1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=CHECK --check-prefix=SSE --check-prefix=SSE-RECIP 3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=AVX-RECIP 4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx,+fma | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=FMA-RECIP 5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=btver2 -print-schedule | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=BTVER2 6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=sandybridge -print-schedule | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=SANDY 7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=haswell -print-schedule | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=HASWELL 8; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=haswell -mattr=-fma | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=HASWELL-NO-FMA 9; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=knl -print-schedule | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=AVX512 --check-prefix=KNL 10; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=skx -print-schedule | FileCheck %s --check-prefix=CHECK --check-prefix=AVX --check-prefix=AVX512 --check-prefix=SKX 11 12; If the target's divss/divps instructions are substantially 13; slower than rcpss/rcpps with a Newton-Raphson refinement, 14; we should generate the estimate sequence. 15 16; See PR21385 ( http://llvm.org/bugs/show_bug.cgi?id=21385 ) 17; for details about the accuracy, speed, and implementation 18; differences of x86 reciprocal estimates. 19 20define float @f32_no_estimate(float %x) #0 { 21; SSE-LABEL: f32_no_estimate: 22; SSE: # %bb.0: 23; SSE-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero 24; SSE-NEXT: divss %xmm0, %xmm1 25; SSE-NEXT: movaps %xmm1, %xmm0 26; SSE-NEXT: retq 27; 28; AVX-RECIP-LABEL: f32_no_estimate: 29; AVX-RECIP: # %bb.0: 30; AVX-RECIP-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero 31; AVX-RECIP-NEXT: vdivss %xmm0, %xmm1, %xmm0 32; AVX-RECIP-NEXT: retq 33; 34; FMA-RECIP-LABEL: f32_no_estimate: 35; FMA-RECIP: # %bb.0: 36; FMA-RECIP-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero 37; FMA-RECIP-NEXT: vdivss %xmm0, %xmm1, %xmm0 38; FMA-RECIP-NEXT: retq 39; 40; BTVER2-LABEL: f32_no_estimate: 41; BTVER2: # %bb.0: 42; BTVER2-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:1.00] 43; BTVER2-NEXT: vdivss %xmm0, %xmm1, %xmm0 # sched: [19:19.00] 44; BTVER2-NEXT: retq # sched: [4:1.00] 45; 46; SANDY-LABEL: f32_no_estimate: 47; SANDY: # %bb.0: 48; SANDY-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [6:0.50] 49; SANDY-NEXT: vdivss %xmm0, %xmm1, %xmm0 # sched: [14:14.00] 50; SANDY-NEXT: retq # sched: [1:1.00] 51; 52; HASWELL-LABEL: f32_no_estimate: 53; HASWELL: # %bb.0: 54; HASWELL-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] 55; HASWELL-NEXT: vdivss %xmm0, %xmm1, %xmm0 # sched: [13:7.00] 56; HASWELL-NEXT: retq # sched: [7:1.00] 57; 58; HASWELL-NO-FMA-LABEL: f32_no_estimate: 59; HASWELL-NO-FMA: # %bb.0: 60; HASWELL-NO-FMA-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero 61; HASWELL-NO-FMA-NEXT: vdivss %xmm0, %xmm1, %xmm0 62; HASWELL-NO-FMA-NEXT: retq 63; 64; KNL-LABEL: f32_no_estimate: 65; KNL: # %bb.0: 66; KNL-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] 67; KNL-NEXT: vdivss %xmm0, %xmm1, %xmm0 # sched: [13:7.00] 68; KNL-NEXT: retq # sched: [7:1.00] 69; 70; SKX-LABEL: f32_no_estimate: 71; SKX: # %bb.0: 72; SKX-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero sched: [5:0.50] 73; SKX-NEXT: vdivss %xmm0, %xmm1, %xmm0 # sched: [11:3.00] 74; SKX-NEXT: retq # sched: [7:1.00] 75 %div = fdiv fast float 1.0, %x 76 ret float %div 77} 78 79define float @f32_one_step(float %x) #1 { 80; SSE-LABEL: f32_one_step: 81; SSE: # %bb.0: 82; SSE-NEXT: rcpss %xmm0, %xmm2 83; SSE-NEXT: mulss %xmm2, %xmm0 84; SSE-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero 85; SSE-NEXT: subss %xmm0, %xmm1 86; SSE-NEXT: mulss %xmm2, %xmm1 87; SSE-NEXT: addss %xmm2, %xmm1 88; SSE-NEXT: movaps %xmm1, %xmm0 89; SSE-NEXT: retq 90; 91; AVX-RECIP-LABEL: f32_one_step: 92; AVX-RECIP: # %bb.0: 93; AVX-RECIP-NEXT: vrcpss %xmm0, %xmm0, %xmm1 94; AVX-RECIP-NEXT: vmulss %xmm1, %xmm0, %xmm0 95; AVX-RECIP-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero 96; AVX-RECIP-NEXT: vsubss %xmm0, %xmm2, %xmm0 97; AVX-RECIP-NEXT: vmulss %xmm0, %xmm1, %xmm0 98; AVX-RECIP-NEXT: vaddss %xmm0, %xmm1, %xmm0 99; AVX-RECIP-NEXT: retq 100; 101; FMA-RECIP-LABEL: f32_one_step: 102; FMA-RECIP: # %bb.0: 103; FMA-RECIP-NEXT: vrcpss %xmm0, %xmm0, %xmm1 104; FMA-RECIP-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm1 * xmm0) + mem 105; FMA-RECIP-NEXT: vfmadd132ss {{.*#+}} xmm0 = (xmm0 * xmm1) + xmm1 106; FMA-RECIP-NEXT: retq 107; 108; BTVER2-LABEL: f32_one_step: 109; BTVER2: # %bb.0: 110; BTVER2-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [5:1.00] 111; BTVER2-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [2:1.00] 112; BTVER2-NEXT: vmulss %xmm1, %xmm0, %xmm0 # sched: [2:1.00] 113; BTVER2-NEXT: vsubss %xmm0, %xmm2, %xmm0 # sched: [3:1.00] 114; BTVER2-NEXT: vmulss %xmm0, %xmm1, %xmm0 # sched: [2:1.00] 115; BTVER2-NEXT: vaddss %xmm0, %xmm1, %xmm0 # sched: [3:1.00] 116; BTVER2-NEXT: retq # sched: [4:1.00] 117; 118; SANDY-LABEL: f32_one_step: 119; SANDY: # %bb.0: 120; SANDY-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [5:1.00] 121; SANDY-NEXT: vmulss %xmm1, %xmm0, %xmm0 # sched: [5:1.00] 122; SANDY-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [6:0.50] 123; SANDY-NEXT: vsubss %xmm0, %xmm2, %xmm0 # sched: [3:1.00] 124; SANDY-NEXT: vmulss %xmm0, %xmm1, %xmm0 # sched: [5:1.00] 125; SANDY-NEXT: vaddss %xmm0, %xmm1, %xmm0 # sched: [3:1.00] 126; SANDY-NEXT: retq # sched: [1:1.00] 127; 128; HASWELL-LABEL: f32_one_step: 129; HASWELL: # %bb.0: 130; HASWELL-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [5:1.00] 131; HASWELL-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm1 * xmm0) + mem sched: [10:0.50] 132; HASWELL-NEXT: vfmadd132ss {{.*#+}} xmm0 = (xmm0 * xmm1) + xmm1 sched: [5:0.50] 133; HASWELL-NEXT: retq # sched: [7:1.00] 134; 135; HASWELL-NO-FMA-LABEL: f32_one_step: 136; HASWELL-NO-FMA: # %bb.0: 137; HASWELL-NO-FMA-NEXT: vrcpss %xmm0, %xmm0, %xmm1 138; HASWELL-NO-FMA-NEXT: vmulss %xmm1, %xmm0, %xmm0 139; HASWELL-NO-FMA-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero 140; HASWELL-NO-FMA-NEXT: vsubss %xmm0, %xmm2, %xmm0 141; HASWELL-NO-FMA-NEXT: vmulss %xmm0, %xmm1, %xmm0 142; HASWELL-NO-FMA-NEXT: vaddss %xmm0, %xmm1, %xmm0 143; HASWELL-NO-FMA-NEXT: retq 144; 145; KNL-LABEL: f32_one_step: 146; KNL: # %bb.0: 147; KNL-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [5:1.00] 148; KNL-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm1 * xmm0) + mem sched: [10:0.50] 149; KNL-NEXT: vfmadd132ss {{.*#+}} xmm0 = (xmm0 * xmm1) + xmm1 sched: [5:0.50] 150; KNL-NEXT: retq # sched: [7:1.00] 151; 152; SKX-LABEL: f32_one_step: 153; SKX: # %bb.0: 154; SKX-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [4:1.00] 155; SKX-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm1 * xmm0) + mem sched: [9:0.50] 156; SKX-NEXT: vfmadd132ss {{.*#+}} xmm0 = (xmm0 * xmm1) + xmm1 sched: [4:0.50] 157; SKX-NEXT: retq # sched: [7:1.00] 158 %div = fdiv fast float 1.0, %x 159 ret float %div 160} 161 162define float @f32_two_step(float %x) #2 { 163; SSE-LABEL: f32_two_step: 164; SSE: # %bb.0: 165; SSE-NEXT: rcpss %xmm0, %xmm2 166; SSE-NEXT: movaps %xmm0, %xmm3 167; SSE-NEXT: mulss %xmm2, %xmm3 168; SSE-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero 169; SSE-NEXT: movaps %xmm1, %xmm4 170; SSE-NEXT: subss %xmm3, %xmm4 171; SSE-NEXT: mulss %xmm2, %xmm4 172; SSE-NEXT: addss %xmm2, %xmm4 173; SSE-NEXT: mulss %xmm4, %xmm0 174; SSE-NEXT: subss %xmm0, %xmm1 175; SSE-NEXT: mulss %xmm4, %xmm1 176; SSE-NEXT: addss %xmm4, %xmm1 177; SSE-NEXT: movaps %xmm1, %xmm0 178; SSE-NEXT: retq 179; 180; AVX-RECIP-LABEL: f32_two_step: 181; AVX-RECIP: # %bb.0: 182; AVX-RECIP-NEXT: vrcpss %xmm0, %xmm0, %xmm1 183; AVX-RECIP-NEXT: vmulss %xmm1, %xmm0, %xmm2 184; AVX-RECIP-NEXT: vmovss {{.*#+}} xmm3 = mem[0],zero,zero,zero 185; AVX-RECIP-NEXT: vsubss %xmm2, %xmm3, %xmm2 186; AVX-RECIP-NEXT: vmulss %xmm2, %xmm1, %xmm2 187; AVX-RECIP-NEXT: vaddss %xmm2, %xmm1, %xmm1 188; AVX-RECIP-NEXT: vmulss %xmm1, %xmm0, %xmm0 189; AVX-RECIP-NEXT: vsubss %xmm0, %xmm3, %xmm0 190; AVX-RECIP-NEXT: vmulss %xmm0, %xmm1, %xmm0 191; AVX-RECIP-NEXT: vaddss %xmm0, %xmm1, %xmm0 192; AVX-RECIP-NEXT: retq 193; 194; FMA-RECIP-LABEL: f32_two_step: 195; FMA-RECIP: # %bb.0: 196; FMA-RECIP-NEXT: vrcpss %xmm0, %xmm0, %xmm1 197; FMA-RECIP-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero 198; FMA-RECIP-NEXT: vmovaps %xmm1, %xmm3 199; FMA-RECIP-NEXT: vfnmadd213ss {{.*#+}} xmm3 = -(xmm0 * xmm3) + xmm2 200; FMA-RECIP-NEXT: vfmadd132ss {{.*#+}} xmm3 = (xmm3 * xmm1) + xmm1 201; FMA-RECIP-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm3 * xmm0) + xmm2 202; FMA-RECIP-NEXT: vfmadd132ss {{.*#+}} xmm0 = (xmm0 * xmm3) + xmm3 203; FMA-RECIP-NEXT: retq 204; 205; BTVER2-LABEL: f32_two_step: 206; BTVER2: # %bb.0: 207; BTVER2-NEXT: vmovss {{.*#+}} xmm3 = mem[0],zero,zero,zero sched: [5:1.00] 208; BTVER2-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [2:1.00] 209; BTVER2-NEXT: vmulss %xmm1, %xmm0, %xmm2 # sched: [2:1.00] 210; BTVER2-NEXT: vsubss %xmm2, %xmm3, %xmm2 # sched: [3:1.00] 211; BTVER2-NEXT: vmulss %xmm2, %xmm1, %xmm2 # sched: [2:1.00] 212; BTVER2-NEXT: vaddss %xmm2, %xmm1, %xmm1 # sched: [3:1.00] 213; BTVER2-NEXT: vmulss %xmm1, %xmm0, %xmm0 # sched: [2:1.00] 214; BTVER2-NEXT: vsubss %xmm0, %xmm3, %xmm0 # sched: [3:1.00] 215; BTVER2-NEXT: vmulss %xmm0, %xmm1, %xmm0 # sched: [2:1.00] 216; BTVER2-NEXT: vaddss %xmm0, %xmm1, %xmm0 # sched: [3:1.00] 217; BTVER2-NEXT: retq # sched: [4:1.00] 218; 219; SANDY-LABEL: f32_two_step: 220; SANDY: # %bb.0: 221; SANDY-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [5:1.00] 222; SANDY-NEXT: vmulss %xmm1, %xmm0, %xmm2 # sched: [5:1.00] 223; SANDY-NEXT: vmovss {{.*#+}} xmm3 = mem[0],zero,zero,zero sched: [6:0.50] 224; SANDY-NEXT: vsubss %xmm2, %xmm3, %xmm2 # sched: [3:1.00] 225; SANDY-NEXT: vmulss %xmm2, %xmm1, %xmm2 # sched: [5:1.00] 226; SANDY-NEXT: vaddss %xmm2, %xmm1, %xmm1 # sched: [3:1.00] 227; SANDY-NEXT: vmulss %xmm1, %xmm0, %xmm0 # sched: [5:1.00] 228; SANDY-NEXT: vsubss %xmm0, %xmm3, %xmm0 # sched: [3:1.00] 229; SANDY-NEXT: vmulss %xmm0, %xmm1, %xmm0 # sched: [5:1.00] 230; SANDY-NEXT: vaddss %xmm0, %xmm1, %xmm0 # sched: [3:1.00] 231; SANDY-NEXT: retq # sched: [1:1.00] 232; 233; HASWELL-LABEL: f32_two_step: 234; HASWELL: # %bb.0: 235; HASWELL-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [5:1.00] 236; HASWELL-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [5:0.50] 237; HASWELL-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] 238; HASWELL-NEXT: vfnmadd213ss {{.*#+}} xmm3 = -(xmm0 * xmm3) + xmm2 sched: [5:0.50] 239; HASWELL-NEXT: vfmadd132ss {{.*#+}} xmm3 = (xmm3 * xmm1) + xmm1 sched: [5:0.50] 240; HASWELL-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm3 * xmm0) + xmm2 sched: [5:0.50] 241; HASWELL-NEXT: vfmadd132ss {{.*#+}} xmm0 = (xmm0 * xmm3) + xmm3 sched: [5:0.50] 242; HASWELL-NEXT: retq # sched: [7:1.00] 243; 244; HASWELL-NO-FMA-LABEL: f32_two_step: 245; HASWELL-NO-FMA: # %bb.0: 246; HASWELL-NO-FMA-NEXT: vrcpss %xmm0, %xmm0, %xmm1 247; HASWELL-NO-FMA-NEXT: vmulss %xmm1, %xmm0, %xmm2 248; HASWELL-NO-FMA-NEXT: vmovss {{.*#+}} xmm3 = mem[0],zero,zero,zero 249; HASWELL-NO-FMA-NEXT: vsubss %xmm2, %xmm3, %xmm2 250; HASWELL-NO-FMA-NEXT: vmulss %xmm2, %xmm1, %xmm2 251; HASWELL-NO-FMA-NEXT: vaddss %xmm2, %xmm1, %xmm1 252; HASWELL-NO-FMA-NEXT: vmulss %xmm1, %xmm0, %xmm0 253; HASWELL-NO-FMA-NEXT: vsubss %xmm0, %xmm3, %xmm0 254; HASWELL-NO-FMA-NEXT: vmulss %xmm0, %xmm1, %xmm0 255; HASWELL-NO-FMA-NEXT: vaddss %xmm0, %xmm1, %xmm0 256; HASWELL-NO-FMA-NEXT: retq 257; 258; KNL-LABEL: f32_two_step: 259; KNL: # %bb.0: 260; KNL-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [5:1.00] 261; KNL-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [5:0.50] 262; KNL-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] 263; KNL-NEXT: vfnmadd213ss {{.*#+}} xmm3 = -(xmm0 * xmm3) + xmm2 sched: [5:0.50] 264; KNL-NEXT: vfmadd132ss {{.*#+}} xmm3 = (xmm3 * xmm1) + xmm1 sched: [5:0.50] 265; KNL-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm3 * xmm0) + xmm2 sched: [5:0.50] 266; KNL-NEXT: vfmadd132ss {{.*#+}} xmm0 = (xmm0 * xmm3) + xmm3 sched: [5:0.50] 267; KNL-NEXT: retq # sched: [7:1.00] 268; 269; SKX-LABEL: f32_two_step: 270; SKX: # %bb.0: 271; SKX-NEXT: vrcpss %xmm0, %xmm0, %xmm1 # sched: [4:1.00] 272; SKX-NEXT: vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero sched: [5:0.50] 273; SKX-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:0.33] 274; SKX-NEXT: vfnmadd213ss {{.*#+}} xmm3 = -(xmm0 * xmm3) + xmm2 sched: [4:0.50] 275; SKX-NEXT: vfmadd132ss {{.*#+}} xmm3 = (xmm3 * xmm1) + xmm1 sched: [4:0.50] 276; SKX-NEXT: vfnmadd213ss {{.*#+}} xmm0 = -(xmm3 * xmm0) + xmm2 sched: [4:0.50] 277; SKX-NEXT: vfmadd132ss {{.*#+}} xmm0 = (xmm0 * xmm3) + xmm3 sched: [4:0.50] 278; SKX-NEXT: retq # sched: [7:1.00] 279 %div = fdiv fast float 1.0, %x 280 ret float %div 281} 282 283define <4 x float> @v4f32_no_estimate(<4 x float> %x) #0 { 284; SSE-LABEL: v4f32_no_estimate: 285; SSE: # %bb.0: 286; SSE-NEXT: movaps {{.*#+}} xmm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 287; SSE-NEXT: divps %xmm0, %xmm1 288; SSE-NEXT: movaps %xmm1, %xmm0 289; SSE-NEXT: retq 290; 291; AVX-RECIP-LABEL: v4f32_no_estimate: 292; AVX-RECIP: # %bb.0: 293; AVX-RECIP-NEXT: vmovaps {{.*#+}} xmm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 294; AVX-RECIP-NEXT: vdivps %xmm0, %xmm1, %xmm0 295; AVX-RECIP-NEXT: retq 296; 297; FMA-RECIP-LABEL: v4f32_no_estimate: 298; FMA-RECIP: # %bb.0: 299; FMA-RECIP-NEXT: vmovaps {{.*#+}} xmm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 300; FMA-RECIP-NEXT: vdivps %xmm0, %xmm1, %xmm0 301; FMA-RECIP-NEXT: retq 302; 303; BTVER2-LABEL: v4f32_no_estimate: 304; BTVER2: # %bb.0: 305; BTVER2-NEXT: vmovaps {{.*#+}} xmm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [5:1.00] 306; BTVER2-NEXT: vdivps %xmm0, %xmm1, %xmm0 # sched: [19:19.00] 307; BTVER2-NEXT: retq # sched: [4:1.00] 308; 309; SANDY-LABEL: v4f32_no_estimate: 310; SANDY: # %bb.0: 311; SANDY-NEXT: vmovaps {{.*#+}} xmm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [6:0.50] 312; SANDY-NEXT: vdivps %xmm0, %xmm1, %xmm0 # sched: [14:14.00] 313; SANDY-NEXT: retq # sched: [1:1.00] 314; 315; HASWELL-LABEL: v4f32_no_estimate: 316; HASWELL: # %bb.0: 317; HASWELL-NEXT: vbroadcastss {{.*#+}} xmm1 = [1,1,1,1] sched: [6:0.50] 318; HASWELL-NEXT: vdivps %xmm0, %xmm1, %xmm0 # sched: [13:7.00] 319; HASWELL-NEXT: retq # sched: [7:1.00] 320; 321; HASWELL-NO-FMA-LABEL: v4f32_no_estimate: 322; HASWELL-NO-FMA: # %bb.0: 323; HASWELL-NO-FMA-NEXT: vbroadcastss {{.*#+}} xmm1 = [1,1,1,1] 324; HASWELL-NO-FMA-NEXT: vdivps %xmm0, %xmm1, %xmm0 325; HASWELL-NO-FMA-NEXT: retq 326; 327; KNL-LABEL: v4f32_no_estimate: 328; KNL: # %bb.0: 329; KNL-NEXT: vbroadcastss {{.*#+}} xmm1 = [1,1,1,1] sched: [6:0.50] 330; KNL-NEXT: vdivps %xmm0, %xmm1, %xmm0 # sched: [13:7.00] 331; KNL-NEXT: retq # sched: [7:1.00] 332; 333; SKX-LABEL: v4f32_no_estimate: 334; SKX: # %bb.0: 335; SKX-NEXT: vbroadcastss {{.*#+}} xmm1 = [1,1,1,1] sched: [6:0.50] 336; SKX-NEXT: vdivps %xmm0, %xmm1, %xmm0 # sched: [11:3.00] 337; SKX-NEXT: retq # sched: [7:1.00] 338 %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x 339 ret <4 x float> %div 340} 341 342define <4 x float> @v4f32_one_step(<4 x float> %x) #1 { 343; SSE-LABEL: v4f32_one_step: 344; SSE: # %bb.0: 345; SSE-NEXT: rcpps %xmm0, %xmm2 346; SSE-NEXT: mulps %xmm2, %xmm0 347; SSE-NEXT: movaps {{.*#+}} xmm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 348; SSE-NEXT: subps %xmm0, %xmm1 349; SSE-NEXT: mulps %xmm2, %xmm1 350; SSE-NEXT: addps %xmm2, %xmm1 351; SSE-NEXT: movaps %xmm1, %xmm0 352; SSE-NEXT: retq 353; 354; AVX-RECIP-LABEL: v4f32_one_step: 355; AVX-RECIP: # %bb.0: 356; AVX-RECIP-NEXT: vrcpps %xmm0, %xmm1 357; AVX-RECIP-NEXT: vmulps %xmm1, %xmm0, %xmm0 358; AVX-RECIP-NEXT: vmovaps {{.*#+}} xmm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 359; AVX-RECIP-NEXT: vsubps %xmm0, %xmm2, %xmm0 360; AVX-RECIP-NEXT: vmulps %xmm0, %xmm1, %xmm0 361; AVX-RECIP-NEXT: vaddps %xmm0, %xmm1, %xmm0 362; AVX-RECIP-NEXT: retq 363; 364; FMA-RECIP-LABEL: v4f32_one_step: 365; FMA-RECIP: # %bb.0: 366; FMA-RECIP-NEXT: vrcpps %xmm0, %xmm1 367; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} xmm0 = -(xmm1 * xmm0) + mem 368; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} xmm0 = (xmm0 * xmm1) + xmm1 369; FMA-RECIP-NEXT: retq 370; 371; BTVER2-LABEL: v4f32_one_step: 372; BTVER2: # %bb.0: 373; BTVER2-NEXT: vmovaps {{.*#+}} xmm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [5:1.00] 374; BTVER2-NEXT: vrcpps %xmm0, %xmm1 # sched: [2:1.00] 375; BTVER2-NEXT: vmulps %xmm1, %xmm0, %xmm0 # sched: [2:1.00] 376; BTVER2-NEXT: vsubps %xmm0, %xmm2, %xmm0 # sched: [3:1.00] 377; BTVER2-NEXT: vmulps %xmm0, %xmm1, %xmm0 # sched: [2:1.00] 378; BTVER2-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00] 379; BTVER2-NEXT: retq # sched: [4:1.00] 380; 381; SANDY-LABEL: v4f32_one_step: 382; SANDY: # %bb.0: 383; SANDY-NEXT: vrcpps %xmm0, %xmm1 # sched: [5:1.00] 384; SANDY-NEXT: vmulps %xmm1, %xmm0, %xmm0 # sched: [5:1.00] 385; SANDY-NEXT: vmovaps {{.*#+}} xmm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [6:0.50] 386; SANDY-NEXT: vsubps %xmm0, %xmm2, %xmm0 # sched: [3:1.00] 387; SANDY-NEXT: vmulps %xmm0, %xmm1, %xmm0 # sched: [5:1.00] 388; SANDY-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00] 389; SANDY-NEXT: retq # sched: [1:1.00] 390; 391; HASWELL-LABEL: v4f32_one_step: 392; HASWELL: # %bb.0: 393; HASWELL-NEXT: vrcpps %xmm0, %xmm1 # sched: [5:1.00] 394; HASWELL-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50] 395; HASWELL-NEXT: vfnmadd213ps {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm2 sched: [5:0.50] 396; HASWELL-NEXT: vfmadd132ps {{.*#+}} xmm0 = (xmm0 * xmm1) + xmm1 sched: [5:0.50] 397; HASWELL-NEXT: retq # sched: [7:1.00] 398; 399; HASWELL-NO-FMA-LABEL: v4f32_one_step: 400; HASWELL-NO-FMA: # %bb.0: 401; HASWELL-NO-FMA-NEXT: vrcpps %xmm0, %xmm1 402; HASWELL-NO-FMA-NEXT: vmulps %xmm1, %xmm0, %xmm0 403; HASWELL-NO-FMA-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] 404; HASWELL-NO-FMA-NEXT: vsubps %xmm0, %xmm2, %xmm0 405; HASWELL-NO-FMA-NEXT: vmulps %xmm0, %xmm1, %xmm0 406; HASWELL-NO-FMA-NEXT: vaddps %xmm0, %xmm1, %xmm0 407; HASWELL-NO-FMA-NEXT: retq 408; 409; KNL-LABEL: v4f32_one_step: 410; KNL: # %bb.0: 411; KNL-NEXT: vrcpps %xmm0, %xmm1 # sched: [5:1.00] 412; KNL-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50] 413; KNL-NEXT: vfnmadd213ps {{.*#+}} xmm0 = -(xmm1 * xmm0) + xmm2 sched: [5:0.50] 414; KNL-NEXT: vfmadd132ps {{.*#+}} xmm0 = (xmm0 * xmm1) + xmm1 sched: [5:0.50] 415; KNL-NEXT: retq # sched: [7:1.00] 416; 417; SKX-LABEL: v4f32_one_step: 418; SKX: # %bb.0: 419; SKX-NEXT: vrcpps %xmm0, %xmm1 # sched: [4:1.00] 420; SKX-NEXT: vfnmadd213ps {{.*#+}} xmm0 = -(xmm1 * xmm0) + mem sched: [10:0.50] 421; SKX-NEXT: vfmadd132ps {{.*#+}} xmm0 = (xmm0 * xmm1) + xmm1 sched: [4:0.50] 422; SKX-NEXT: retq # sched: [7:1.00] 423 %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x 424 ret <4 x float> %div 425} 426 427define <4 x float> @v4f32_two_step(<4 x float> %x) #2 { 428; SSE-LABEL: v4f32_two_step: 429; SSE: # %bb.0: 430; SSE-NEXT: rcpps %xmm0, %xmm2 431; SSE-NEXT: movaps %xmm0, %xmm3 432; SSE-NEXT: mulps %xmm2, %xmm3 433; SSE-NEXT: movaps {{.*#+}} xmm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 434; SSE-NEXT: movaps %xmm1, %xmm4 435; SSE-NEXT: subps %xmm3, %xmm4 436; SSE-NEXT: mulps %xmm2, %xmm4 437; SSE-NEXT: addps %xmm2, %xmm4 438; SSE-NEXT: mulps %xmm4, %xmm0 439; SSE-NEXT: subps %xmm0, %xmm1 440; SSE-NEXT: mulps %xmm4, %xmm1 441; SSE-NEXT: addps %xmm4, %xmm1 442; SSE-NEXT: movaps %xmm1, %xmm0 443; SSE-NEXT: retq 444; 445; AVX-RECIP-LABEL: v4f32_two_step: 446; AVX-RECIP: # %bb.0: 447; AVX-RECIP-NEXT: vrcpps %xmm0, %xmm1 448; AVX-RECIP-NEXT: vmulps %xmm1, %xmm0, %xmm2 449; AVX-RECIP-NEXT: vmovaps {{.*#+}} xmm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 450; AVX-RECIP-NEXT: vsubps %xmm2, %xmm3, %xmm2 451; AVX-RECIP-NEXT: vmulps %xmm2, %xmm1, %xmm2 452; AVX-RECIP-NEXT: vaddps %xmm2, %xmm1, %xmm1 453; AVX-RECIP-NEXT: vmulps %xmm1, %xmm0, %xmm0 454; AVX-RECIP-NEXT: vsubps %xmm0, %xmm3, %xmm0 455; AVX-RECIP-NEXT: vmulps %xmm0, %xmm1, %xmm0 456; AVX-RECIP-NEXT: vaddps %xmm0, %xmm1, %xmm0 457; AVX-RECIP-NEXT: retq 458; 459; FMA-RECIP-LABEL: v4f32_two_step: 460; FMA-RECIP: # %bb.0: 461; FMA-RECIP-NEXT: vrcpps %xmm0, %xmm1 462; FMA-RECIP-NEXT: vmovaps {{.*#+}} xmm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 463; FMA-RECIP-NEXT: vmovaps %xmm1, %xmm3 464; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} xmm3 = -(xmm0 * xmm3) + xmm2 465; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} xmm3 = (xmm3 * xmm1) + xmm1 466; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} xmm0 = -(xmm3 * xmm0) + xmm2 467; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} xmm0 = (xmm0 * xmm3) + xmm3 468; FMA-RECIP-NEXT: retq 469; 470; BTVER2-LABEL: v4f32_two_step: 471; BTVER2: # %bb.0: 472; BTVER2-NEXT: vmovaps {{.*#+}} xmm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [5:1.00] 473; BTVER2-NEXT: vrcpps %xmm0, %xmm1 # sched: [2:1.00] 474; BTVER2-NEXT: vmulps %xmm1, %xmm0, %xmm2 # sched: [2:1.00] 475; BTVER2-NEXT: vsubps %xmm2, %xmm3, %xmm2 # sched: [3:1.00] 476; BTVER2-NEXT: vmulps %xmm2, %xmm1, %xmm2 # sched: [2:1.00] 477; BTVER2-NEXT: vaddps %xmm2, %xmm1, %xmm1 # sched: [3:1.00] 478; BTVER2-NEXT: vmulps %xmm1, %xmm0, %xmm0 # sched: [2:1.00] 479; BTVER2-NEXT: vsubps %xmm0, %xmm3, %xmm0 # sched: [3:1.00] 480; BTVER2-NEXT: vmulps %xmm0, %xmm1, %xmm0 # sched: [2:1.00] 481; BTVER2-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00] 482; BTVER2-NEXT: retq # sched: [4:1.00] 483; 484; SANDY-LABEL: v4f32_two_step: 485; SANDY: # %bb.0: 486; SANDY-NEXT: vrcpps %xmm0, %xmm1 # sched: [5:1.00] 487; SANDY-NEXT: vmulps %xmm1, %xmm0, %xmm2 # sched: [5:1.00] 488; SANDY-NEXT: vmovaps {{.*#+}} xmm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [6:0.50] 489; SANDY-NEXT: vsubps %xmm2, %xmm3, %xmm2 # sched: [3:1.00] 490; SANDY-NEXT: vmulps %xmm2, %xmm1, %xmm2 # sched: [5:1.00] 491; SANDY-NEXT: vaddps %xmm2, %xmm1, %xmm1 # sched: [3:1.00] 492; SANDY-NEXT: vmulps %xmm1, %xmm0, %xmm0 # sched: [5:1.00] 493; SANDY-NEXT: vsubps %xmm0, %xmm3, %xmm0 # sched: [3:1.00] 494; SANDY-NEXT: vmulps %xmm0, %xmm1, %xmm0 # sched: [5:1.00] 495; SANDY-NEXT: vaddps %xmm0, %xmm1, %xmm0 # sched: [3:1.00] 496; SANDY-NEXT: retq # sched: [1:1.00] 497; 498; HASWELL-LABEL: v4f32_two_step: 499; HASWELL: # %bb.0: 500; HASWELL-NEXT: vrcpps %xmm0, %xmm1 # sched: [5:1.00] 501; HASWELL-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50] 502; HASWELL-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] 503; HASWELL-NEXT: vfnmadd213ps {{.*#+}} xmm3 = -(xmm0 * xmm3) + xmm2 sched: [5:0.50] 504; HASWELL-NEXT: vfmadd132ps {{.*#+}} xmm3 = (xmm3 * xmm1) + xmm1 sched: [5:0.50] 505; HASWELL-NEXT: vfnmadd213ps {{.*#+}} xmm0 = -(xmm3 * xmm0) + xmm2 sched: [5:0.50] 506; HASWELL-NEXT: vfmadd132ps {{.*#+}} xmm0 = (xmm0 * xmm3) + xmm3 sched: [5:0.50] 507; HASWELL-NEXT: retq # sched: [7:1.00] 508; 509; HASWELL-NO-FMA-LABEL: v4f32_two_step: 510; HASWELL-NO-FMA: # %bb.0: 511; HASWELL-NO-FMA-NEXT: vrcpps %xmm0, %xmm1 512; HASWELL-NO-FMA-NEXT: vmulps %xmm1, %xmm0, %xmm2 513; HASWELL-NO-FMA-NEXT: vbroadcastss {{.*#+}} xmm3 = [1,1,1,1] 514; HASWELL-NO-FMA-NEXT: vsubps %xmm2, %xmm3, %xmm2 515; HASWELL-NO-FMA-NEXT: vmulps %xmm2, %xmm1, %xmm2 516; HASWELL-NO-FMA-NEXT: vaddps %xmm2, %xmm1, %xmm1 517; HASWELL-NO-FMA-NEXT: vmulps %xmm1, %xmm0, %xmm0 518; HASWELL-NO-FMA-NEXT: vsubps %xmm0, %xmm3, %xmm0 519; HASWELL-NO-FMA-NEXT: vmulps %xmm0, %xmm1, %xmm0 520; HASWELL-NO-FMA-NEXT: vaddps %xmm0, %xmm1, %xmm0 521; HASWELL-NO-FMA-NEXT: retq 522; 523; KNL-LABEL: v4f32_two_step: 524; KNL: # %bb.0: 525; KNL-NEXT: vrcpps %xmm0, %xmm1 # sched: [5:1.00] 526; KNL-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50] 527; KNL-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:1.00] 528; KNL-NEXT: vfnmadd213ps {{.*#+}} xmm3 = -(xmm0 * xmm3) + xmm2 sched: [5:0.50] 529; KNL-NEXT: vfmadd132ps {{.*#+}} xmm3 = (xmm3 * xmm1) + xmm1 sched: [5:0.50] 530; KNL-NEXT: vfnmadd213ps {{.*#+}} xmm0 = -(xmm3 * xmm0) + xmm2 sched: [5:0.50] 531; KNL-NEXT: vfmadd132ps {{.*#+}} xmm0 = (xmm0 * xmm3) + xmm3 sched: [5:0.50] 532; KNL-NEXT: retq # sched: [7:1.00] 533; 534; SKX-LABEL: v4f32_two_step: 535; SKX: # %bb.0: 536; SKX-NEXT: vrcpps %xmm0, %xmm1 # sched: [4:1.00] 537; SKX-NEXT: vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50] 538; SKX-NEXT: vmovaps %xmm1, %xmm3 # sched: [1:0.33] 539; SKX-NEXT: vfnmadd213ps {{.*#+}} xmm3 = -(xmm0 * xmm3) + xmm2 sched: [4:0.50] 540; SKX-NEXT: vfmadd132ps {{.*#+}} xmm3 = (xmm3 * xmm1) + xmm1 sched: [4:0.50] 541; SKX-NEXT: vfnmadd213ps {{.*#+}} xmm0 = -(xmm3 * xmm0) + xmm2 sched: [4:0.50] 542; SKX-NEXT: vfmadd132ps {{.*#+}} xmm0 = (xmm0 * xmm3) + xmm3 sched: [4:0.50] 543; SKX-NEXT: retq # sched: [7:1.00] 544 %div = fdiv fast <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, %x 545 ret <4 x float> %div 546} 547 548define <8 x float> @v8f32_no_estimate(<8 x float> %x) #0 { 549; SSE-LABEL: v8f32_no_estimate: 550; SSE: # %bb.0: 551; SSE-NEXT: movaps {{.*#+}} xmm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 552; SSE-NEXT: movaps %xmm2, %xmm3 553; SSE-NEXT: divps %xmm0, %xmm3 554; SSE-NEXT: divps %xmm1, %xmm2 555; SSE-NEXT: movaps %xmm3, %xmm0 556; SSE-NEXT: movaps %xmm2, %xmm1 557; SSE-NEXT: retq 558; 559; AVX-RECIP-LABEL: v8f32_no_estimate: 560; AVX-RECIP: # %bb.0: 561; AVX-RECIP-NEXT: vmovaps {{.*#+}} ymm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 562; AVX-RECIP-NEXT: vdivps %ymm0, %ymm1, %ymm0 563; AVX-RECIP-NEXT: retq 564; 565; FMA-RECIP-LABEL: v8f32_no_estimate: 566; FMA-RECIP: # %bb.0: 567; FMA-RECIP-NEXT: vmovaps {{.*#+}} ymm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 568; FMA-RECIP-NEXT: vdivps %ymm0, %ymm1, %ymm0 569; FMA-RECIP-NEXT: retq 570; 571; BTVER2-LABEL: v8f32_no_estimate: 572; BTVER2: # %bb.0: 573; BTVER2-NEXT: vmovaps {{.*#+}} ymm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [5:1.00] 574; BTVER2-NEXT: vdivps %ymm0, %ymm1, %ymm0 # sched: [38:38.00] 575; BTVER2-NEXT: retq # sched: [4:1.00] 576; 577; SANDY-LABEL: v8f32_no_estimate: 578; SANDY: # %bb.0: 579; SANDY-NEXT: vmovaps {{.*#+}} ymm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [7:0.50] 580; SANDY-NEXT: vdivps %ymm0, %ymm1, %ymm0 # sched: [29:28.00] 581; SANDY-NEXT: retq # sched: [1:1.00] 582; 583; HASWELL-LABEL: v8f32_no_estimate: 584; HASWELL: # %bb.0: 585; HASWELL-NEXT: vbroadcastss {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 586; HASWELL-NEXT: vdivps %ymm0, %ymm1, %ymm0 # sched: [21:14.00] 587; HASWELL-NEXT: retq # sched: [7:1.00] 588; 589; HASWELL-NO-FMA-LABEL: v8f32_no_estimate: 590; HASWELL-NO-FMA: # %bb.0: 591; HASWELL-NO-FMA-NEXT: vbroadcastss {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1] 592; HASWELL-NO-FMA-NEXT: vdivps %ymm0, %ymm1, %ymm0 593; HASWELL-NO-FMA-NEXT: retq 594; 595; KNL-LABEL: v8f32_no_estimate: 596; KNL: # %bb.0: 597; KNL-NEXT: vbroadcastss {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 598; KNL-NEXT: vdivps %ymm0, %ymm1, %ymm0 # sched: [21:14.00] 599; KNL-NEXT: retq # sched: [7:1.00] 600; 601; SKX-LABEL: v8f32_no_estimate: 602; SKX: # %bb.0: 603; SKX-NEXT: vbroadcastss {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 604; SKX-NEXT: vdivps %ymm0, %ymm1, %ymm0 # sched: [11:5.00] 605; SKX-NEXT: retq # sched: [7:1.00] 606 %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x 607 ret <8 x float> %div 608} 609 610define <8 x float> @v8f32_one_step(<8 x float> %x) #1 { 611; SSE-LABEL: v8f32_one_step: 612; SSE: # %bb.0: 613; SSE-NEXT: rcpps %xmm0, %xmm4 614; SSE-NEXT: mulps %xmm4, %xmm0 615; SSE-NEXT: movaps {{.*#+}} xmm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 616; SSE-NEXT: movaps %xmm2, %xmm3 617; SSE-NEXT: subps %xmm0, %xmm3 618; SSE-NEXT: mulps %xmm4, %xmm3 619; SSE-NEXT: addps %xmm4, %xmm3 620; SSE-NEXT: rcpps %xmm1, %xmm0 621; SSE-NEXT: mulps %xmm0, %xmm1 622; SSE-NEXT: subps %xmm1, %xmm2 623; SSE-NEXT: mulps %xmm0, %xmm2 624; SSE-NEXT: addps %xmm0, %xmm2 625; SSE-NEXT: movaps %xmm3, %xmm0 626; SSE-NEXT: movaps %xmm2, %xmm1 627; SSE-NEXT: retq 628; 629; AVX-RECIP-LABEL: v8f32_one_step: 630; AVX-RECIP: # %bb.0: 631; AVX-RECIP-NEXT: vrcpps %ymm0, %ymm1 632; AVX-RECIP-NEXT: vmulps %ymm1, %ymm0, %ymm0 633; AVX-RECIP-NEXT: vmovaps {{.*#+}} ymm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 634; AVX-RECIP-NEXT: vsubps %ymm0, %ymm2, %ymm0 635; AVX-RECIP-NEXT: vmulps %ymm0, %ymm1, %ymm0 636; AVX-RECIP-NEXT: vaddps %ymm0, %ymm1, %ymm0 637; AVX-RECIP-NEXT: retq 638; 639; FMA-RECIP-LABEL: v8f32_one_step: 640; FMA-RECIP: # %bb.0: 641; FMA-RECIP-NEXT: vrcpps %ymm0, %ymm1 642; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) + mem 643; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm1) + ymm1 644; FMA-RECIP-NEXT: retq 645; 646; BTVER2-LABEL: v8f32_one_step: 647; BTVER2: # %bb.0: 648; BTVER2-NEXT: vmovaps {{.*#+}} ymm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [5:1.00] 649; BTVER2-NEXT: vrcpps %ymm0, %ymm1 # sched: [2:2.00] 650; BTVER2-NEXT: vmulps %ymm1, %ymm0, %ymm0 # sched: [2:2.00] 651; BTVER2-NEXT: vsubps %ymm0, %ymm2, %ymm0 # sched: [3:2.00] 652; BTVER2-NEXT: vmulps %ymm0, %ymm1, %ymm0 # sched: [2:2.00] 653; BTVER2-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:2.00] 654; BTVER2-NEXT: retq # sched: [4:1.00] 655; 656; SANDY-LABEL: v8f32_one_step: 657; SANDY: # %bb.0: 658; SANDY-NEXT: vrcpps %ymm0, %ymm1 # sched: [7:2.00] 659; SANDY-NEXT: vmulps %ymm1, %ymm0, %ymm0 # sched: [5:1.00] 660; SANDY-NEXT: vmovaps {{.*#+}} ymm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [7:0.50] 661; SANDY-NEXT: vsubps %ymm0, %ymm2, %ymm0 # sched: [3:1.00] 662; SANDY-NEXT: vmulps %ymm0, %ymm1, %ymm0 # sched: [5:1.00] 663; SANDY-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00] 664; SANDY-NEXT: retq # sched: [1:1.00] 665; 666; HASWELL-LABEL: v8f32_one_step: 667; HASWELL: # %bb.0: 668; HASWELL-NEXT: vrcpps %ymm0, %ymm1 # sched: [11:2.00] 669; HASWELL-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 670; HASWELL-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm2 sched: [5:0.50] 671; HASWELL-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm1) + ymm1 sched: [5:0.50] 672; HASWELL-NEXT: retq # sched: [7:1.00] 673; 674; HASWELL-NO-FMA-LABEL: v8f32_one_step: 675; HASWELL-NO-FMA: # %bb.0: 676; HASWELL-NO-FMA-NEXT: vrcpps %ymm0, %ymm1 677; HASWELL-NO-FMA-NEXT: vmulps %ymm1, %ymm0, %ymm0 678; HASWELL-NO-FMA-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] 679; HASWELL-NO-FMA-NEXT: vsubps %ymm0, %ymm2, %ymm0 680; HASWELL-NO-FMA-NEXT: vmulps %ymm0, %ymm1, %ymm0 681; HASWELL-NO-FMA-NEXT: vaddps %ymm0, %ymm1, %ymm0 682; HASWELL-NO-FMA-NEXT: retq 683; 684; KNL-LABEL: v8f32_one_step: 685; KNL: # %bb.0: 686; KNL-NEXT: vrcpps %ymm0, %ymm1 # sched: [11:2.00] 687; KNL-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 688; KNL-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm2 sched: [5:0.50] 689; KNL-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm1) + ymm1 sched: [5:0.50] 690; KNL-NEXT: retq # sched: [7:1.00] 691; 692; SKX-LABEL: v8f32_one_step: 693; SKX: # %bb.0: 694; SKX-NEXT: vrcpps %ymm0, %ymm1 # sched: [4:1.00] 695; SKX-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) + mem sched: [11:0.50] 696; SKX-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm1) + ymm1 sched: [4:0.50] 697; SKX-NEXT: retq # sched: [7:1.00] 698 %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x 699 ret <8 x float> %div 700} 701 702define <8 x float> @v8f32_two_step(<8 x float> %x) #2 { 703; SSE-LABEL: v8f32_two_step: 704; SSE: # %bb.0: 705; SSE-NEXT: movaps %xmm1, %xmm2 706; SSE-NEXT: rcpps %xmm0, %xmm3 707; SSE-NEXT: movaps %xmm0, %xmm4 708; SSE-NEXT: mulps %xmm3, %xmm4 709; SSE-NEXT: movaps {{.*#+}} xmm1 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 710; SSE-NEXT: movaps %xmm1, %xmm5 711; SSE-NEXT: subps %xmm4, %xmm5 712; SSE-NEXT: mulps %xmm3, %xmm5 713; SSE-NEXT: addps %xmm3, %xmm5 714; SSE-NEXT: mulps %xmm5, %xmm0 715; SSE-NEXT: movaps %xmm1, %xmm3 716; SSE-NEXT: subps %xmm0, %xmm3 717; SSE-NEXT: mulps %xmm5, %xmm3 718; SSE-NEXT: addps %xmm5, %xmm3 719; SSE-NEXT: rcpps %xmm2, %xmm0 720; SSE-NEXT: movaps %xmm2, %xmm4 721; SSE-NEXT: mulps %xmm0, %xmm4 722; SSE-NEXT: movaps %xmm1, %xmm5 723; SSE-NEXT: subps %xmm4, %xmm5 724; SSE-NEXT: mulps %xmm0, %xmm5 725; SSE-NEXT: addps %xmm0, %xmm5 726; SSE-NEXT: mulps %xmm5, %xmm2 727; SSE-NEXT: subps %xmm2, %xmm1 728; SSE-NEXT: mulps %xmm5, %xmm1 729; SSE-NEXT: addps %xmm5, %xmm1 730; SSE-NEXT: movaps %xmm3, %xmm0 731; SSE-NEXT: retq 732; 733; AVX-RECIP-LABEL: v8f32_two_step: 734; AVX-RECIP: # %bb.0: 735; AVX-RECIP-NEXT: vrcpps %ymm0, %ymm1 736; AVX-RECIP-NEXT: vmulps %ymm1, %ymm0, %ymm2 737; AVX-RECIP-NEXT: vmovaps {{.*#+}} ymm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 738; AVX-RECIP-NEXT: vsubps %ymm2, %ymm3, %ymm2 739; AVX-RECIP-NEXT: vmulps %ymm2, %ymm1, %ymm2 740; AVX-RECIP-NEXT: vaddps %ymm2, %ymm1, %ymm1 741; AVX-RECIP-NEXT: vmulps %ymm1, %ymm0, %ymm0 742; AVX-RECIP-NEXT: vsubps %ymm0, %ymm3, %ymm0 743; AVX-RECIP-NEXT: vmulps %ymm0, %ymm1, %ymm0 744; AVX-RECIP-NEXT: vaddps %ymm0, %ymm1, %ymm0 745; AVX-RECIP-NEXT: retq 746; 747; FMA-RECIP-LABEL: v8f32_two_step: 748; FMA-RECIP: # %bb.0: 749; FMA-RECIP-NEXT: vrcpps %ymm0, %ymm1 750; FMA-RECIP-NEXT: vmovaps {{.*#+}} ymm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 751; FMA-RECIP-NEXT: vmovaps %ymm1, %ymm3 752; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} ymm3 = -(ymm0 * ymm3) + ymm2 753; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} ymm3 = (ymm3 * ymm1) + ymm1 754; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm3 * ymm0) + ymm2 755; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm3) + ymm3 756; FMA-RECIP-NEXT: retq 757; 758; BTVER2-LABEL: v8f32_two_step: 759; BTVER2: # %bb.0: 760; BTVER2-NEXT: vmovaps {{.*#+}} ymm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [5:1.00] 761; BTVER2-NEXT: vrcpps %ymm0, %ymm1 # sched: [2:2.00] 762; BTVER2-NEXT: vmulps %ymm1, %ymm0, %ymm2 # sched: [2:2.00] 763; BTVER2-NEXT: vsubps %ymm2, %ymm3, %ymm2 # sched: [3:2.00] 764; BTVER2-NEXT: vmulps %ymm2, %ymm1, %ymm2 # sched: [2:2.00] 765; BTVER2-NEXT: vaddps %ymm2, %ymm1, %ymm1 # sched: [3:2.00] 766; BTVER2-NEXT: vmulps %ymm1, %ymm0, %ymm0 # sched: [2:2.00] 767; BTVER2-NEXT: vsubps %ymm0, %ymm3, %ymm0 # sched: [3:2.00] 768; BTVER2-NEXT: vmulps %ymm0, %ymm1, %ymm0 # sched: [2:2.00] 769; BTVER2-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:2.00] 770; BTVER2-NEXT: retq # sched: [4:1.00] 771; 772; SANDY-LABEL: v8f32_two_step: 773; SANDY: # %bb.0: 774; SANDY-NEXT: vrcpps %ymm0, %ymm1 # sched: [7:2.00] 775; SANDY-NEXT: vmulps %ymm1, %ymm0, %ymm2 # sched: [5:1.00] 776; SANDY-NEXT: vmovaps {{.*#+}} ymm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [7:0.50] 777; SANDY-NEXT: vsubps %ymm2, %ymm3, %ymm2 # sched: [3:1.00] 778; SANDY-NEXT: vmulps %ymm2, %ymm1, %ymm2 # sched: [5:1.00] 779; SANDY-NEXT: vaddps %ymm2, %ymm1, %ymm1 # sched: [3:1.00] 780; SANDY-NEXT: vmulps %ymm1, %ymm0, %ymm0 # sched: [5:1.00] 781; SANDY-NEXT: vsubps %ymm0, %ymm3, %ymm0 # sched: [3:1.00] 782; SANDY-NEXT: vmulps %ymm0, %ymm1, %ymm0 # sched: [5:1.00] 783; SANDY-NEXT: vaddps %ymm0, %ymm1, %ymm0 # sched: [3:1.00] 784; SANDY-NEXT: retq # sched: [1:1.00] 785; 786; HASWELL-LABEL: v8f32_two_step: 787; HASWELL: # %bb.0: 788; HASWELL-NEXT: vrcpps %ymm0, %ymm1 # sched: [11:2.00] 789; HASWELL-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 790; HASWELL-NEXT: vmovaps %ymm1, %ymm3 # sched: [1:1.00] 791; HASWELL-NEXT: vfnmadd213ps {{.*#+}} ymm3 = -(ymm0 * ymm3) + ymm2 sched: [5:0.50] 792; HASWELL-NEXT: vfmadd132ps {{.*#+}} ymm3 = (ymm3 * ymm1) + ymm1 sched: [5:0.50] 793; HASWELL-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm3 * ymm0) + ymm2 sched: [5:0.50] 794; HASWELL-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm3) + ymm3 sched: [5:0.50] 795; HASWELL-NEXT: retq # sched: [7:1.00] 796; 797; HASWELL-NO-FMA-LABEL: v8f32_two_step: 798; HASWELL-NO-FMA: # %bb.0: 799; HASWELL-NO-FMA-NEXT: vrcpps %ymm0, %ymm1 800; HASWELL-NO-FMA-NEXT: vmulps %ymm1, %ymm0, %ymm2 801; HASWELL-NO-FMA-NEXT: vbroadcastss {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1] 802; HASWELL-NO-FMA-NEXT: vsubps %ymm2, %ymm3, %ymm2 803; HASWELL-NO-FMA-NEXT: vmulps %ymm2, %ymm1, %ymm2 804; HASWELL-NO-FMA-NEXT: vaddps %ymm2, %ymm1, %ymm1 805; HASWELL-NO-FMA-NEXT: vmulps %ymm1, %ymm0, %ymm0 806; HASWELL-NO-FMA-NEXT: vsubps %ymm0, %ymm3, %ymm0 807; HASWELL-NO-FMA-NEXT: vmulps %ymm0, %ymm1, %ymm0 808; HASWELL-NO-FMA-NEXT: vaddps %ymm0, %ymm1, %ymm0 809; HASWELL-NO-FMA-NEXT: retq 810; 811; KNL-LABEL: v8f32_two_step: 812; KNL: # %bb.0: 813; KNL-NEXT: vrcpps %ymm0, %ymm1 # sched: [11:2.00] 814; KNL-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 815; KNL-NEXT: vmovaps %ymm1, %ymm3 # sched: [1:1.00] 816; KNL-NEXT: vfnmadd213ps {{.*#+}} ymm3 = -(ymm0 * ymm3) + ymm2 sched: [5:0.50] 817; KNL-NEXT: vfmadd132ps {{.*#+}} ymm3 = (ymm3 * ymm1) + ymm1 sched: [5:0.50] 818; KNL-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm3 * ymm0) + ymm2 sched: [5:0.50] 819; KNL-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm3) + ymm3 sched: [5:0.50] 820; KNL-NEXT: retq # sched: [7:1.00] 821; 822; SKX-LABEL: v8f32_two_step: 823; SKX: # %bb.0: 824; SKX-NEXT: vrcpps %ymm0, %ymm1 # sched: [4:1.00] 825; SKX-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 826; SKX-NEXT: vmovaps %ymm1, %ymm3 # sched: [1:0.33] 827; SKX-NEXT: vfnmadd213ps {{.*#+}} ymm3 = -(ymm0 * ymm3) + ymm2 sched: [4:0.50] 828; SKX-NEXT: vfmadd132ps {{.*#+}} ymm3 = (ymm3 * ymm1) + ymm1 sched: [4:0.50] 829; SKX-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm3 * ymm0) + ymm2 sched: [4:0.50] 830; SKX-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm3) + ymm3 sched: [4:0.50] 831; SKX-NEXT: retq # sched: [7:1.00] 832 %div = fdiv fast <8 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x 833 ret <8 x float> %div 834} 835 836define <16 x float> @v16f32_no_estimate(<16 x float> %x) #0 { 837; SSE-LABEL: v16f32_no_estimate: 838; SSE: # %bb.0: 839; SSE-NEXT: movaps {{.*#+}} xmm4 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 840; SSE-NEXT: movaps %xmm4, %xmm5 841; SSE-NEXT: divps %xmm0, %xmm5 842; SSE-NEXT: movaps %xmm4, %xmm6 843; SSE-NEXT: divps %xmm1, %xmm6 844; SSE-NEXT: movaps %xmm4, %xmm7 845; SSE-NEXT: divps %xmm2, %xmm7 846; SSE-NEXT: divps %xmm3, %xmm4 847; SSE-NEXT: movaps %xmm5, %xmm0 848; SSE-NEXT: movaps %xmm6, %xmm1 849; SSE-NEXT: movaps %xmm7, %xmm2 850; SSE-NEXT: movaps %xmm4, %xmm3 851; SSE-NEXT: retq 852; 853; AVX-RECIP-LABEL: v16f32_no_estimate: 854; AVX-RECIP: # %bb.0: 855; AVX-RECIP-NEXT: vmovaps {{.*#+}} ymm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 856; AVX-RECIP-NEXT: vdivps %ymm0, %ymm2, %ymm0 857; AVX-RECIP-NEXT: vdivps %ymm1, %ymm2, %ymm1 858; AVX-RECIP-NEXT: retq 859; 860; FMA-RECIP-LABEL: v16f32_no_estimate: 861; FMA-RECIP: # %bb.0: 862; FMA-RECIP-NEXT: vmovaps {{.*#+}} ymm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 863; FMA-RECIP-NEXT: vdivps %ymm0, %ymm2, %ymm0 864; FMA-RECIP-NEXT: vdivps %ymm1, %ymm2, %ymm1 865; FMA-RECIP-NEXT: retq 866; 867; BTVER2-LABEL: v16f32_no_estimate: 868; BTVER2: # %bb.0: 869; BTVER2-NEXT: vmovaps {{.*#+}} ymm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [5:1.00] 870; BTVER2-NEXT: vdivps %ymm0, %ymm2, %ymm0 # sched: [38:38.00] 871; BTVER2-NEXT: vdivps %ymm1, %ymm2, %ymm1 # sched: [38:38.00] 872; BTVER2-NEXT: retq # sched: [4:1.00] 873; 874; SANDY-LABEL: v16f32_no_estimate: 875; SANDY: # %bb.0: 876; SANDY-NEXT: vmovaps {{.*#+}} ymm2 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [7:0.50] 877; SANDY-NEXT: vdivps %ymm0, %ymm2, %ymm0 # sched: [29:28.00] 878; SANDY-NEXT: vdivps %ymm1, %ymm2, %ymm1 # sched: [29:28.00] 879; SANDY-NEXT: retq # sched: [1:1.00] 880; 881; HASWELL-LABEL: v16f32_no_estimate: 882; HASWELL: # %bb.0: 883; HASWELL-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 884; HASWELL-NEXT: vdivps %ymm0, %ymm2, %ymm0 # sched: [21:14.00] 885; HASWELL-NEXT: vdivps %ymm1, %ymm2, %ymm1 # sched: [21:14.00] 886; HASWELL-NEXT: retq # sched: [7:1.00] 887; 888; HASWELL-NO-FMA-LABEL: v16f32_no_estimate: 889; HASWELL-NO-FMA: # %bb.0: 890; HASWELL-NO-FMA-NEXT: vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] 891; HASWELL-NO-FMA-NEXT: vdivps %ymm0, %ymm2, %ymm0 892; HASWELL-NO-FMA-NEXT: vdivps %ymm1, %ymm2, %ymm1 893; HASWELL-NO-FMA-NEXT: retq 894; 895; KNL-LABEL: v16f32_no_estimate: 896; KNL: # %bb.0: 897; KNL-NEXT: vbroadcastss {{.*#+}} zmm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] sched: [10:1.00] 898; KNL-NEXT: vdivps %zmm0, %zmm1, %zmm0 # sched: [21:14.00] 899; KNL-NEXT: retq # sched: [7:1.00] 900; 901; SKX-LABEL: v16f32_no_estimate: 902; SKX: # %bb.0: 903; SKX-NEXT: vbroadcastss {{.*#+}} zmm1 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] sched: [8:0.50] 904; SKX-NEXT: vdivps %zmm0, %zmm1, %zmm0 # sched: [18:10.00] 905; SKX-NEXT: retq # sched: [7:1.00] 906 %div = fdiv fast <16 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x 907 ret <16 x float> %div 908} 909 910define <16 x float> @v16f32_one_step(<16 x float> %x) #1 { 911; SSE-LABEL: v16f32_one_step: 912; SSE: # %bb.0: 913; SSE-NEXT: movaps %xmm3, %xmm4 914; SSE-NEXT: movaps %xmm0, %xmm5 915; SSE-NEXT: rcpps %xmm0, %xmm6 916; SSE-NEXT: mulps %xmm6, %xmm5 917; SSE-NEXT: movaps {{.*#+}} xmm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 918; SSE-NEXT: movaps %xmm3, %xmm0 919; SSE-NEXT: subps %xmm5, %xmm0 920; SSE-NEXT: mulps %xmm6, %xmm0 921; SSE-NEXT: addps %xmm6, %xmm0 922; SSE-NEXT: rcpps %xmm1, %xmm6 923; SSE-NEXT: mulps %xmm6, %xmm1 924; SSE-NEXT: movaps %xmm3, %xmm5 925; SSE-NEXT: subps %xmm1, %xmm5 926; SSE-NEXT: mulps %xmm6, %xmm5 927; SSE-NEXT: addps %xmm6, %xmm5 928; SSE-NEXT: rcpps %xmm2, %xmm1 929; SSE-NEXT: mulps %xmm1, %xmm2 930; SSE-NEXT: movaps %xmm3, %xmm6 931; SSE-NEXT: subps %xmm2, %xmm6 932; SSE-NEXT: mulps %xmm1, %xmm6 933; SSE-NEXT: addps %xmm1, %xmm6 934; SSE-NEXT: rcpps %xmm4, %xmm1 935; SSE-NEXT: mulps %xmm1, %xmm4 936; SSE-NEXT: subps %xmm4, %xmm3 937; SSE-NEXT: mulps %xmm1, %xmm3 938; SSE-NEXT: addps %xmm1, %xmm3 939; SSE-NEXT: movaps %xmm5, %xmm1 940; SSE-NEXT: movaps %xmm6, %xmm2 941; SSE-NEXT: retq 942; 943; AVX-RECIP-LABEL: v16f32_one_step: 944; AVX-RECIP: # %bb.0: 945; AVX-RECIP-NEXT: vrcpps %ymm0, %ymm2 946; AVX-RECIP-NEXT: vmulps %ymm2, %ymm0, %ymm0 947; AVX-RECIP-NEXT: vmovaps {{.*#+}} ymm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 948; AVX-RECIP-NEXT: vsubps %ymm0, %ymm3, %ymm0 949; AVX-RECIP-NEXT: vmulps %ymm0, %ymm2, %ymm0 950; AVX-RECIP-NEXT: vaddps %ymm0, %ymm2, %ymm0 951; AVX-RECIP-NEXT: vrcpps %ymm1, %ymm2 952; AVX-RECIP-NEXT: vmulps %ymm2, %ymm1, %ymm1 953; AVX-RECIP-NEXT: vsubps %ymm1, %ymm3, %ymm1 954; AVX-RECIP-NEXT: vmulps %ymm1, %ymm2, %ymm1 955; AVX-RECIP-NEXT: vaddps %ymm1, %ymm2, %ymm1 956; AVX-RECIP-NEXT: retq 957; 958; FMA-RECIP-LABEL: v16f32_one_step: 959; FMA-RECIP: # %bb.0: 960; FMA-RECIP-NEXT: vrcpps %ymm0, %ymm2 961; FMA-RECIP-NEXT: vmovaps {{.*#+}} ymm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 962; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm2 * ymm0) + ymm3 963; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm2) + ymm2 964; FMA-RECIP-NEXT: vrcpps %ymm1, %ymm2 965; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} ymm1 = -(ymm2 * ymm1) + ymm3 966; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} ymm1 = (ymm1 * ymm2) + ymm2 967; FMA-RECIP-NEXT: retq 968; 969; BTVER2-LABEL: v16f32_one_step: 970; BTVER2: # %bb.0: 971; BTVER2-NEXT: vmovaps {{.*#+}} ymm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [5:1.00] 972; BTVER2-NEXT: vrcpps %ymm0, %ymm2 # sched: [2:2.00] 973; BTVER2-NEXT: vrcpps %ymm1, %ymm4 # sched: [2:2.00] 974; BTVER2-NEXT: vmulps %ymm2, %ymm0, %ymm0 # sched: [2:2.00] 975; BTVER2-NEXT: vmulps %ymm4, %ymm1, %ymm1 # sched: [2:2.00] 976; BTVER2-NEXT: vsubps %ymm0, %ymm3, %ymm0 # sched: [3:2.00] 977; BTVER2-NEXT: vsubps %ymm1, %ymm3, %ymm1 # sched: [3:2.00] 978; BTVER2-NEXT: vmulps %ymm0, %ymm2, %ymm0 # sched: [2:2.00] 979; BTVER2-NEXT: vmulps %ymm1, %ymm4, %ymm1 # sched: [2:2.00] 980; BTVER2-NEXT: vaddps %ymm0, %ymm2, %ymm0 # sched: [3:2.00] 981; BTVER2-NEXT: vaddps %ymm1, %ymm4, %ymm1 # sched: [3:2.00] 982; BTVER2-NEXT: retq # sched: [4:1.00] 983; 984; SANDY-LABEL: v16f32_one_step: 985; SANDY: # %bb.0: 986; SANDY-NEXT: vrcpps %ymm0, %ymm2 # sched: [7:2.00] 987; SANDY-NEXT: vmulps %ymm2, %ymm0, %ymm0 # sched: [5:1.00] 988; SANDY-NEXT: vmovaps {{.*#+}} ymm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [7:0.50] 989; SANDY-NEXT: vsubps %ymm0, %ymm3, %ymm0 # sched: [3:1.00] 990; SANDY-NEXT: vmulps %ymm0, %ymm2, %ymm0 # sched: [5:1.00] 991; SANDY-NEXT: vaddps %ymm0, %ymm2, %ymm0 # sched: [3:1.00] 992; SANDY-NEXT: vrcpps %ymm1, %ymm2 # sched: [7:2.00] 993; SANDY-NEXT: vmulps %ymm2, %ymm1, %ymm1 # sched: [5:1.00] 994; SANDY-NEXT: vsubps %ymm1, %ymm3, %ymm1 # sched: [3:1.00] 995; SANDY-NEXT: vmulps %ymm1, %ymm2, %ymm1 # sched: [5:1.00] 996; SANDY-NEXT: vaddps %ymm1, %ymm2, %ymm1 # sched: [3:1.00] 997; SANDY-NEXT: retq # sched: [1:1.00] 998; 999; HASWELL-LABEL: v16f32_one_step: 1000; HASWELL: # %bb.0: 1001; HASWELL-NEXT: vrcpps %ymm0, %ymm2 # sched: [11:2.00] 1002; HASWELL-NEXT: vbroadcastss {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 1003; HASWELL-NEXT: vrcpps %ymm1, %ymm4 # sched: [11:2.00] 1004; HASWELL-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm2 * ymm0) + ymm3 sched: [5:0.50] 1005; HASWELL-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm2) + ymm2 sched: [5:0.50] 1006; HASWELL-NEXT: vfnmadd213ps {{.*#+}} ymm1 = -(ymm4 * ymm1) + ymm3 sched: [5:0.50] 1007; HASWELL-NEXT: vfmadd132ps {{.*#+}} ymm1 = (ymm1 * ymm4) + ymm4 sched: [5:0.50] 1008; HASWELL-NEXT: retq # sched: [7:1.00] 1009; 1010; HASWELL-NO-FMA-LABEL: v16f32_one_step: 1011; HASWELL-NO-FMA: # %bb.0: 1012; HASWELL-NO-FMA-NEXT: vrcpps %ymm0, %ymm2 1013; HASWELL-NO-FMA-NEXT: vmulps %ymm2, %ymm0, %ymm0 1014; HASWELL-NO-FMA-NEXT: vbroadcastss {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1] 1015; HASWELL-NO-FMA-NEXT: vsubps %ymm0, %ymm3, %ymm0 1016; HASWELL-NO-FMA-NEXT: vmulps %ymm0, %ymm2, %ymm0 1017; HASWELL-NO-FMA-NEXT: vaddps %ymm0, %ymm2, %ymm0 1018; HASWELL-NO-FMA-NEXT: vrcpps %ymm1, %ymm2 1019; HASWELL-NO-FMA-NEXT: vmulps %ymm2, %ymm1, %ymm1 1020; HASWELL-NO-FMA-NEXT: vsubps %ymm1, %ymm3, %ymm1 1021; HASWELL-NO-FMA-NEXT: vmulps %ymm1, %ymm2, %ymm1 1022; HASWELL-NO-FMA-NEXT: vaddps %ymm1, %ymm2, %ymm1 1023; HASWELL-NO-FMA-NEXT: retq 1024; 1025; KNL-LABEL: v16f32_one_step: 1026; KNL: # %bb.0: 1027; KNL-NEXT: vrcp14ps %zmm0, %zmm1 # sched: [11:2.00] 1028; KNL-NEXT: vfnmadd213ps {{.*#+}} zmm0 = -(zmm1 * zmm0) + mem sched: [12:0.50] 1029; KNL-NEXT: vfmadd132ps {{.*#+}} zmm0 = (zmm0 * zmm1) + zmm1 sched: [5:0.50] 1030; KNL-NEXT: retq # sched: [7:1.00] 1031; 1032; SKX-LABEL: v16f32_one_step: 1033; SKX: # %bb.0: 1034; SKX-NEXT: vrcp14ps %zmm0, %zmm1 # sched: [4:2.00] 1035; SKX-NEXT: vfnmadd213ps {{.*#+}} zmm0 = -(zmm1 * zmm0) + mem sched: [11:0.50] 1036; SKX-NEXT: vfmadd132ps {{.*#+}} zmm0 = (zmm0 * zmm1) + zmm1 sched: [4:0.50] 1037; SKX-NEXT: retq # sched: [7:1.00] 1038 %div = fdiv fast <16 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x 1039 ret <16 x float> %div 1040} 1041 1042define <16 x float> @v16f32_two_step(<16 x float> %x) #2 { 1043; SSE-LABEL: v16f32_two_step: 1044; SSE: # %bb.0: 1045; SSE-NEXT: movaps %xmm3, %xmm4 1046; SSE-NEXT: movaps %xmm1, %xmm5 1047; SSE-NEXT: movaps %xmm0, %xmm1 1048; SSE-NEXT: rcpps %xmm0, %xmm0 1049; SSE-NEXT: movaps %xmm1, %xmm6 1050; SSE-NEXT: mulps %xmm0, %xmm6 1051; SSE-NEXT: movaps {{.*#+}} xmm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 1052; SSE-NEXT: movaps %xmm3, %xmm7 1053; SSE-NEXT: subps %xmm6, %xmm7 1054; SSE-NEXT: mulps %xmm0, %xmm7 1055; SSE-NEXT: addps %xmm0, %xmm7 1056; SSE-NEXT: mulps %xmm7, %xmm1 1057; SSE-NEXT: movaps %xmm3, %xmm0 1058; SSE-NEXT: subps %xmm1, %xmm0 1059; SSE-NEXT: mulps %xmm7, %xmm0 1060; SSE-NEXT: addps %xmm7, %xmm0 1061; SSE-NEXT: rcpps %xmm5, %xmm1 1062; SSE-NEXT: movaps %xmm5, %xmm6 1063; SSE-NEXT: mulps %xmm1, %xmm6 1064; SSE-NEXT: movaps %xmm3, %xmm7 1065; SSE-NEXT: subps %xmm6, %xmm7 1066; SSE-NEXT: mulps %xmm1, %xmm7 1067; SSE-NEXT: addps %xmm1, %xmm7 1068; SSE-NEXT: mulps %xmm7, %xmm5 1069; SSE-NEXT: movaps %xmm3, %xmm1 1070; SSE-NEXT: subps %xmm5, %xmm1 1071; SSE-NEXT: mulps %xmm7, %xmm1 1072; SSE-NEXT: addps %xmm7, %xmm1 1073; SSE-NEXT: rcpps %xmm2, %xmm5 1074; SSE-NEXT: movaps %xmm2, %xmm6 1075; SSE-NEXT: mulps %xmm5, %xmm6 1076; SSE-NEXT: movaps %xmm3, %xmm7 1077; SSE-NEXT: subps %xmm6, %xmm7 1078; SSE-NEXT: mulps %xmm5, %xmm7 1079; SSE-NEXT: addps %xmm5, %xmm7 1080; SSE-NEXT: mulps %xmm7, %xmm2 1081; SSE-NEXT: movaps %xmm3, %xmm5 1082; SSE-NEXT: subps %xmm2, %xmm5 1083; SSE-NEXT: mulps %xmm7, %xmm5 1084; SSE-NEXT: addps %xmm7, %xmm5 1085; SSE-NEXT: rcpps %xmm4, %xmm2 1086; SSE-NEXT: movaps %xmm4, %xmm6 1087; SSE-NEXT: mulps %xmm2, %xmm6 1088; SSE-NEXT: movaps %xmm3, %xmm7 1089; SSE-NEXT: subps %xmm6, %xmm7 1090; SSE-NEXT: mulps %xmm2, %xmm7 1091; SSE-NEXT: addps %xmm2, %xmm7 1092; SSE-NEXT: mulps %xmm7, %xmm4 1093; SSE-NEXT: subps %xmm4, %xmm3 1094; SSE-NEXT: mulps %xmm7, %xmm3 1095; SSE-NEXT: addps %xmm7, %xmm3 1096; SSE-NEXT: movaps %xmm5, %xmm2 1097; SSE-NEXT: retq 1098; 1099; AVX-RECIP-LABEL: v16f32_two_step: 1100; AVX-RECIP: # %bb.0: 1101; AVX-RECIP-NEXT: vrcpps %ymm0, %ymm2 1102; AVX-RECIP-NEXT: vmulps %ymm2, %ymm0, %ymm3 1103; AVX-RECIP-NEXT: vmovaps {{.*#+}} ymm4 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 1104; AVX-RECIP-NEXT: vsubps %ymm3, %ymm4, %ymm3 1105; AVX-RECIP-NEXT: vmulps %ymm3, %ymm2, %ymm3 1106; AVX-RECIP-NEXT: vaddps %ymm3, %ymm2, %ymm2 1107; AVX-RECIP-NEXT: vmulps %ymm2, %ymm0, %ymm0 1108; AVX-RECIP-NEXT: vsubps %ymm0, %ymm4, %ymm0 1109; AVX-RECIP-NEXT: vmulps %ymm0, %ymm2, %ymm0 1110; AVX-RECIP-NEXT: vaddps %ymm0, %ymm2, %ymm0 1111; AVX-RECIP-NEXT: vrcpps %ymm1, %ymm2 1112; AVX-RECIP-NEXT: vmulps %ymm2, %ymm1, %ymm3 1113; AVX-RECIP-NEXT: vsubps %ymm3, %ymm4, %ymm3 1114; AVX-RECIP-NEXT: vmulps %ymm3, %ymm2, %ymm3 1115; AVX-RECIP-NEXT: vaddps %ymm3, %ymm2, %ymm2 1116; AVX-RECIP-NEXT: vmulps %ymm2, %ymm1, %ymm1 1117; AVX-RECIP-NEXT: vsubps %ymm1, %ymm4, %ymm1 1118; AVX-RECIP-NEXT: vmulps %ymm1, %ymm2, %ymm1 1119; AVX-RECIP-NEXT: vaddps %ymm1, %ymm2, %ymm1 1120; AVX-RECIP-NEXT: retq 1121; 1122; FMA-RECIP-LABEL: v16f32_two_step: 1123; FMA-RECIP: # %bb.0: 1124; FMA-RECIP-NEXT: vrcpps %ymm0, %ymm2 1125; FMA-RECIP-NEXT: vmovaps {{.*#+}} ymm3 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] 1126; FMA-RECIP-NEXT: vmovaps %ymm2, %ymm4 1127; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} ymm4 = -(ymm0 * ymm4) + ymm3 1128; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} ymm4 = (ymm4 * ymm2) + ymm2 1129; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm4 * ymm0) + ymm3 1130; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm4) + ymm4 1131; FMA-RECIP-NEXT: vrcpps %ymm1, %ymm2 1132; FMA-RECIP-NEXT: vmovaps %ymm2, %ymm4 1133; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} ymm4 = -(ymm1 * ymm4) + ymm3 1134; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} ymm4 = (ymm4 * ymm2) + ymm2 1135; FMA-RECIP-NEXT: vfnmadd213ps {{.*#+}} ymm1 = -(ymm4 * ymm1) + ymm3 1136; FMA-RECIP-NEXT: vfmadd132ps {{.*#+}} ymm1 = (ymm1 * ymm4) + ymm4 1137; FMA-RECIP-NEXT: retq 1138; 1139; BTVER2-LABEL: v16f32_two_step: 1140; BTVER2: # %bb.0: 1141; BTVER2-NEXT: vmovaps {{.*#+}} ymm4 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [5:1.00] 1142; BTVER2-NEXT: vrcpps %ymm0, %ymm2 # sched: [2:2.00] 1143; BTVER2-NEXT: vmulps %ymm2, %ymm0, %ymm3 # sched: [2:2.00] 1144; BTVER2-NEXT: vsubps %ymm3, %ymm4, %ymm3 # sched: [3:2.00] 1145; BTVER2-NEXT: vmulps %ymm3, %ymm2, %ymm3 # sched: [2:2.00] 1146; BTVER2-NEXT: vaddps %ymm3, %ymm2, %ymm2 # sched: [3:2.00] 1147; BTVER2-NEXT: vmulps %ymm2, %ymm0, %ymm0 # sched: [2:2.00] 1148; BTVER2-NEXT: vsubps %ymm0, %ymm4, %ymm0 # sched: [3:2.00] 1149; BTVER2-NEXT: vmulps %ymm0, %ymm2, %ymm0 # sched: [2:2.00] 1150; BTVER2-NEXT: vaddps %ymm0, %ymm2, %ymm0 # sched: [3:2.00] 1151; BTVER2-NEXT: vrcpps %ymm1, %ymm2 # sched: [2:2.00] 1152; BTVER2-NEXT: vmulps %ymm2, %ymm1, %ymm3 # sched: [2:2.00] 1153; BTVER2-NEXT: vsubps %ymm3, %ymm4, %ymm3 # sched: [3:2.00] 1154; BTVER2-NEXT: vmulps %ymm3, %ymm2, %ymm3 # sched: [2:2.00] 1155; BTVER2-NEXT: vaddps %ymm3, %ymm2, %ymm2 # sched: [3:2.00] 1156; BTVER2-NEXT: vmulps %ymm2, %ymm1, %ymm1 # sched: [2:2.00] 1157; BTVER2-NEXT: vsubps %ymm1, %ymm4, %ymm1 # sched: [3:2.00] 1158; BTVER2-NEXT: vmulps %ymm1, %ymm2, %ymm1 # sched: [2:2.00] 1159; BTVER2-NEXT: vaddps %ymm1, %ymm2, %ymm1 # sched: [3:2.00] 1160; BTVER2-NEXT: retq # sched: [4:1.00] 1161; 1162; SANDY-LABEL: v16f32_two_step: 1163; SANDY: # %bb.0: 1164; SANDY-NEXT: vrcpps %ymm0, %ymm2 # sched: [7:2.00] 1165; SANDY-NEXT: vmulps %ymm2, %ymm0, %ymm3 # sched: [5:1.00] 1166; SANDY-NEXT: vmovaps {{.*#+}} ymm4 = [1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00,1.000000e+00] sched: [7:0.50] 1167; SANDY-NEXT: vsubps %ymm3, %ymm4, %ymm3 # sched: [3:1.00] 1168; SANDY-NEXT: vmulps %ymm3, %ymm2, %ymm3 # sched: [5:1.00] 1169; SANDY-NEXT: vaddps %ymm3, %ymm2, %ymm2 # sched: [3:1.00] 1170; SANDY-NEXT: vmulps %ymm2, %ymm0, %ymm0 # sched: [5:1.00] 1171; SANDY-NEXT: vsubps %ymm0, %ymm4, %ymm0 # sched: [3:1.00] 1172; SANDY-NEXT: vmulps %ymm0, %ymm2, %ymm0 # sched: [5:1.00] 1173; SANDY-NEXT: vaddps %ymm0, %ymm2, %ymm0 # sched: [3:1.00] 1174; SANDY-NEXT: vrcpps %ymm1, %ymm2 # sched: [7:2.00] 1175; SANDY-NEXT: vmulps %ymm2, %ymm1, %ymm3 # sched: [5:1.00] 1176; SANDY-NEXT: vsubps %ymm3, %ymm4, %ymm3 # sched: [3:1.00] 1177; SANDY-NEXT: vmulps %ymm3, %ymm2, %ymm3 # sched: [5:1.00] 1178; SANDY-NEXT: vaddps %ymm3, %ymm2, %ymm2 # sched: [3:1.00] 1179; SANDY-NEXT: vmulps %ymm2, %ymm1, %ymm1 # sched: [5:1.00] 1180; SANDY-NEXT: vsubps %ymm1, %ymm4, %ymm1 # sched: [3:1.00] 1181; SANDY-NEXT: vmulps %ymm1, %ymm2, %ymm1 # sched: [5:1.00] 1182; SANDY-NEXT: vaddps %ymm1, %ymm2, %ymm1 # sched: [3:1.00] 1183; SANDY-NEXT: retq # sched: [1:1.00] 1184; 1185; HASWELL-LABEL: v16f32_two_step: 1186; HASWELL: # %bb.0: 1187; HASWELL-NEXT: vrcpps %ymm0, %ymm2 # sched: [11:2.00] 1188; HASWELL-NEXT: vbroadcastss {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1] sched: [7:0.50] 1189; HASWELL-NEXT: vmovaps %ymm2, %ymm4 # sched: [1:1.00] 1190; HASWELL-NEXT: vfnmadd213ps {{.*#+}} ymm4 = -(ymm0 * ymm4) + ymm3 sched: [5:0.50] 1191; HASWELL-NEXT: vfmadd132ps {{.*#+}} ymm4 = (ymm4 * ymm2) + ymm2 sched: [5:0.50] 1192; HASWELL-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm4 * ymm0) + ymm3 sched: [5:0.50] 1193; HASWELL-NEXT: vfmadd132ps {{.*#+}} ymm0 = (ymm0 * ymm4) + ymm4 sched: [5:0.50] 1194; HASWELL-NEXT: vrcpps %ymm1, %ymm2 # sched: [11:2.00] 1195; HASWELL-NEXT: vmovaps %ymm2, %ymm4 # sched: [1:1.00] 1196; HASWELL-NEXT: vfnmadd213ps {{.*#+}} ymm4 = -(ymm1 * ymm4) + ymm3 sched: [5:0.50] 1197; HASWELL-NEXT: vfmadd132ps {{.*#+}} ymm4 = (ymm4 * ymm2) + ymm2 sched: [5:0.50] 1198; HASWELL-NEXT: vfnmadd213ps {{.*#+}} ymm1 = -(ymm4 * ymm1) + ymm3 sched: [5:0.50] 1199; HASWELL-NEXT: vfmadd132ps {{.*#+}} ymm1 = (ymm1 * ymm4) + ymm4 sched: [5:0.50] 1200; HASWELL-NEXT: retq # sched: [7:1.00] 1201; 1202; HASWELL-NO-FMA-LABEL: v16f32_two_step: 1203; HASWELL-NO-FMA: # %bb.0: 1204; HASWELL-NO-FMA-NEXT: vrcpps %ymm0, %ymm2 1205; HASWELL-NO-FMA-NEXT: vmulps %ymm2, %ymm0, %ymm3 1206; HASWELL-NO-FMA-NEXT: vbroadcastss {{.*#+}} ymm4 = [1,1,1,1,1,1,1,1] 1207; HASWELL-NO-FMA-NEXT: vsubps %ymm3, %ymm4, %ymm3 1208; HASWELL-NO-FMA-NEXT: vmulps %ymm3, %ymm2, %ymm3 1209; HASWELL-NO-FMA-NEXT: vaddps %ymm3, %ymm2, %ymm2 1210; HASWELL-NO-FMA-NEXT: vmulps %ymm2, %ymm0, %ymm0 1211; HASWELL-NO-FMA-NEXT: vsubps %ymm0, %ymm4, %ymm0 1212; HASWELL-NO-FMA-NEXT: vmulps %ymm0, %ymm2, %ymm0 1213; HASWELL-NO-FMA-NEXT: vaddps %ymm0, %ymm2, %ymm0 1214; HASWELL-NO-FMA-NEXT: vrcpps %ymm1, %ymm2 1215; HASWELL-NO-FMA-NEXT: vmulps %ymm2, %ymm1, %ymm3 1216; HASWELL-NO-FMA-NEXT: vsubps %ymm3, %ymm4, %ymm3 1217; HASWELL-NO-FMA-NEXT: vmulps %ymm3, %ymm2, %ymm3 1218; HASWELL-NO-FMA-NEXT: vaddps %ymm3, %ymm2, %ymm2 1219; HASWELL-NO-FMA-NEXT: vmulps %ymm2, %ymm1, %ymm1 1220; HASWELL-NO-FMA-NEXT: vsubps %ymm1, %ymm4, %ymm1 1221; HASWELL-NO-FMA-NEXT: vmulps %ymm1, %ymm2, %ymm1 1222; HASWELL-NO-FMA-NEXT: vaddps %ymm1, %ymm2, %ymm1 1223; HASWELL-NO-FMA-NEXT: retq 1224; 1225; KNL-LABEL: v16f32_two_step: 1226; KNL: # %bb.0: 1227; KNL-NEXT: vrcp14ps %zmm0, %zmm1 # sched: [11:2.00] 1228; KNL-NEXT: vbroadcastss {{.*#+}} zmm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] sched: [10:1.00] 1229; KNL-NEXT: vmovaps %zmm1, %zmm3 # sched: [1:1.00] 1230; KNL-NEXT: vfnmadd213ps {{.*#+}} zmm3 = -(zmm0 * zmm3) + zmm2 sched: [5:0.50] 1231; KNL-NEXT: vfmadd132ps {{.*#+}} zmm3 = (zmm3 * zmm1) + zmm1 sched: [5:0.50] 1232; KNL-NEXT: vfnmadd213ps {{.*#+}} zmm0 = -(zmm3 * zmm0) + zmm2 sched: [5:0.50] 1233; KNL-NEXT: vfmadd132ps {{.*#+}} zmm0 = (zmm0 * zmm3) + zmm3 sched: [5:0.50] 1234; KNL-NEXT: retq # sched: [7:1.00] 1235; 1236; SKX-LABEL: v16f32_two_step: 1237; SKX: # %bb.0: 1238; SKX-NEXT: vrcp14ps %zmm0, %zmm1 # sched: [4:2.00] 1239; SKX-NEXT: vbroadcastss {{.*#+}} zmm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1] sched: [8:0.50] 1240; SKX-NEXT: vmovaps %zmm1, %zmm3 # sched: [1:0.33] 1241; SKX-NEXT: vfnmadd213ps {{.*#+}} zmm3 = -(zmm0 * zmm3) + zmm2 sched: [4:0.50] 1242; SKX-NEXT: vfmadd132ps {{.*#+}} zmm3 = (zmm3 * zmm1) + zmm1 sched: [4:0.50] 1243; SKX-NEXT: vfnmadd213ps {{.*#+}} zmm0 = -(zmm3 * zmm0) + zmm2 sched: [4:0.50] 1244; SKX-NEXT: vfmadd132ps {{.*#+}} zmm0 = (zmm0 * zmm3) + zmm3 sched: [4:0.50] 1245; SKX-NEXT: retq # sched: [7:1.00] 1246 %div = fdiv fast <16 x float> <float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0, float 1.0>, %x 1247 ret <16 x float> %div 1248} 1249 1250attributes #0 = { "unsafe-fp-math"="true" "reciprocal-estimates"="!divf,!vec-divf" } 1251attributes #1 = { "unsafe-fp-math"="true" "reciprocal-estimates"="divf,vec-divf" } 1252attributes #2 = { "unsafe-fp-math"="true" "reciprocal-estimates"="divf:2,vec-divf:2" } 1253 1254