88 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=X863; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=X644 5%v4_varying_complex = type { <4 x float>, <4 x float> }6 7define void @FFT(ptr noalias nocapture %destination, ptr noalias %re, ptr noalias nocapture %ptr_cast_for_load) nounwind {8; X86-LABEL: FFT:9; X86: # %bb.0: # %begin10; X86-NEXT: pushl %ebx11; X86-NEXT: pushl %edi12; X86-NEXT: pushl %esi13; X86-NEXT: movl {{[0-9]+}}(%esp), %eax14; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx15; X86-NEXT: movl {{[0-9]+}}(%esp), %edx16; X86-NEXT: movdqu (%edx), %xmm017; X86-NEXT: pslld $4, %xmm018; X86-NEXT: movd %xmm0, %edx19; X86-NEXT: pextrd $1, %xmm0, %esi20; X86-NEXT: pextrd $2, %xmm0, %edi21; X86-NEXT: pextrd $3, %xmm0, %ebx22; X86-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero23; X86-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero24; X86-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero25; X86-NEXT: movss {{.*#+}} xmm3 = mem[0],zero,zero,zero26; X86-NEXT: movss %xmm0, 128(%eax)27; X86-NEXT: movss %xmm1, 164(%eax)28; X86-NEXT: movss %xmm2, 200(%eax)29; X86-NEXT: movss %xmm3, 236(%eax)30; X86-NEXT: popl %esi31; X86-NEXT: popl %edi32; X86-NEXT: popl %ebx33; X86-NEXT: retl34;35; X64-LABEL: FFT:36; X64: # %bb.0: # %begin37; X64-NEXT: movdqu (%rdx), %xmm038; X64-NEXT: pslld $4, %xmm039; X64-NEXT: movd %xmm0, %eax40; X64-NEXT: cltq41; X64-NEXT: pextrd $1, %xmm0, %ecx42; X64-NEXT: movslq %ecx, %rcx43; X64-NEXT: pextrd $2, %xmm0, %edx44; X64-NEXT: movslq %edx, %rdx45; X64-NEXT: pextrd $3, %xmm0, %r8d46; X64-NEXT: movslq %r8d, %r847; X64-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero48; X64-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero49; X64-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero50; X64-NEXT: movss {{.*#+}} xmm3 = mem[0],zero,zero,zero51; X64-NEXT: movss %xmm0, 128(%rdi)52; X64-NEXT: movss %xmm1, 164(%rdi)53; X64-NEXT: movss %xmm2, 200(%rdi)54; X64-NEXT: movss %xmm3, 236(%rdi)55; X64-NEXT: retq56begin:57 %ptr_masked_load79 = load <4 x i32>, ptr %ptr_cast_for_load, align 458 %mul__bitReversedProgramIndex_load = shl <4 x i32> %ptr_masked_load79, <i32 4, i32 4, i32 4, i32 4>59 60 %offset32_1 = extractelement <4 x i32> %mul__bitReversedProgramIndex_load, i32 061 %ptroffset_1 = sext i32 %offset32_1 to i6462 %offset32_2 = extractelement <4 x i32> %mul__bitReversedProgramIndex_load, i32 163 %ptroffset_2 = sext i32 %offset32_2 to i6464 %offset32_3 = extractelement <4 x i32> %mul__bitReversedProgramIndex_load, i32 265 %ptroffset_3 = sext i32 %offset32_3 to i6466 %offset32_4 = extractelement <4 x i32> %mul__bitReversedProgramIndex_load, i32 367 %ptroffset_4 = sext i32 %offset32_4 to i6468 69 %ptrcast_1 = getelementptr float, ptr %re, i64 %ptroffset_170 %val_1 = load float, ptr %ptrcast_1, align 471 %ptrcast_2 = getelementptr float, ptr %re, i64 %ptroffset_272 %val_2 = load float, ptr %ptrcast_2, align 473 %ptrcast_3 = getelementptr float, ptr %re, i64 %ptroffset_374 %val_3 = load float, ptr %ptrcast_3, align 475 %ptrcast_4 = getelementptr float, ptr %re, i64 %ptroffset_476 %val_4 = load float, ptr %ptrcast_4, align 477 78 %ptrcast1_1 = getelementptr inbounds %v4_varying_complex, ptr %destination, i64 4, i32 0, i64 079 store float %val_1, ptr %ptrcast1_1, align 480 %finalptr_2 = getelementptr i8, ptr %destination, i64 16481 store float %val_2, ptr %finalptr_2, align 482 %finalptr_3 = getelementptr i8, ptr %destination, i64 20083 store float %val_3, ptr %finalptr_3, align 484 %finalptr_4 = getelementptr i8, ptr %destination, i64 23685 store float %val_4, ptr %finalptr_4, align 486 ret void87}88