[mips] Fix inefficient code generation.

This patch eliminates the need to emit a constant move instruction when this
pattern is matched:

(select (setgt a, Constant), T, F)

The pattern above effectively turns into this:

(conditional-move (setlt a, Constant + 1), F, T)



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@176384 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 156 insertions, 4 deletions

diff --git a/lib/Target/Mips/MipsCondMov.td b/lib/Target/Mips/MipsCondMov.td
index 1c68f2407e..42e4c99f05 100644
--- a/lib/Target/Mips/MipsCondMov.td
+++ b/lib/Target/Mips/MipsCondMov.td
@@ -68,6 +68,13 @@ multiclass MovzPats0<RegisterClass CRC, RegisterClass DRC,
                 (MOVZInst DRC:$T, (SLTOp CRC:$rhs, CRC:$lhs), DRC:$F)>;
   def : MipsPat<(select (i32 (setule CRC:$lhs, CRC:$rhs)), DRC:$T, DRC:$F),
                 (MOVZInst DRC:$T, (SLTuOp CRC:$rhs, CRC:$lhs), DRC:$F)>;
+  def : MipsPat<(select (i32 (setgt CRC:$lhs, immSExt16Plus1:$rhs)),
+                        DRC:$T, DRC:$F),
+                (MOVZInst DRC:$T, (SLTiOp CRC:$lhs, (Plus1 imm:$rhs)), DRC:$F)>;
+  def : MipsPat<(select (i32 (setugt CRC:$lhs, immSExt16Plus1:$rhs)),
+                        DRC:$T, DRC:$F),
+                (MOVZInst DRC:$T, (SLTiuOp CRC:$lhs, (Plus1 imm:$rhs)),
+                          DRC:$F)>;
 }
 
 multiclass MovzPats1<RegisterClass CRC, RegisterClass DRC,
diff --git a/lib/Target/Mips/MipsISelDAGToDAG.cpp b/lib/Target/Mips/MipsISelDAGToDAG.cpp
index 78c74ef879..6dff548505 100644
--- a/lib/Target/Mips/MipsISelDAGToDAG.cpp
+++ b/lib/Target/Mips/MipsISelDAGToDAG.cpp
@@ -109,7 +109,7 @@ private:
        SDValue &Alias);
 
   // getImm - Return a target constant with the specified value.
-  inline SDValue getImm(const SDNode *Node, unsigned Imm) {
+  inline SDValue getImm(const SDNode *Node, uint64_t Imm) {
     return CurDAG->getTargetConstant(Imm, Node->getValueType(0));
   }
 
diff --git a/lib/Target/Mips/MipsInstrInfo.td b/lib/Target/Mips/MipsInstrInfo.td
index de09c9ea92..b71ced1493 100644
--- a/lib/Target/Mips/MipsInstrInfo.td
+++ b/lib/Target/Mips/MipsInstrInfo.td
@@ -299,6 +299,9 @@ def HI16 : SDNodeXForm<imm, [{
   return getImm(N, (N->getZExtValue() >> 16) & 0xFFFF);
 }]>;
 
+// Plus 1.
+def Plus1 : SDNodeXForm<imm, [{ return getImm(N, N->getSExtValue() + 1); }]>;
+
 // Node immediate fits as 16-bit sign extended on target immediate.
 // e.g. addi, andi
 def immSExt8  : PatLeaf<(imm), [{ return isInt<8>(N->getSExtValue()); }]>;
@@ -331,6 +334,11 @@ def immLow16Zero : PatLeaf<(imm), [{
 // shamt field must fit in 5 bits.
 def immZExt5 : ImmLeaf<i32, [{return Imm == (Imm & 0x1f);}]>;
 
+// True if (N + 1) fits in 16-bit field.
+def immSExt16Plus1 : PatLeaf<(imm), [{
+  return isInt<17>(N->getSExtValue()) && isInt<16>(N->getSExtValue() + 1);
+}]>;
+
 // Mips Address Mode! SDNode frameindex could possibily be a match
 // since load and store instructions from stack used it.
 def addr :
diff --git a/test/CodeGen/Mips/cmov.ll b/test/CodeGen/Mips/cmov.ll
index 3af899a4e2..81925a4953 100755
--- a/test/CodeGen/Mips/cmov.ll
+++ b/test/CodeGen/Mips/cmov.ll
@@ -59,3 +59,140 @@ entry:
   ret i64 %cond
 }
 
+; slti and conditional move.
+;
+; Check that, pattern
+;  (select (setgt a, N), t, f)
+; turns into
+;  (movz t, (setlt a, N + 1), f)
+; if N + 1 fits in 16-bit.
+
+; O32: slti0:
+; O32: slti $[[R0:[0-9]+]], ${{[0-9]+}}, 32767
+; O32: movz ${{[0-9]+}}, ${{[0-9]+}}, $[[R0]]
+
+define i32 @slti0(i32 %a) {
+entry:
+  %cmp = icmp sgt i32 %a, 32766
+  %cond = select i1 %cmp, i32 3, i32 4
+  ret i32 %cond
+}
+
+; O32: slti1:
+; O32: slt ${{[0-9]+}}
+
+define i32 @slti1(i32 %a) {
+entry:
+  %cmp = icmp sgt i32 %a, 32767
+  %cond = select i1 %cmp, i32 3, i32 4
+  ret i32 %cond
+}
+
+; O32: slti2:
+; O32: slti $[[R0:[0-9]+]], ${{[0-9]+}}, -32768
+; O32: movz ${{[0-9]+}}, ${{[0-9]+}}, $[[R0]]
+
+define i32 @slti2(i32 %a) {
+entry:
+  %cmp = icmp sgt i32 %a, -32769
+  %cond = select i1 %cmp, i32 3, i32 4
+  ret i32 %cond
+}
+
+; O32: slti3:
+; O32: slt ${{[0-9]+}}
+
+define i32 @slti3(i32 %a) {
+entry:
+  %cmp = icmp sgt i32 %a, -32770
+  %cond = select i1 %cmp, i32 3, i32 4
+  ret i32 %cond
+}
+
+; 64-bit patterns.
+
+; N64: slti64_0:
+; N64: slti $[[R0:[0-9]+]], ${{[0-9]+}}, 32767
+; N64: movz ${{[0-9]+}}, ${{[0-9]+}}, $[[R0]]
+
+define i64 @slti64_0(i64 %a) {
+entry:
+  %cmp = icmp sgt i64 %a, 32766
+  %conv = select i1 %cmp, i64 3, i64 4
+  ret i64 %conv
+}
+
+; N64: slti64_1:
+; N64: slt ${{[0-9]+}}
+
+define i64 @slti64_1(i64 %a) {
+entry:
+  %cmp = icmp sgt i64 %a, 32767
+  %conv = select i1 %cmp, i64 3, i64 4
+  ret i64 %conv
+}
+
+; N64: slti64_2:
+; N64: slti $[[R0:[0-9]+]], ${{[0-9]+}}, -32768
+; N64: movz ${{[0-9]+}}, ${{[0-9]+}}, $[[R0]]
+
+define i64 @slti64_2(i64 %a) {
+entry:
+  %cmp = icmp sgt i64 %a, -32769
+  %conv = select i1 %cmp, i64 3, i64 4
+  ret i64 %conv
+}
+
+; N64: slti64_3:
+; N64: slt ${{[0-9]+}}
+
+define i64 @slti64_3(i64 %a) {
+entry:
+  %cmp = icmp sgt i64 %a, -32770
+  %conv = select i1 %cmp, i64 3, i64 4
+  ret i64 %conv
+}
+
+; sltiu instructions.
+
+; O32: sltiu0:
+; O32: sltiu $[[R0:[0-9]+]], ${{[0-9]+}}, 32767
+; O32: movz ${{[0-9]+}}, ${{[0-9]+}}, $[[R0]]
+
+define i32 @sltiu0(i32 %a) {
+entry:
+  %cmp = icmp ugt i32 %a, 32766
+  %cond = select i1 %cmp, i32 3, i32 4
+  ret i32 %cond
+}
+
+; O32: sltiu1:
+; O32: sltu ${{[0-9]+}}
+
+define i32 @sltiu1(i32 %a) {
+entry:
+  %cmp = icmp ugt i32 %a, 32767
+  %cond = select i1 %cmp, i32 3, i32 4
+  ret i32 %cond
+}
+
+; O32: sltiu2:
+; O32: sltiu $[[R0:[0-9]+]], ${{[0-9]+}}, -32768
+; O32: movz ${{[0-9]+}}, ${{[0-9]+}}, $[[R0]]
+
+define i32 @sltiu2(i32 %a) {
+entry:
+  %cmp = icmp ugt i32 %a, -32769
+  %cond = select i1 %cmp, i32 3, i32 4
+  ret i32 %cond
+}
+
+; O32: sltiu3:
+; O32: sltu ${{[0-9]+}}
+
+define i32 @sltiu3(i32 %a) {
+entry:
+  %cmp = icmp ugt i32 %a, -32770
+  %cond = select i1 %cmp, i32 3, i32 4
+  ret i32 %cond
+}
diff --git a/test/CodeGen/Mips/mips64-f128.ll b/test/CodeGen/Mips/mips64-f128.ll
index 9aa8b8e1df..c6dd434dd7 100644
--- a/test/CodeGen/Mips/mips64-f128.ll
+++ b/test/CodeGen/Mips/mips64-f128.ll
@@ -632,9 +632,9 @@ entry:
 ; CHECK: or   $[[R3:[0-9]+]], $8, $zero
 ; CHECK: ld   $25, %call16(__gttf2)($gp)
 ; CHECK: jalr $25
-; CHECK: slt  $1, $zero, $2
-; CHECK: movn $[[R1]], $[[R3]], $1
-; CHECK: movn $[[R0]], $[[R2]], $1
+; CHECK: slti $1, $2, 1
+; CHECK: movz $[[R1]], $[[R3]], $1
+; CHECK: movz $[[R0]], $[[R2]], $1
 ; CHECK: or   $2, $[[R1]], $zero
 ; CHECK: or   $3, $[[R0]], $zero