 = getelementptr inbounds i8, ptr %b0, i64 40229 %c10 = getelementptr inbounds i8, ptr %c0, i64 40230 %a12 = getelementptr inbounds i8, ptr %a0, i64 48231 %b12 = getelementptr inbounds i8, ptr %b0, i64 48232 %c12 = getelementptr inbounds i8, ptr %c0, i64 48233 %a14 = getelementptr inbounds i8, ptr %a0, i64 56234 %b14 = getelementptr inbounds i8, ptr %b0, i64 56235 %c14 = getelementptr inbounds i8, ptr %c0, i64 56236 %va0 = load <2 x float>, ptr %a0, align 4237 %vb0 = load <2 x float>, ptr %b0, align 4238 %va2 = load <2 x float>, ptr %a2, align 4239 %vb2 = load <2 x float>, ptr %b2, align 4240 %va4 = load <2 x float>, ptr %a4, align 4241 %vb4 = load <2 x float>, ptr %b4, align 4242 %va6 = load <2 x float>, ptr %a6, align 4243 %vb6 = load <2 x float>, ptr %b6, align 4244 %va8 = load <2 x float>, ptr %a8, align 4245 %vb8 = load <2 x float>, ptr %b8, align 4246 %va10 = load <2 x float>, ptr %a10, align 4247 %vb10 = load <2 x float>, ptr %b10, align 4248 %va12 = load <2 x float>, ptr %a12, align 4249 %vb12 = load <2 x float>, ptr %b12, align 4250 %va14 = load <2 x float>, ptr %a14, align 4251 %vb14 = load <2 x float>, ptr %b14, align 4252 %vc0 = fdiv <2 x float> %va0, %vb0253 %vc2 = fdiv <2 x float> %va2, %vb2254 %vc4 = fdiv <2 x float> %va4, %vb4255 %vc6 = fdiv <2 x float> %va6, %vb6256 %vc8 = fdiv <2 x float> %va8, %vb8257 %vc10 = fdiv <2 x float> %va10, %vb10258 %vc12 = fdiv <2 x float> %va12, %vb12259 %vc14 = fdiv <2 x float> %va14, %vb14260 store <2 x float> %vc0, ptr %c0, align 4261 store <2 x float> %vc2, ptr %c2, align 4262 store <2 x float> %vc4, ptr %c4, align 4263 store <2 x float> %vc6, ptr %c6, align 4264 store <2 x float> %vc8, ptr %c8, align 4265 store <2 x float> %vc10, ptr %c10, align 4266 store <2 x float> %vc12, ptr %c12, align 4267 store <2 x float> %vc14, ptr %c14, align 4268 ret void269}270