1; RUN: llc < %s -mtriple=i686-apple-darwin -mcpu=knl | FileCheck -check-prefix=X32 %s 2; RUN: llc < %s -mtriple=i386-pc-win32 -mcpu=knl | FileCheck -check-prefix=X32 %s 3; RUN: llc < %s -mtriple=x86_64-win32 -mcpu=knl | FileCheck -check-prefix=WIN64 %s 4; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl | FileCheck -check-prefix=X64 %s 5 6declare <16 x float> @func_float16_ptr(<16 x float>, <16 x float> *) 7declare <16 x float> @func_float16(<16 x float>, <16 x float>) 8declare i32 @func_int(i32, i32) 9 10; WIN64-LABEL: testf16_inp 11; WIN64: vaddps {{.*}}, {{%zmm[0-1]}} 12; WIN64: leaq {{.*}}(%rsp), %rcx 13; WIN64: call 14; WIN64: ret 15 16; X32-LABEL: testf16_inp 17; X32: vaddps {{.*}}, {{%zmm[0-1]}} 18; Push is not deemed profitable if we're realigning the stack. 19; X32: {{pushl|movl}} %eax 20; X32: call 21; X32: ret 22 23; X64-LABEL: testf16_inp 24; X64: vaddps {{.*}}, {{%zmm[0-1]}} 25; X64: leaq {{.*}}(%rsp), %rdi 26; X64: call 27; X64: ret 28 29;test calling conventions - input parameters 30define <16 x float> @testf16_inp(<16 x float> %a, <16 x float> %b) nounwind { 31 %y = alloca <16 x float>, align 16 32 %x = fadd <16 x float> %a, %b 33 %1 = call intel_ocl_bicc <16 x float> @func_float16_ptr(<16 x float> %x, <16 x float>* %y) 34 %2 = load <16 x float>, <16 x float>* %y, align 16 35 %3 = fadd <16 x float> %2, %1 36 ret <16 x float> %3 37} 38 39;test calling conventions - preserved registers 40 41; preserved zmm16- 42; WIN64-LABEL: testf16_regs 43; WIN64: call 44; WIN64: vaddps %zmm16, %zmm0, %zmm0 45; WIN64: ret 46 47; preserved zmm16- 48; X64-LABEL: testf16_regs 49; X64: call 50; X64: vaddps %zmm16, %zmm0, %zmm0 51; X64: ret 52 53define <16 x float> @testf16_regs(<16 x float> %a, <16 x float> %b) nounwind { 54 %y = alloca <16 x float>, align 16 55 %x = fadd <16 x float> %a, %b 56 %1 = call intel_ocl_bicc <16 x float> @func_float16_ptr(<16 x float> %x, <16 x float>* %y) 57 %2 = load <16 x float>, <16 x float>* %y, align 16 58 %3 = fadd <16 x float> %1, %b 59 %4 = fadd <16 x float> %2, %3 60 ret <16 x float> %4 61} 62 63; test calling conventions - prolog and epilog 64; WIN64-LABEL: test_prolog_epilog 65; WIN64: vmovups %zmm21, {{.*(%rbp).*}} # 64-byte Spill 66; WIN64: vmovups %zmm6, {{.*(%rbp).*}} # 64-byte Spill 67; WIN64: call 68; WIN64: vmovups {{.*(%rbp).*}}, %zmm6 # 64-byte Reload 69; WIN64: vmovups {{.*(%rbp).*}}, %zmm21 # 64-byte Reload 70 71; X64-LABEL: test_prolog_epilog 72; X64: kmovq %k7, {{.*}}(%rsp) ## 8-byte Folded Spill 73; X64: kmovq %k6, {{.*}}(%rsp) ## 8-byte Folded Spill 74; X64: kmovq %k5, {{.*}}(%rsp) ## 8-byte Folded Spill 75; X64: kmovq %k4, {{.*}}(%rsp) ## 8-byte Folded Spill 76; X64: vmovups %zmm31, {{.*}}(%rsp) ## 64-byte Spill 77; X64: vmovups %zmm16, {{.*}}(%rsp) ## 64-byte Spill 78; X64: call 79; X64: vmovups {{.*}}(%rsp), %zmm16 ## 64-byte Reload 80; X64: vmovups {{.*}}(%rsp), %zmm31 ## 64-byte Reload 81define intel_ocl_bicc <16 x float> @test_prolog_epilog(<16 x float> %a, <16 x float> %b) nounwind { 82 %c = call <16 x float> @func_float16(<16 x float> %a, <16 x float> %b) 83 ret <16 x float> %c 84} 85 86 87declare <16 x float> @func_float16_mask(<16 x float>, <16 x i1>) 88 89; X64-LABEL: testf16_inp_mask 90; X64: kmovw %edi, %k1 91; X64: call 92define <16 x float> @testf16_inp_mask(<16 x float> %a, i16 %mask) { 93 %imask = bitcast i16 %mask to <16 x i1> 94 %1 = call intel_ocl_bicc <16 x float> @func_float16_mask(<16 x float> %a, <16 x i1> %imask) 95 ret <16 x float> %1 96} 97 98; X64-LABEL: test_prolog_epilog_with_mask 99; X64: kxorw %k{{.*}}, %k{{.*}}, %k1 100; X64: call 101define intel_ocl_bicc <16 x float> @test_prolog_epilog_with_mask(<16 x float> %a, <16 x i32> %x1, <16 x i32>%x2, <16 x i1> %mask) nounwind { 102 %cmp_res = icmp eq <16 x i32>%x1, %x2 103 %mask1 = xor <16 x i1> %cmp_res, %mask 104 %c = call intel_ocl_bicc <16 x float> @func_float16_mask(<16 x float> %a, <16 x i1>%mask1) 105 ret <16 x float> %c 106} 107