ptx: add shift instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@120982 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 92 insertions, 0 deletions

diff --git a/lib/Target/PTX/PTXInstrInfo.td b/lib/Target/PTX/PTXInstrInfo.td
index 01e0f2a..ccd77df 100644
--- a/lib/Target/PTX/PTXInstrInfo.td
+++ b/lib/Target/PTX/PTXInstrInfo.td
@@ -46,6 +46,11 @@ def MEMii : Operand<i32> {
 // PTX Specific Node Definitions
 //===----------------------------------------------------------------------===//
 
+// PTX allow generic 3-reg shifts like shl r0, r1, r2
+def PTXshl : SDNode<"ISD::SHL", SDTIntBinOp>;
+def PTXsrl : SDNode<"ISD::SRL", SDTIntBinOp>;
+def PTXsra : SDNode<"ISD::SRA", SDTIntBinOp>;
+
 def PTXexit
   : SDNode<"PTXISD::EXIT", SDTNone, [SDNPHasChain]>;
 def PTXret
@@ -66,6 +71,22 @@ multiclass INT3<string opcstr, SDNode opnode> {
                    [(set RRegs32:$d, (opnode RRegs32:$a, imm:$b))]>;
 }
 
+// no %type directive, non-communtable
+multiclass INT3ntnc<string opcstr, SDNode opnode> {
+  def rr : InstPTX<(outs RRegs32:$d),
+                   (ins RRegs32:$a, RRegs32:$b),
+                   !strconcat(opcstr, "\t$d, $a, $b"),
+                   [(set RRegs32:$d, (opnode RRegs32:$a, RRegs32:$b))]>;
+  def ri : InstPTX<(outs RRegs32:$d),
+                   (ins RRegs32:$a, i32imm:$b),
+                   !strconcat(opcstr, "\t$d, $a, $b"),
+                   [(set RRegs32:$d, (opnode RRegs32:$a, imm:$b))]>;
+  def ir : InstPTX<(outs RRegs32:$d),
+                   (ins i32imm:$a, RRegs32:$b),
+                   !strconcat(opcstr, "\t$d, $a, $b"),
+                   [(set RRegs32:$d, (opnode imm:$a, RRegs32:$b))]>;
+}
+
 multiclass PTX_LD<string opstr, RegisterClass RC, PatFrag pat_load> {
   def ri : InstPTX<(outs RC:$d),
                    (ins MEMri:$a),
@@ -86,6 +107,12 @@ multiclass PTX_LD<string opstr, RegisterClass RC, PatFrag pat_load> {
 defm ADD : INT3<"add", add>;
 defm SUB : INT3<"sub", sub>;
 
+///===- Logic and Shift Instructions --------------------------------------===//
+
+defm SHL : INT3ntnc<"shl.b32", PTXshl>;
+defm SRL : INT3ntnc<"shr.u32", PTXsrl>;
+defm SRA : INT3ntnc<"shr.s32", PTXsra>;
+
 ///===- Data Movement and Conversion Instructions -------------------------===//
 
 let neverHasSideEffects = 1 in {
diff --git a/test/CodeGen/PTX/shl.ll b/test/CodeGen/PTX/shl.ll
new file mode 100644
index 0000000..b564b43
--- /dev/null
+++ b/test/CodeGen/PTX/shl.ll
@@ -0,0 +1,22 @@
+; RUN: llc < %s -march=ptx | FileCheck %s
+
+define ptx_device i32 @t1(i32 %x, i32 %y) {
+; CHECK: shl.b32 r0, r1, r2
+	%z = shl i32 %x, %y
+; CHECK: ret;
+	ret i32 %z
+}
+
+define ptx_device i32 @t2(i32 %x) {
+; CHECK: shl.b32 r0, r1, 3
+	%z = shl i32 %x, 3
+; CHECK: ret;
+	ret i32 %z
+}
+
+define ptx_device i32 @t3(i32 %x) {
+; CHECK: shl.b32 r0, 3, r1
+	%z = shl i32 3, %x
+; CHECK: ret;
+	ret i32 %z
+}
diff --git a/test/CodeGen/PTX/shr.ll b/test/CodeGen/PTX/shr.ll
new file mode 100644
index 0000000..3f8ade8
--- /dev/null
+++ b/test/CodeGen/PTX/shr.ll
@@ -0,0 +1,43 @@
+; RUN: llc < %s -march=ptx | FileCheck %s
+
+define ptx_device i32 @t1(i32 %x, i32 %y) {
+; CHECK: shr.u32 r0, r1, r2
+	%z = lshr i32 %x, %y
+; CHECK: ret;
+	ret i32 %z
+}
+
+define ptx_device i32 @t2(i32 %x) {
+; CHECK: shr.u32 r0, r1, 3
+	%z = lshr i32 %x, 3
+; CHECK: ret;
+	ret i32 %z
+}
+
+define ptx_device i32 @t3(i32 %x) {
+; CHECK: shr.u32 r0, 3, r1
+	%z = lshr i32 3, %x
+; CHECK: ret;
+	ret i32 %z
+}
+
+define ptx_device i32 @t4(i32 %x, i32 %y) {
+; CHECK: shr.s32 r0, r1, r2
+	%z = ashr i32 %x, %y
+; CHECK: ret;
+	ret i32 %z
+}
+
+define ptx_device i32 @t5(i32 %x) {
+; CHECK: shr.s32 r0, r1, 3
+	%z = ashr i32 %x, 3
+; CHECK: ret;
+	ret i32 %z
+}
+
+define ptx_device i32 @t6(i32 %x) {
+; CHECK: shr.s32 r0, -3, r1
+	%z = ashr i32 -3, %x
+; CHECK: ret;
+	ret i32 %z
+}