R600/SI: Select V_BFE_U32 for and+shift with a non-literal offset

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@233079 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/AMDGPUInstructions.td b/lib/Target/R600/AMDGPUInstructions.td
index 849b241..4d08201 100644 (file)
--- a/lib/Target/R600/AMDGPUInstructions.td
+++ b/lib/Target/R600/AMDGPUInstructions.td
@@ -578,22 +578,20 @@ class SHA256MaPattern <Instruction BFI_INT, Instruction XOR> : Pat <
 
 // Bitfield extract patterns
 
-/*
-
-XXX: The BFE pattern is not working correctly because the XForm is not being
-applied.
+def IMMZeroBasedBitfieldMask : PatLeaf <(imm), [{
+  return isMask_32(N->getZExtValue());
+}]>;
 
-def legalshift32 : ImmLeaf <i32, [{return Imm >=0 && Imm < 32;}]>;
-def bfemask : PatLeaf <(imm), [{return isMask_32(N->getZExtValue());}],
-                            SDNodeXForm<imm, [{ return CurDAG->getTargetConstant(countTrailingOnes(N->getZExtValue()), MVT::i32);}]>>;
+def IMMPopCount : SDNodeXForm<imm, [{
+  return CurDAG->getTargetConstant(countPopulation(N->getZExtValue()),
+                                   MVT::i32);
+}]>;
 
-class BFEPattern <Instruction BFE> : Pat <
-  (and (srl i32:$x, legalshift32:$y), bfemask:$z),
-  (BFE $x, $y, $z)
+class BFEPattern <Instruction BFE, Instruction MOV> : Pat <
+  (i32 (and (i32 (srl i32:$src, i32:$rshift)), IMMZeroBasedBitfieldMask:$mask)),
+  (BFE $src, $rshift, (MOV (i32 (IMMPopCount $mask))))
 >;
 
-*/
-
 // rotr pattern
 class ROTRPattern <Instruction BIT_ALIGN> : Pat <
   (rotr i32:$src0, i32:$src1),

diff --git a/lib/Target/R600/EvergreenInstructions.td b/lib/Target/R600/EvergreenInstructions.td
index 9f9472c..5560146 100644 (file)
--- a/lib/Target/R600/EvergreenInstructions.td
+++ b/lib/Target/R600/EvergreenInstructions.td
@@ -287,9 +287,8 @@ def BFE_INT_eg : R600_3OP <0x5, "BFE_INT",
   VecALU
 >;
 
-// XXX: This pattern is broken, disabling for now.  See comment in
-// AMDGPUInstructions.td for more info.
-//  def : BFEPattern <BFE_UINT_eg>;
+def : BFEPattern <BFE_UINT_eg, MOV_IMM_I32>;
+
 def BFI_INT_eg : R600_3OP <0x06, "BFI_INT",
   [(set i32:$dst, (AMDGPUbfi i32:$src0, i32:$src1, i32:$src2))],
   VecALU

diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 764e79b..95b2470 100644 (file)
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -3339,6 +3339,8 @@ multiclass BFMPatterns <ValueType vt, InstSI BFM, InstSI MOV> {
 defm : BFMPatterns <i32, S_BFM_B32, S_MOV_B32>;
 // FIXME: defm : BFMPatterns <i64, S_BFM_B64, S_MOV_B64>;
 
+def : BFEPattern <V_BFE_U32, S_MOV_B32>;
+
 //===----------------------------------------------------------------------===//
 // Fract Patterns
 //===----------------------------------------------------------------------===//

diff --git a/test/CodeGen/R600/bfe_uint.ll b/test/CodeGen/R600/bfe_uint.ll
index 6fe23e9..32e3fc2 100644 (file)
--- a/test/CodeGen/R600/bfe_uint.ll
+++ b/test/CodeGen/R600/bfe_uint.ll
@@ -1,7 +1,5 @@
 ; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
 
-; XFAIL: *
-
 ; CHECK: {{^}}bfe_def:
 ; CHECK: BFE_UINT
 define void @bfe_def(i32 addrspace(1)* %out, i32 %x) {

diff --git a/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll b/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
index 0426b3a..5411192 100644 (file)
--- a/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
+++ b/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
@@ -586,6 +586,16 @@ define void @lshr_and(i32 addrspace(1)* %out, i32 %a) nounwind {
   ret void
 }
 
+; FUNC-LABEL: {{^}}v_lshr_and:
+; SI: v_bfe_u32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}, 3
+; SI: buffer_store_dword
+define void @v_lshr_and(i32 addrspace(1)* %out, i32 %a, i32 %b) nounwind {
+  %c = lshr i32 %a, %b
+  %d = and i32 %c, 7
+  store i32 %d, i32 addrspace(1)* %out, align 8
+  ret void
+}
+
 ; FUNC-LABEL: {{^}}and_lshr:
 ; SI: s_bfe_u32 {{s[0-9]+}}, {{s[0-9]+}}, 0x30006
 ; SI: buffer_store_dword