maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

Move all PSHUF* patterns close to the PSHUF* definitions. Also be 

explicit about which subtarget they refer to, and add AVX versions of
the ones we currently don't. Remove old and now wrong comments!

llvm-svn: 138515
This commit is contained in:
Bruno Cardoso Lopes 2011-08-24 23:17:57 +00:00
parent 2953d7b320
commit 9566a66a7c
1 changed files with 56 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

103
llvm/lib/Target/X86/X86InstrSSE.td
View File

@ -2986,6 +2986,34 @@ let Predicates = [HasAVX] in {
// SSE2 with ImmT == Imm8 and XD prefix. // SSE2 with ImmT == Imm8 and XD prefix.
defm VPSHUFLW : sse2_pshuffle<"vpshuflw", v8i16, pshuflw, bc_v8i16>, XD, defm VPSHUFLW : sse2_pshuffle<"vpshuflw", v8i16, pshuflw, bc_v8i16>, XD,
VEX; VEX;
let AddedComplexity = 5 in
def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
(VPSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>;
// Unary v4f32 shuffle with VPSHUF* in order to fold a load.
def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
(VPSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>;
def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv2i64 addr:$src1)),
(i8 imm:$imm))),
(VPSHUFDmi addr:$src1, imm:$imm)>, Requires<[HasAVX]>;
def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv4f32 addr:$src1)),
(i8 imm:$imm))),
(VPSHUFDmi addr:$src1, imm:$imm)>;
def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
(VPSHUFDri VR128:$src1, imm:$imm)>, Requires<[HasAVX]>;
def : Pat<(v4i32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
(VPSHUFDri VR128:$src1, imm:$imm)>, Requires<[HasAVX]>;
def : Pat<(v8i16 (X86PShufhw VR128:$src, (i8 imm:$imm))),
(VPSHUFHWri VR128:$src, imm:$imm)>;
def : Pat<(v8i16 (X86PShufhw (bc_v8i16 (memopv2i64 addr:$src)),
(i8 imm:$imm))),
(VPSHUFHWmi addr:$src, imm:$imm)>;
def : Pat<(v8i16 (X86PShuflw VR128:$src, (i8 imm:$imm))),
(VPSHUFLWri VR128:$src, imm:$imm)>;
def : Pat<(v8i16 (X86PShuflw (bc_v8i16 (memopv2i64 addr:$src)),
(i8 imm:$imm))),
(VPSHUFLWmi addr:$src, imm:$imm)>;
} }
let Predicates = [HasSSE2] in { let Predicates = [HasSSE2] in {
@ -2997,6 +3025,34 @@ let Predicates = [HasSSE2] in {
// SSE2 with ImmT == Imm8 and XD prefix. // SSE2 with ImmT == Imm8 and XD prefix.
defm PSHUFLW : sse2_pshuffle<"pshuflw", v8i16, pshuflw, bc_v8i16>, XD; defm PSHUFLW : sse2_pshuffle<"pshuflw", v8i16, pshuflw, bc_v8i16>, XD;
let AddedComplexity = 5 in
def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
(PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>;
// Unary v4f32 shuffle with PSHUF* in order to fold a load.
def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
(PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>;
def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv2i64 addr:$src1)),
(i8 imm:$imm))),
(PSHUFDmi addr:$src1, imm:$imm)>;
def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv4f32 addr:$src1)),
(i8 imm:$imm))),
(PSHUFDmi addr:$src1, imm:$imm)>;
def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
(PSHUFDri VR128:$src1, imm:$imm)>;
def : Pat<(v4i32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
(PSHUFDri VR128:$src1, imm:$imm)>;
def : Pat<(v8i16 (X86PShufhw VR128:$src, (i8 imm:$imm))),
(PSHUFHWri VR128:$src, imm:$imm)>;
def : Pat<(v8i16 (X86PShufhw (bc_v8i16 (memopv2i64 addr:$src)),
(i8 imm:$imm))),
(PSHUFHWmi addr:$src, imm:$imm)>;
def : Pat<(v8i16 (X86PShuflw VR128:$src, (i8 imm:$imm))),
(PSHUFLWri VR128:$src, imm:$imm)>;
def : Pat<(v8i16 (X86PShuflw (bc_v8i16 (memopv2i64 addr:$src)),
(i8 imm:$imm))),
(PSHUFLWmi addr:$src, imm:$imm)>;
} }
//===---------------------------------------------------------------------===// //===---------------------------------------------------------------------===//
@ -4150,15 +4206,6 @@ def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
(PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>; (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
} }
let AddedComplexity = 5 in
def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
(PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
Requires<[HasSSE2]>;
// Unary v4f32 shuffle with PSHUF* in order to fold a load.
def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
(PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
Requires<[HasSSE2]>;
let AddedComplexity = 20 in { let AddedComplexity = 20 in {
// vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))), def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
@ -5964,32 +6011,6 @@ def VZEROUPPER : I<0x77, RawFrm, (outs), (ins), "vzeroupper",
// The AVX version of some but not all of them are described here, and more // The AVX version of some but not all of them are described here, and more
// should come in a near future. // should come in a near future.
// Shuffle with PSHUFD instruction folding loads. The first two patterns match
// SSE2 loads, which are always promoted to v2i64. The last one should match
// the SSE1 case, where the only legal load is v4f32, but there is no PSHUFD
// in SSE2, how does it ever worked? Anyway, the pattern will remain here until
// we investigate further.
def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv2i64 addr:$src1)),
(i8 imm:$imm))),
(VPSHUFDmi addr:$src1, imm:$imm)>, Requires<[HasAVX]>;
def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv2i64 addr:$src1)),
(i8 imm:$imm))),
(PSHUFDmi addr:$src1, imm:$imm)>;
def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv4f32 addr:$src1)),
(i8 imm:$imm))),
(PSHUFDmi addr:$src1, imm:$imm)>; // FIXME: has this ever worked?
// Shuffle with PSHUFD instruction.
def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
(VPSHUFDri VR128:$src1, imm:$imm)>, Requires<[HasAVX]>;
def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
(PSHUFDri VR128:$src1, imm:$imm)>;
def : Pat<(v4i32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
(VPSHUFDri VR128:$src1, imm:$imm)>, Requires<[HasAVX]>;
def : Pat<(v4i32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
(PSHUFDri VR128:$src1, imm:$imm)>;
// Shuffle with MOVHLPS instruction // Shuffle with MOVHLPS instruction
def : Pat<(v4f32 (X86Movhlps VR128:$src1, VR128:$src2)), def : Pat<(v4f32 (X86Movhlps VR128:$src1, VR128:$src2)),
(MOVHLPSrr VR128:$src1, VR128:$src2)>; (MOVHLPSrr VR128:$src1, VR128:$src2)>;
@ -6155,18 +6176,6 @@ def : Pat<(v4f32 (X86Movsd VR128:$src1, VR128:$src2)),
def : Pat<(v4i32 (X86Movsd VR128:$src1, VR128:$src2)), def : Pat<(v4i32 (X86Movsd VR128:$src1, VR128:$src2)),
(MOVSDrr VR128:$src1, (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_sd))>; (MOVSDrr VR128:$src1, (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_sd))>;
// Shuffle with PSHUFHW
def : Pat<(v8i16 (X86PShufhw VR128:$src, (i8 imm:$imm))),
(PSHUFHWri VR128:$src, imm:$imm)>;
def : Pat<(v8i16 (X86PShufhw (bc_v8i16 (memopv2i64 addr:$src)), (i8 imm:$imm))),
(PSHUFHWmi addr:$src, imm:$imm)>;
// Shuffle with PSHUFLW
def : Pat<(v8i16 (X86PShuflw VR128:$src, (i8 imm:$imm))),
(PSHUFLWri VR128:$src, imm:$imm)>;
def : Pat<(v8i16 (X86PShuflw (bc_v8i16 (memopv2i64 addr:$src)), (i8 imm:$imm))),
(PSHUFLWmi addr:$src, imm:$imm)>;
// Shuffle with MOVLPS // Shuffle with MOVLPS
def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))), def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
(MOVLPSrm VR128:$src1, addr:$src2)>; (MOVLPSrm VR128:$src1, addr:$src2)>;