R600/SI: Add 32-bit encoding of v_cndmask_b32

This was done by refactoring the v_cndmask_b32 tablegen definition
to use inherit from VOP2Inst.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231795 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/SIInstrInfo.td b/lib/Target/R600/SIInstrInfo.td
index 7a313326bd65906f1dcdce5dba64908331006f39..8bb5f30bb6828a6fee815a179403e49e50301950 100644 (file)
--- a/lib/Target/R600/SIInstrInfo.td
+++ b/lib/Target/R600/SIInstrInfo.td
@@ -802,6 +802,11 @@ def VOP_I1_F64_I32 : VOPProfile <[i1, f64, i32, untyped]> {
 def VOP_I64_I64_I32 : VOPProfile <[i64, i64, i32, untyped]>;
 def VOP_I64_I32_I64 : VOPProfile <[i64, i32, i64, untyped]>;
 def VOP_I64_I64_I64 : VOPProfile <[i64, i64, i64, untyped]>;
+def VOP_CNDMASK : VOPProfile <[i32, i32, i32, untyped]> {
+  let Ins32 = (ins Src0RC32:$src0, Src1RC32:$src1, VCCReg:$src2);
+  let Ins64 = (ins Src0RC64:$src0, Src1RC64:$src1, SSrc_64:$src2);
+  let Asm64 = " $dst, $src0, $src1, $src2";
+}
 
 def VOP_F32_F32_F32_F32 : VOPProfile <[f32, f32, f32, f32]>;
 def VOP_MADK : VOPProfile <[f32, f32, f32, f32]> {

diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 8ba89a47b2753b27ec3ef92c154fcc7ea46f8b63..5cdde4388123d8a8fdde97a6a11576117f9070bf 100644 (file)
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1483,13 +1483,17 @@ defm V_INTERP_MOV_F32 : VINTRP_m <
 // VOP2 Instructions
 //===----------------------------------------------------------------------===//
 
-defm V_CNDMASK_B32_e64 : VOP3_m_nomods <vop3<0x100>, (outs VGPR_32:$dst),
-  (ins VSrc_32:$src0, VSrc_32:$src1, SSrc_64:$src2),
-  "v_cndmask_b32_e64 $dst, $src0, $src1, $src2",
-  [(set i32:$dst, (select i1:$src2, i32:$src1, i32:$src0))],
-  "v_cndmask_b32_e64", 3
->;
+multiclass V_CNDMASK <vop2 op, string name> {
+  defm _e32 : VOP2_m <
+      op, VOP_CNDMASK.Outs, VOP_CNDMASK.Ins32, VOP_CNDMASK.Asm32, [],
+      name, name>;
+
+  defm _e64  : VOP3_m <
+      op, VOP_CNDMASK.Outs, VOP_CNDMASK.Ins64,
+      name#"_e64"#!cast<string>(VOP_CNDMASK.Asm64), [], name, 3>;
+}
 
+defm V_CNDMASK_B32 : V_CNDMASK<vop2<0x0>, "v_cndmask_b32">;
 
 let isCommutable = 1 in {
 defm V_ADD_F32 : VOP2Inst <vop2<0x3, 0x1>, "v_add_f32",
@@ -2261,6 +2265,11 @@ def : Pat <
   (V_BCNT_U32_B32_e64 $popcnt, $val)
 >;
 
+def : Pat <
+  (i32 (select i1:$src0, i32:$src1, i32:$src2)),
+  (V_CNDMASK_B32_e64 $src2, $src1, $src0)
+>;
+
 /********** ======================= **********/
 /********** Image sampling patterns **********/
 /********** ======================= **********/

diff --git a/lib/Target/R600/SIShrinkInstructions.cpp b/lib/Target/R600/SIShrinkInstructions.cpp
index 97bbd78d621fd2b346704db0ae6eb88b612e4481..b115882a84d9e9937337141f122e76b1f5c5970b 100644 (file)
--- a/lib/Target/R600/SIShrinkInstructions.cpp
+++ b/lib/Target/R600/SIShrinkInstructions.cpp
@@ -88,6 +88,11 @@ static bool canShrink(MachineInstr &MI, const SIInstrInfo *TII,
 
   const MachineOperand *Src2 = TII->getNamedOperand(MI, AMDGPU::OpName::src2);
   // Can't shrink instruction with three operands.
+  // FIXME: v_cndmask_b32 has 3 operands and is shrinkable, but we need to add
+  // a special case for it.  It can only be shrunk if the third operand
+  // is vcc.  We should handle this the same way we handle vopc, by addding
+  // a register allocation hint pre-regalloc and then do the shrining
+  // post-regalloc.
   if (Src2)
     return false;