Introduce a new technique for merging BasicBlock with Instruction sentinel by superpo...

[oota-llvm.git] / lib / Target / X86 / X86InstrMMX.td

diff --git a/lib/Target/X86/X86InstrMMX.td b/lib/Target/X86/X86InstrMMX.td
index bf0dcd20e25111d551c635cca0392cf1b0f040c0..71f2cb164d51176bec10508af3deb63370819953 100644 (file)
--- a/lib/Target/X86/X86InstrMMX.td
+++ b/lib/Target/X86/X86InstrMMX.td
@@ -146,7 +146,7 @@ def MMX_FEMMS : MMXI<0x0E, RawFrm, (outs), (ins), "femms", [(int_x86_mmx_femms)]
 def MMX_MOVD64rr : MMXI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR32:$src),
                         "movd\t{$src, $dst|$dst, $src}",
                         [(set VR64:$dst, (v2i32 (scalar_to_vector GR32:$src)))]>;
-let isSimpleLoad = 1, isReMaterializable = 1 in
+let canFoldAsLoad = 1, isReMaterializable = 1 in
 def MMX_MOVD64rm : MMXI<0x6E, MRMSrcMem, (outs VR64:$dst), (ins i32mem:$src),
                         "movd\t{$src, $dst|$dst, $src}",
               [(set VR64:$dst, (v2i32 (scalar_to_vector (loadi32 addr:$src))))]>;
@@ -156,16 +156,18 @@ def MMX_MOVD64mr : MMXI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR64:$src),
 
 let neverHasSideEffects = 1 in
 def MMX_MOVD64to64rr : MMXRI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR64:$src),
-                             "movd\t{$src, $dst|$dst, $src}", []>;
+                             "movd\t{$src, $dst|$dst, $src}",
+                             []>;
 
 let neverHasSideEffects = 1 in
-def MMX_MOVD64from64rr : MMXRI<0x7E, MRMSrcReg, (outs GR64:$dst), (ins VR64:$src),
+def MMX_MOVD64from64rr : MMXRI<0x7E, MRMSrcReg,
+                               (outs GR64:$dst), (ins VR64:$src),
                                "movd\t{$src, $dst|$dst, $src}", []>;
 
 let neverHasSideEffects = 1 in
 def MMX_MOVQ64rr : MMXI<0x6F, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
                         "movq\t{$src, $dst|$dst, $src}", []>;
-let isSimpleLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
+let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
 def MMX_MOVQ64rm : MMXI<0x6F, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
                         "movq\t{$src, $dst|$dst, $src}",
                         [(set VR64:$dst, (load_mmx addr:$src))]>;
@@ -187,6 +189,10 @@ def MMX_MOVQ2DQrr : SSDIi8<0xD6, MRMDestMem, (outs VR128:$dst), (ins VR64:$src),
                         (v2i64 (scalar_to_vector (i64 (bitconvert VR64:$src)))),
                         MOVL_shuffle_mask)))]>;
 
+let neverHasSideEffects = 1 in
+def MMX_MOVQ2FR64rr: SSDIi8<0xD6, MRMDestMem, (outs FR64:$dst), (ins VR64:$src),
+                           "movq2dq\t{$src, $dst|$dst, $src}", []>;
+
 def MMX_MOVNTQmr  : MMXI<0xE7, MRMDestMem, (outs), (ins i64mem:$dst, VR64:$src),
                          "movntq\t{$src, $dst|$dst, $src}",
                          [(int_x86_mmx_movnt_dq addr:$dst, VR64:$src)]>;
@@ -575,9 +581,17 @@ def : Pat<(i64 (bitconvert (v4i16 VR64:$src))),
           (MMX_MOVD64from64rr VR64:$src)>;
 def : Pat<(i64  (bitconvert (v8i8 VR64:$src))),
           (MMX_MOVD64from64rr VR64:$src)>;
-
-// Move scalar to XMM zero-extended
-// movd to XMM register zero-extends
+def : Pat<(f64 (bitconvert (v1i64 VR64:$src))),
+          (MMX_MOVQ2FR64rr VR64:$src)>;
+def : Pat<(f64 (bitconvert (v2i32 VR64:$src))),
+          (MMX_MOVQ2FR64rr VR64:$src)>;
+def : Pat<(f64 (bitconvert (v4i16 VR64:$src))),
+          (MMX_MOVQ2FR64rr VR64:$src)>;
+def : Pat<(f64 (bitconvert (v8i8 VR64:$src))),
+          (MMX_MOVQ2FR64rr VR64:$src)>;
+
+// Move scalar to MMX zero-extended
+// movd to MMX register zero-extends
 let AddedComplexity = 15 in {
   def : Pat<(v8i8 (X86vzmovl (bc_v8i8 (v2i32 (scalar_to_vector GR32:$src))))),
            (MMX_MOVZDI2PDIrr GR32:$src)>; 
@@ -585,6 +599,25 @@ let AddedComplexity = 15 in {
            (MMX_MOVZDI2PDIrr GR32:$src)>; 
 }
 
+let AddedComplexity = 20 in {
+  def : Pat<(v8i8 (X86vzmovl (bc_v8i8 (load_mmx addr:$src)))),
+            (MMX_MOVZDI2PDIrm addr:$src)>; 
+  def : Pat<(v4i16 (X86vzmovl (bc_v4i16 (load_mmx addr:$src)))),
+            (MMX_MOVZDI2PDIrm addr:$src)>; 
+  def : Pat<(v2i32 (X86vzmovl (bc_v2i32 (load_mmx addr:$src)))),
+            (MMX_MOVZDI2PDIrm addr:$src)>; 
+}
+
+// Clear top half.
+let AddedComplexity = 15 in {
+  def : Pat<(v8i8 (X86vzmovl VR64:$src)),
+            (MMX_PUNPCKLDQrr VR64:$src, (MMX_V_SET0))>;
+  def : Pat<(v4i16 (X86vzmovl VR64:$src)),
+            (MMX_PUNPCKLDQrr VR64:$src, (MMX_V_SET0))>;
+  def : Pat<(v2i32 (X86vzmovl VR64:$src)),
+            (MMX_PUNPCKLDQrr VR64:$src, (MMX_V_SET0))>;
+}
+
 // Scalar to v4i16 / v8i8. The source may be a GR32, but only the lower
 // 8 or 16-bits matter.
 def : Pat<(bc_v8i8  (v2i32 (scalar_to_vector GR32:$src))),
@@ -648,7 +681,13 @@ def : Pat<(v1i64 (and (xor VR64:$src1, (bc_v1i64 (v8i8  immAllOnesV_bc))),
           (MMX_PANDNrm VR64:$src1, addr:$src2)>;
 
 // Move MMX to lower 64-bit of XMM
-def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert VR64:$src)))),
+def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert (v8i8 VR64:$src))))),
+          (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;
+def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert (v4i16 VR64:$src))))),
+          (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;
+def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert (v2i32 VR64:$src))))),
+          (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;
+def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert (v1i64 VR64:$src))))),
           (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;
 
 // Move lower 64-bit of XMM to MMX.
@@ -662,3 +701,14 @@ def : Pat<(v8i8 (bitconvert (i64 (vector_extract (v2i64 VR128:$src),
                                                   (iPTR 0))))),
           (v8i8 (MMX_MOVDQ2Qrr VR128:$src))>;
 
+// CMOV* - Used to implement the SELECT DAG operation.  Expanded by the
+// scheduler into a branch sequence.
+// These are expanded by the scheduler.
+let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
+  def CMOV_V1I64 : I<0, Pseudo,
+                    (outs VR64:$dst), (ins VR64:$t, VR64:$f, i8imm:$cond),
+                    "#CMOV_V1I64 PSEUDO!",
+                    [(set VR64:$dst,
+                      (v1i64 (X86cmov VR64:$t, VR64:$f, imm:$cond,
+                                          EFLAGS)))]>;
+}