diff options
| author | Craig Topper <craig.topper@gmail.com> | 2012-09-06 05:15:01 +0000 |
|---|---|---|
| committer | Craig Topper <craig.topper@gmail.com> | 2012-09-06 05:15:01 +0000 |
| commit | 07149fe7150043dbc62893cbc1e1733b7eb210da (patch) | |
| tree | d5d86f4f663f902be94ce80f70df8dbca0c94fc0 /lib/Target | |
| parent | ae6a2e224802778dd07a8ed8c2513d5ad6eece1f (diff) | |
| download | external_llvm-07149fe7150043dbc62893cbc1e1733b7eb210da.zip external_llvm-07149fe7150043dbc62893cbc1e1733b7eb210da.tar.gz external_llvm-07149fe7150043dbc62893cbc1e1733b7eb210da.tar.bz2 | |
Add patterns for converting stores of subvector_extracts of lower 128-bits of a 256-bit vector to VMOVAPSmr/VMOVUPSmr.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@163292 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/Target')
| -rw-r--r-- | lib/Target/X86/X86InstrSSE.td | 40 |
1 files changed, 40 insertions, 0 deletions
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td index 959ae36..1276bda 100644 --- a/lib/Target/X86/X86InstrSSE.td +++ b/lib/Target/X86/X86InstrSSE.td @@ -1019,6 +1019,46 @@ let Predicates = [HasAVX] in { (VMOVUPSYmr addr:$dst, VR256:$src)>; def : Pat<(store (v32i8 VR256:$src), addr:$dst), (VMOVUPSYmr addr:$dst, VR256:$src)>; + + // Special patterns for storing subvector extracts of lower 128-bits + // Its cheaper to just use VMOVAPS/VMOVUPS instead of VEXTRACTF128mr + def : Pat<(alignedstore (v2f64 (extract_subvector + (v4f64 VR256:$src), (i32 0))), addr:$dst), + (VMOVAPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(alignedstore (v4f32 (extract_subvector + (v8f32 VR256:$src), (i32 0))), addr:$dst), + (VMOVAPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(alignedstore (v2i64 (extract_subvector + (v4i64 VR256:$src), (i32 0))), addr:$dst), + (VMOVAPDmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(alignedstore (v4i32 (extract_subvector + (v8i32 VR256:$src), (i32 0))), addr:$dst), + (VMOVAPSmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(alignedstore (v8i16 (extract_subvector + (v16i16 VR256:$src), (i32 0))), addr:$dst), + (VMOVAPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(alignedstore (v16i8 (extract_subvector + (v32i8 VR256:$src), (i32 0))), addr:$dst), + (VMOVAPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + + def : Pat<(store (v2f64 (extract_subvector + (v4f64 VR256:$src), (i32 0))), addr:$dst), + (VMOVUPDmr addr:$dst, (v2f64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(store (v4f32 (extract_subvector + (v8f32 VR256:$src), (i32 0))), addr:$dst), + (VMOVUPSmr addr:$dst, (v4f32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(store (v2i64 (extract_subvector + (v4i64 VR256:$src), (i32 0))), addr:$dst), + (VMOVUPDmr addr:$dst, (v2i64 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(store (v4i32 (extract_subvector + (v8i32 VR256:$src), (i32 0))), addr:$dst), + (VMOVUPSmr addr:$dst, (v4i32 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(store (v8i16 (extract_subvector + (v16i16 VR256:$src), (i32 0))), addr:$dst), + (VMOVAPSmr addr:$dst, (v8i16 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; + def : Pat<(store (v16i8 (extract_subvector + (v32i8 VR256:$src), (i32 0))), addr:$dst), + (VMOVUPSmr addr:$dst, (v16i8 (EXTRACT_SUBREG VR256:$src,sub_xmm)))>; } // Use movaps / movups for SSE integer load / store (one byte shorter). |