1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2 ; RUN: llc < %s -mcpu=skx | FileCheck %s --check-prefixes=CHECK,CHECK-SKX 3 ; RUN: llc < %s -mcpu=knl | FileCheck %s --check-prefixes=CHECK,CHECK-KNL 4 5 target triple = "x86_64-unknown-unknown" 6 7 define <8 x i64> @test1(<8 x i64> %m, <8 x i64> %a, <8 x i64> %b) { 8 ; CHECK-SKX-LABEL: test1: 9 ; CHECK-SKX: # %bb.0: # %entry 10 ; CHECK-SKX-NEXT: vpsllq $63, %zmm0, %zmm0 11 ; CHECK-SKX-NEXT: vpmovq2m %zmm0, %k1 12 ; CHECK-SKX-NEXT: vpblendmq %zmm1, %zmm2, %zmm0 {%k1} 13 ; CHECK-SKX-NEXT: retq 14 ; 15 ; CHECK-KNL-LABEL: test1: 16 ; CHECK-KNL: # %bb.0: # %entry 17 ; CHECK-KNL-NEXT: vpsllq $63, %zmm0, %zmm0 18 ; CHECK-KNL-NEXT: vptestmq %zmm0, %zmm0, %k1 19 ; CHECK-KNL-NEXT: vpblendmq %zmm1, %zmm2, %zmm0 {%k1} 20 ; CHECK-KNL-NEXT: retq 21 entry: 22 %m.trunc = trunc <8 x i64> %m to <8 x i1> 23 %ret = select <8 x i1> %m.trunc, <8 x i64> %a, <8 x i64> %b 24 ret <8 x i64> %ret 25 } 26 27 ; This is a very contrived test case to trick the legalizer into splitting the 28 ; v16i1 masks in the select during type legalization, and in so doing extend them 29 ; into two v8i64 types. This lets us ensure that the lowering code can handle 30 ; both formulations of vselect. All of this trickery is because we can't 31 ; directly form an SDAG input to the lowering. 32 define <16 x double> @test2(<16 x float> %x, <16 x float> %y, <16 x double> %a, <16 x double> %b) { 33 ; CHECK-LABEL: test2: 34 ; CHECK: # %bb.0: # %entry 35 ; CHECK-NEXT: vxorps %xmm6, %xmm6, %xmm6 36 ; CHECK-NEXT: vcmpltps %zmm0, %zmm6, %k0 37 ; CHECK-NEXT: vcmpltps %zmm6, %zmm1, %k1 38 ; CHECK-NEXT: korw %k1, %k0, %k1 39 ; CHECK-NEXT: vblendmpd %zmm2, %zmm4, %zmm0 {%k1} 40 ; CHECK-NEXT: kshiftrw $8, %k1, %k1 41 ; CHECK-NEXT: vblendmpd %zmm3, %zmm5, %zmm1 {%k1} 42 ; CHECK-NEXT: retq 43 entry: 44 %gt.m = fcmp ogt <16 x float> %x, zeroinitializer 45 %lt.m = fcmp olt <16 x float> %y, zeroinitializer 46 %m.or = or <16 x i1> %gt.m, %lt.m 47 %ret = select <16 x i1> %m.or, <16 x double> %a, <16 x double> %b 48 ret <16 x double> %ret 49 } 50
