[AMDGPU] Split R600 and GCN bfe patterns

[foad]

Summary

This is in preparation for making the GCN patterns divergence-aware.
NFC.

[arsenm]

llvm/lib/Target/AMDGPU/SIInstructions.td

2360

This and these others can directly use 0

[foad]

Remove unnecessary S_MOV_B32 from patterns.

[arsenm]

LGTM. We do need a set for the scalar cases though

[foad]

In D88579#2306193, @arsenm wrote:

We do need a set for the scalar cases though

It's currently being done by AMDGPUDAGToDAGISel::SelectS_BFE instead of with patterns, right? I guess you'd prefer patterns so GlobalISel can use them too?

[arsenm]

In D88579#2306199, @foad wrote:

In D88579#2306193, @arsenm wrote:

We do need a set for the scalar cases though

It's currently being done by AMDGPUDAGToDAGISel::SelectS_BFE instead of with patterns, right? I guess you'd prefer patterns so GlobalISel can use them too?

Yes

llvm/lib/Target/AMDGPU/AMDGPUInstructions.td

 
 // XXX: Convert to new syntax and use COPY_TO_REG, once the DFAPacketizer
 // can handle COPY instructions.
 class DwordAddrPat<ValueType vt, RegisterClass rc> : AMDGPUPat <
   (vt (AMDGPUdwordaddr (vt rc:$addr))),
   (vt rc:$addr)
 >;
 
-// Bitfield extract patterns
-
-def IMMZeroBasedBitfieldMask : ImmLeaf <i32, [{
-  return isMask_32(Imm);
-}]>;
-
-def IMMPopCount : SDNodeXForm<imm, [{
-  return CurDAG->getTargetConstant(countPopulation(N->getZExtValue()), SDLoc(N),
-                                   MVT::i32);
-}]>;
-
-multiclass BFEPattern <Instruction UBFE, Instruction SBFE, Instruction MOV> {
-  def : AMDGPUPat <
-    (i32 (and (i32 (srl i32:$src, i32:$rshift)), IMMZeroBasedBitfieldMask:$mask)),
-    (UBFE $src, $rshift, (MOV (i32 (IMMPopCount $mask))))
-  >;
-
-  // x & ((1 << y) - 1)
-  def : AMDGPUPat <
-    (and i32:$src, (add_oneuse (shl_oneuse 1, i32:$width), -1)),
-    (UBFE $src, (MOV (i32 0)), $width)
-  >;
-
-  // x & ~(-1 << y)
-  def : AMDGPUPat <
-    (and i32:$src, (xor_oneuse (shl_oneuse -1, i32:$width), -1)),
-    (UBFE $src, (MOV (i32 0)), $width)
-  >;
-
-  // x & (-1 >> (bitwidth - y))
-  def : AMDGPUPat <
-    (and i32:$src, (srl_oneuse -1, (sub 32, i32:$width))),
-    (UBFE $src, (MOV (i32 0)), $width)
-  >;
-
-  // x << (bitwidth - y) >> (bitwidth - y)
-  def : AMDGPUPat <
-    (srl (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
-    (UBFE $src, (MOV (i32 0)), $width)
-  >;
-
-  def : AMDGPUPat <
-    (sra (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
-    (SBFE $src, (MOV (i32 0)), $width)
-  >;
-}
-
 // fshr pattern
 class FSHRPattern <Instruction BIT_ALIGN> : AMDGPUPat <
   (fshr i32:$src0, i32:$src1, i32:$src2),
   (BIT_ALIGN $src0, $src1, $src2)
 >;
 
 // rotr pattern
 class ROTRPattern <Instruction BIT_ALIGN> : AMDGPUPat <

llvm/lib/Target/AMDGPU/EvergreenInstructions.td

 class EGPat<dag pattern, dag result> : AMDGPUPat<pattern, result> {
   let SubtargetPredicate = isEG;
 }
 
 class EGOrCaymanPat<dag pattern, dag result> : AMDGPUPat<pattern, result> {
   let SubtargetPredicate = isEGorCayman;
 }
 
+def IMMZeroBasedBitfieldMask : ImmLeaf <i32, [{
+  return isMask_32(Imm);
+}]>;
+
+def IMMPopCount : SDNodeXForm<imm, [{
+  return CurDAG->getTargetConstant(countPopulation(N->getZExtValue()), SDLoc(N),
+                                   MVT::i32);
+}]>;
+
 //===----------------------------------------------------------------------===//
 // Evergreen / Cayman store instructions
 //===----------------------------------------------------------------------===//
 
 let SubtargetPredicate = isEGorCayman in {
 
 class CF_MEM_RAT_CACHELESS <bits<6> rat_inst, bits<4> rat_id, bits<4> mask, dag ins,
                            string name, list<dag> pattern>
def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
   VecALU
 >;
 
 def BFE_INT_eg : R600_3OP <0x5, "BFE_INT",
   [(set i32:$dst, (AMDGPUbfe_i32 i32:$src0, i32:$src1, i32:$src2))],
   VecALU
 >;
 
-defm : BFEPattern <BFE_UINT_eg, BFE_INT_eg, MOV_IMM_I32>;
+// Bitfield extract patterns
+
+def : AMDGPUPat <
+  (and (i32 (srl i32:$src, i32:$rshift)), IMMZeroBasedBitfieldMask:$mask),
+  (BFE_UINT_eg $src, $rshift, (MOV_IMM_I32 (i32 (IMMPopCount $mask))))
+>;
+
+// x & ((1 << y) - 1)
+def : AMDGPUPat <
+  (and i32:$src, (add_oneuse (shl_oneuse 1, i32:$width), -1)),
+  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
+>;
+
+// x & ~(-1 << y)
+def : AMDGPUPat <
+  (and i32:$src, (xor_oneuse (shl_oneuse -1, i32:$width), -1)),
+  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
+>;
+
+// x & (-1 >> (bitwidth - y))
+def : AMDGPUPat <
+  (and i32:$src, (srl_oneuse -1, (sub 32, i32:$width))),
+  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
+>;
+
+// x << (bitwidth - y) >> (bitwidth - y)
+def : AMDGPUPat <
+  (srl (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
+  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
+>;
+
+def : AMDGPUPat <
+  (sra (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
+  (BFE_INT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
+>;
 
 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT",
   [(set i32:$dst, (AMDGPUbfi i32:$src0, i32:$src1, i32:$src2))],
   VecALU
 >;
 
 def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i1)),
   (BFE_INT_eg i32:$src, (i32 ZERO), (i32 ONE_INT))>;

llvm/lib/Target/AMDGPU/SIInstructions.td

def : GCNPat <
     (vt (add (vt (shl 1, vt:$a)), -1)),
     (BFM $a, (MOV (i32 0)))
   >;
 }
 
 defm : BFMPatterns <i32, S_BFM_B32, S_MOV_B32>;
 // FIXME: defm : BFMPatterns <i64, S_BFM_B64, S_MOV_B64>;
 
-defm : BFEPattern <V_BFE_U32, V_BFE_I32, S_MOV_B32>;
+// Bitfield extract patterns
+
+def IMMZeroBasedBitfieldMask : ImmLeaf <i32, [{
+  return isMask_32(Imm);
+}]>;
+
+def IMMPopCount : SDNodeXForm<imm, [{
+  return CurDAG->getTargetConstant(countPopulation(N->getZExtValue()), SDLoc(N),
+                                   MVT::i32);
+}]>;
+
+def : AMDGPUPat <
+  (and (i32 (srl i32:$src, i32:$rshift)), IMMZeroBasedBitfieldMask:$mask),
+  (V_BFE_U32 $src, $rshift, (i32 (IMMPopCount $mask)))
+>;
+
+// x & ((1 << y) - 1)
+def : AMDGPUPat <
+  (and i32:$src, (add_oneuse (shl_oneuse 1, i32:$width), -1)),
+  (V_BFE_U32 $src, (i32 0), $width)
+>;
+
+// x & ~(-1 << y)
+def : AMDGPUPat <
+  (and i32:$src, (xor_oneuse (shl_oneuse -1, i32:$width), -1)),
+  (V_BFE_U32 $src, (i32 0), $width)
+>;
+
+// x & (-1 >> (bitwidth - y))
+def : AMDGPUPat <
+  (and i32:$src, (srl_oneuse -1, (sub 32, i32:$width))),
+  (V_BFE_U32 $src, (i32 0), $width)
+>;
+
+// x << (bitwidth - y) >> (bitwidth - y)
+def : AMDGPUPat <
+  (srl (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
+  (V_BFE_U32 $src, (i32 0), $width)
+>;
+
+def : AMDGPUPat <
+  (sra (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
+  (V_BFE_I32 $src, (i32 0), $width)
+>;
 
 // SHA-256 Ma patterns
 
 // ((x & z) | (y & (x | z))) -> BFI (XOR x, y), z, y
 def : AMDGPUPat <
   (DivergentBinFrag<or> (and i32:$x, i32:$z),
                         (and i32:$y, (or i32:$x, i32:$z))),
   (V_BFI_B32 (V_XOR_B32_e64 i32:$x, i32:$y), i32:$z, i32:$y)