AVX-512: added extend and truncate instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189580 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 117 insertions, 0 deletions

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 95b0de41fa..6b2f1608ca 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -2805,6 +2805,123 @@ def : Pat<(v8f64 (frint VR512:$src)),
 def : Pat<(v8f64 (ftrunc VR512:$src)),
           (VRNDSCALEZPDr VR512:$src, (i32 0x3))>;
 
+//-------------------------------------------------
+// Integer truncate and extend operations
+//-------------------------------------------------
+
+multiclass avx512_trunc_sat<bits<8> opc, string OpcodeStr,
+                          RegisterClass dstRC, RegisterClass srcRC,
+                          RegisterClass KRC, X86MemOperand x86memop> {
+  def rr : AVX512XS8I<opc, MRMDestReg, (outs dstRC:$dst),
+               (ins srcRC:$src),
+               !strconcat(OpcodeStr,"\t{$src, $dst|$dst, $src}"),
+               []>, EVEX;
+
+  def krr : AVX512XS8I<opc, MRMDestReg, (outs dstRC:$dst),
+               (ins KRC:$mask, srcRC:$src),
+               !strconcat(OpcodeStr,
+                 "\t{$src, ${dst}{${mask}}{z}|${dst}{${mask}}{z}, $src}"),
+               []>, EVEX, EVEX_KZ;
+
+  def mr : AVX512XS8I<opc, MRMDestMem, (outs), (ins x86memop:$dst, srcRC:$src),
+               !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
+               []>, EVEX;
+}
+defm VPMOVQB    : avx512_trunc_sat<0x32, "vpmovqb",   VR128X, VR512, VK8WM, 
+                                 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VO>;
+defm VPMOVSQB   : avx512_trunc_sat<0x22, "vpmovsqb",  VR128X, VR512, VK8WM,
+                                 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VO>;
+defm VPMOVUSQB  : avx512_trunc_sat<0x12, "vpmovusqb", VR128X, VR512, VK8WM,
+                                 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VO>;
+defm VPMOVQW    : avx512_trunc_sat<0x34, "vpmovqw",   VR128X, VR512, VK8WM,
+                                 i128mem>, EVEX_V512, EVEX_CD8<16, CD8VQ>;
+defm VPMOVSQW   : avx512_trunc_sat<0x24, "vpmovsqw",  VR128X, VR512, VK8WM,
+                                 i128mem>, EVEX_V512, EVEX_CD8<16, CD8VQ>;
+defm VPMOVUSQW  : avx512_trunc_sat<0x14, "vpmovusqw", VR128X, VR512, VK8WM,
+                                 i128mem>, EVEX_V512, EVEX_CD8<16, CD8VQ>;
+defm VPMOVQD    : avx512_trunc_sat<0x35, "vpmovqd",   VR256X, VR512, VK8WM,
+                                 i256mem>, EVEX_V512, EVEX_CD8<32, CD8VH>;
+defm VPMOVSQD   : avx512_trunc_sat<0x25, "vpmovsqd",  VR256X, VR512, VK8WM,
+                                 i256mem>, EVEX_V512, EVEX_CD8<32, CD8VH>;
+defm VPMOVUSQD  : avx512_trunc_sat<0x15, "vpmovusqd", VR256X, VR512, VK8WM,
+                                 i256mem>, EVEX_V512, EVEX_CD8<32, CD8VH>;
+defm VPMOVDW    : avx512_trunc_sat<0x33, "vpmovdw",   VR256X, VR512, VK16WM,
+                                 i256mem>, EVEX_V512, EVEX_CD8<16, CD8VH>;
+defm VPMOVSDW   : avx512_trunc_sat<0x23, "vpmovsdw",  VR256X, VR512, VK16WM,
+                                 i256mem>, EVEX_V512, EVEX_CD8<16, CD8VH>;
+defm VPMOVUSDW  : avx512_trunc_sat<0x13, "vpmovusdw", VR256X, VR512, VK16WM,
+                                 i256mem>, EVEX_V512, EVEX_CD8<16, CD8VH>;
+defm VPMOVDB    : avx512_trunc_sat<0x31, "vpmovdb",   VR128X, VR512, VK16WM,
+                                 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VQ>;
+defm VPMOVSDB   : avx512_trunc_sat<0x21, "vpmovsdb",  VR128X, VR512, VK16WM,
+                                 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VQ>;
+defm VPMOVUSDB  : avx512_trunc_sat<0x11, "vpmovusdb", VR128X, VR512, VK16WM,
+                                 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VQ>;
+
+def : Pat<(v16i8  (X86vtrunc (v8i64  VR512:$src))), (VPMOVQBrr  VR512:$src)>;
+def : Pat<(v8i16  (X86vtrunc (v8i64  VR512:$src))), (VPMOVQWrr  VR512:$src)>;
+def : Pat<(v16i16 (X86vtrunc (v16i32 VR512:$src))), (VPMOVDWrr  VR512:$src)>;
+def : Pat<(v16i8  (X86vtrunc (v16i32 VR512:$src))), (VPMOVDBrr  VR512:$src)>;
+def : Pat<(v8i32  (X86vtrunc (v8i64  VR512:$src))), (VPMOVQDrr  VR512:$src)>;
+
+def : Pat<(v16i8  (X86vtruncm VK16WM:$mask, (v16i32 VR512:$src))),
+                  (VPMOVDBkrr VK16WM:$mask, VR512:$src)>;
+def : Pat<(v16i16 (X86vtruncm VK16WM:$mask, (v16i32 VR512:$src))),
+                  (VPMOVDWkrr VK16WM:$mask, VR512:$src)>;
+def : Pat<(v8i16  (X86vtruncm VK8WM:$mask,  (v8i64 VR512:$src))),
+                  (VPMOVQWkrr  VK8WM:$mask, VR512:$src)>;
+def : Pat<(v8i32  (X86vtruncm VK8WM:$mask,  (v8i64 VR512:$src))),
+                  (VPMOVQDkrr  VK8WM:$mask, VR512:$src)>;
+
+
+multiclass avx512_extend<bits<8> opc, string OpcodeStr, RegisterClass DstRC,
+                      RegisterClass SrcRC, SDNode OpNode, PatFrag mem_frag, 
+                      X86MemOperand x86memop, ValueType OpVT, ValueType InVT> {
+
+  def rr : AVX5128I<opc, MRMSrcReg, (outs DstRC:$dst),
+              (ins SrcRC:$src),
+              !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
+              [(set DstRC:$dst, (OpVT (OpNode (InVT SrcRC:$src))))]>, EVEX;
+  def rm : AVX5128I<opc, MRMSrcMem, (outs DstRC:$dst),
+              (ins x86memop:$src),
+              !strconcat(OpcodeStr,"\t{$src, $dst|$dst, $src}"),
+              [(set DstRC:$dst,
+                (OpVT (OpNode (InVT (bitconvert (mem_frag addr:$src))))))]>,
+              EVEX;
+}
+
+defm VPMOVZXBDZ: avx512_extend<0x31, "vpmovzxbd", VR512, VR128X, X86vzext, 
+                             memopv2i64, i128mem, v16i32, v16i8>, EVEX_V512,
+                             EVEX_CD8<8, CD8VQ>;
+defm VPMOVZXBQZ: avx512_extend<0x32, "vpmovzxbq", VR512, VR128X, X86vzext, 
+                             memopv2i64, i128mem, v8i64, v16i8>, EVEX_V512,
+                             EVEX_CD8<8, CD8VO>;
+defm VPMOVZXWDZ: avx512_extend<0x33, "vpmovzxwd", VR512, VR256X, X86vzext, 
+                             memopv4i64, i256mem, v16i32, v16i16>, EVEX_V512,
+                             EVEX_CD8<16, CD8VH>;
+defm VPMOVZXWQZ: avx512_extend<0x34, "vpmovzxwq", VR512, VR128X, X86vzext, 
+                             memopv2i64, i128mem, v8i64, v8i16>, EVEX_V512,
+                             EVEX_CD8<16, CD8VQ>;
+defm VPMOVZXDQZ: avx512_extend<0x35, "vpmovzxdq", VR512, VR256X, X86vzext, 
+                             memopv4i64, i256mem, v8i64, v8i32>, EVEX_V512,
+                             EVEX_CD8<32, CD8VH>;
+                             
+defm VPMOVSXBDZ: avx512_extend<0x21, "vpmovsxbd", VR512, VR128X, X86vsext, 
+                             memopv2i64, i128mem, v16i32, v16i8>, EVEX_V512,
+                             EVEX_CD8<8, CD8VQ>;
+defm VPMOVSXBQZ: avx512_extend<0x22, "vpmovsxbq", VR512, VR128X, X86vsext, 
+                             memopv2i64, i128mem, v8i64, v16i8>, EVEX_V512,
+                             EVEX_CD8<8, CD8VO>;
+defm VPMOVSXWDZ: avx512_extend<0x23, "vpmovsxwd", VR512, VR256X, X86vsext, 
+                             memopv4i64, i256mem, v16i32, v16i16>, EVEX_V512,
+                             EVEX_CD8<16, CD8VH>;
+defm VPMOVSXWQZ: avx512_extend<0x24, "vpmovsxwq", VR512, VR128X, X86vsext, 
+                             memopv2i64, i128mem, v8i64, v8i16>, EVEX_V512,
+                             EVEX_CD8<16, CD8VQ>;
+defm VPMOVSXDQZ: avx512_extend<0x25, "vpmovsxdq", VR512, VR256X, X86vsext, 
+                             memopv4i64, i256mem, v8i64, v8i32>, EVEX_V512,
+                             EVEX_CD8<32, CD8VH>;
+
 //===----------------------------------------------------------------------===//
 // VSHUFPS - VSHUFPD Operations