94 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=X863; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=X644 5; truncate v2i64 to v2i326 7define void @convert_v2i64_to_v2i32(ptr %dst.addr, <2 x i64> %src) nounwind {8; X86-LABEL: convert_v2i64_to_v2i32:9; X86: # %bb.0: # %entry10; X86-NEXT: movl {{[0-9]+}}(%esp), %eax11; X86-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]12; X86-NEXT: pcmpeqd %xmm1, %xmm113; X86-NEXT: psubd %xmm1, %xmm014; X86-NEXT: movq %xmm0, (%eax)15; X86-NEXT: retl16;17; X64-LABEL: convert_v2i64_to_v2i32:18; X64: # %bb.0: # %entry19; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]20; X64-NEXT: pcmpeqd %xmm1, %xmm121; X64-NEXT: psubd %xmm1, %xmm022; X64-NEXT: movq %xmm0, (%rdi)23; X64-NEXT: retq24entry:25 %val = trunc <2 x i64> %src to <2 x i32>26 %add = add <2 x i32> %val, < i32 1, i32 1 >27 store <2 x i32> %add, ptr %dst.addr28 ret void29}30 31; truncate v3i32 to v3i832 33define void @convert_v3i32_to_v3i8(ptr %dst.addr, ptr %src.addr) nounwind {34; X86-LABEL: convert_v3i32_to_v3i8:35; X86: # %bb.0: # %entry36; X86-NEXT: movl {{[0-9]+}}(%esp), %eax37; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx38; X86-NEXT: movdqa (%ecx), %xmm039; X86-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]40; X86-NEXT: pcmpeqd %xmm1, %xmm141; X86-NEXT: psubb %xmm1, %xmm042; X86-NEXT: pextrb $2, %xmm0, 2(%eax)43; X86-NEXT: pextrw $0, %xmm0, (%eax)44; X86-NEXT: retl45;46; X64-LABEL: convert_v3i32_to_v3i8:47; X64: # %bb.0: # %entry48; X64-NEXT: movdqa (%rsi), %xmm049; X64-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u]50; X64-NEXT: pcmpeqd %xmm1, %xmm151; X64-NEXT: psubb %xmm1, %xmm052; X64-NEXT: pextrb $2, %xmm0, 2(%rdi)53; X64-NEXT: pextrw $0, %xmm0, (%rdi)54; X64-NEXT: retq55entry:56 %load = load <3 x i32>, ptr %src.addr57 %val = trunc <3 x i32> %load to <3 x i8>58 %add = add <3 x i8> %val, < i8 1, i8 1, i8 1 >59 store <3 x i8> %add, ptr %dst.addr60 ret void61}62 63; truncate v5i16 to v5i864 65define void @convert_v5i16_to_v5i8(ptr %dst.addr, ptr %src.addr) nounwind {66; X86-LABEL: convert_v5i16_to_v5i8:67; X86: # %bb.0: # %entry68; X86-NEXT: movl {{[0-9]+}}(%esp), %eax69; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx70; X86-NEXT: movdqa (%ecx), %xmm071; X86-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u]72; X86-NEXT: pcmpeqd %xmm1, %xmm173; X86-NEXT: psubb %xmm1, %xmm074; X86-NEXT: pextrb $4, %xmm0, 4(%eax)75; X86-NEXT: movd %xmm0, (%eax)76; X86-NEXT: retl77;78; X64-LABEL: convert_v5i16_to_v5i8:79; X64: # %bb.0: # %entry80; X64-NEXT: movdqa (%rsi), %xmm081; X64-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u]82; X64-NEXT: pcmpeqd %xmm1, %xmm183; X64-NEXT: psubb %xmm1, %xmm084; X64-NEXT: pextrb $4, %xmm0, 4(%rdi)85; X64-NEXT: movd %xmm0, (%rdi)86; X64-NEXT: retq87entry:88 %load = load <5 x i16>, ptr %src.addr89 %val = trunc <5 x i16> %load to <5 x i8>90 %add = add <5 x i8> %val, < i8 1, i8 1, i8 1, i8 1, i8 1 >91 store <5 x i8> %add, ptr %dst.addr92 ret void93}94