[AMDGPU] Add pattern for v_alignbit_b32 with immediate

If immediate in shift is less than 32 we can use alignbit too.

Differential Revision: https://reviews.llvm.org/D34729

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@306500 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/AMDGPU/SIInstrInfo.td

@@ -336,6 +336,10 @@ def NegSubInlineConst16 : ImmLeaf<i16, [{
   return Imm < -16 && Imm >= -64;
 }], NegateImm>;
 
+def ShiftAmt32Imm : PatLeaf <(imm), [{
+  return N->getZExtValue() < 32;
+}]>;
+
 //===----------------------------------------------------------------------===//
 // Custom Operands
 //===----------------------------------------------------------------------===//

lib/Target/AMDGPU/SIInstructions.td

@@ -933,10 +933,9 @@ def : Pat<(i32 (trunc (srl i64:$src0, (and i32:$src1, (i32 31))))),
           (V_ALIGNBIT_B32 (i32 (EXTRACT_SUBREG (i64 $src0), sub1)),
                           (i32 (EXTRACT_SUBREG (i64 $src0), sub0)), $src1)>;
 
-def : Pat<(i32 (trunc (shl i64:$src0, (and i32:$src1, (i32 31))))),
+def : Pat<(i32 (trunc (srl i64:$src0, (i32 ShiftAmt32Imm:$src1)))),
           (V_ALIGNBIT_B32 (i32 (EXTRACT_SUBREG (i64 $src0), sub1)),
-                          (i32 (EXTRACT_SUBREG (i64 $src0), sub0)),
-                          (S_SUB_I32 (i32 32), $src1))>;
+                          (i32 (EXTRACT_SUBREG (i64 $src0), sub0)), $src1)>;
 
 /********** ====================== **********/
 /**********   Indirect addressing  **********/

test/CodeGen/AMDGPU/alignbit-pat.ll

@@ -68,6 +68,33 @@ bb:
   ret void
 }
 
+; GCN-LABEL: {{^}}alignbit_shr_pat_const30:
+; GCN: load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}
+; GCN: v_alignbit_b32 v{{[0-9]+}}, v[[HI]], v[[LO]], 30
+
+define amdgpu_kernel void @alignbit_shr_pat_const30(i64 addrspace(1)* nocapture readonly %arg, i32 addrspace(1)* nocapture %arg1) {
+bb:
+  %tmp = load i64, i64 addrspace(1)* %arg, align 8
+  %tmp5 = lshr i64 %tmp, 30
+  %tmp6 = trunc i64 %tmp5 to i32
+  store i32 %tmp6, i32 addrspace(1)* %arg1, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}alignbit_shr_pat_wrong_const33:
+; Negative test, shift amount more than 31
+; GCN: v_lshrrev_b32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
+; GCN-NOT: v_alignbit_b32
+
+define amdgpu_kernel void @alignbit_shr_pat_wrong_const33(i64 addrspace(1)* nocapture readonly %arg, i32 addrspace(1)* nocapture %arg1) {
+bb:
+  %tmp = load i64, i64 addrspace(1)* %arg, align 8
+  %tmp5 = lshr i64 %tmp, 33
+  %tmp6 = trunc i64 %tmp5 to i32
+  store i32 %tmp6, i32 addrspace(1)* %arg1, align 4
+  ret void
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x() #0
 
 attributes #0 = { nounwind readnone speculatable }

test/CodeGen/AMDGPU/cgp-bitfield-extract.ll

@@ -176,14 +176,13 @@ ret:
 ; OPT: ret
 
 ; GCN-LABEL: {{^}}sink_ubfe_i64_span_midpoint:
-; GCN: s_cbranch_scc1 BB3_2
 
-; GCN: s_lshr_b64 s{{\[}}[[LO:[0-9]+]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 30
-; GCN: s_and_b32 s{{[0-9]+}}, s[[LO]], 0xff
+; GCN: v_alignbit_b32 v[[LO:[0-9]+]], s{{[0-9]+}}, v{{[0-9]+}}, 30
+; GCN: s_cbranch_scc1 BB3_2
+; GCN: v_and_b32_e32 v{{[0-9]+}}, 0xff, v[[LO]]
 
 ; GCN: BB3_2:
-; GCN: s_lshr_b64 s{{\[}}[[LO:[0-9]+]]:{{[0-9]+}}], s{{\[[0-9]+:[0-9]+\]}}, 30
-; GCN: s_and_b32 s{{[0-9]+}}, s[[LO]], 0x7f
+; GCN: v_and_b32_e32 v{{[0-9]+}}, 0x7f, v[[LO]]
 
 ; GCN: BB3_3:
 ; GCN: buffer_store_dwordx2

test/CodeGen/AMDGPU/shift-and-i64-ubfe.ll

@@ -151,10 +151,11 @@ define amdgpu_kernel void @v_uextract_bit_1_31_i64(i64 addrspace(1)* %out, i64 a
   ret void
 }
 
-; Spans the dword boundary, so requires full shift
+; Spans the dword boundary, so requires full shift.
+; Truncated after the shift, so only low shift result is used.
 ; GCN-LABEL: {{^}}v_uextract_bit_31_32_i64:
-; GCN: buffer_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]]
-; GCN: v_lshr_b64 v{{\[}}[[SHRLO:[0-9]+]]:[[SHRHI:[0-9]+]]{{\]}}, [[VAL]], 31
+; GCN: buffer_load_dwordx2 v{{\[}}[[VALLO:[0-9]+]]:[[VALHI:[0-9]+]]{{\]}}
+; GCN: v_alignbit_b32 v[[SHRLO:[0-9]+]], v[[VALHI]], v[[VALLO]], 31
 ; GCN-DAG: v_and_b32_e32 v[[AND:[0-9]+]], 3, v[[SHRLO]]{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[ZERO:[0-9]+]], 0{{$}}
 ; GCN: buffer_store_dwordx2 v{{\[}}[[AND]]:[[ZERO]]{{\]}}
@@ -188,8 +189,8 @@ define amdgpu_kernel void @v_uextract_bit_32_33_i64(i64 addrspace(1)* %out, i64
 
 ; GCN-LABEL: {{^}}v_uextract_bit_30_60_i64:
 ; GCN-DAG: v_mov_b32_e32 v[[ZERO:[0-9]+]], 0{{$}}
-; GCN: buffer_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]]
-; GCN: v_lshr_b64 v{{\[}}[[SHRLO:[0-9]+]]:[[SHRHI:[0-9]+]]{{\]}}, [[VAL]], 30
+; GCN: buffer_load_dwordx2 v{{\[}}[[VALLO:[0-9]+]]:[[VALHI:[0-9]+]]{{\]}}
+; GCN: v_alignbit_b32 v[[SHRLO:[0-9]+]], v[[VALHI]], v[[VALLO]], 30
 ; GCN-DAG: v_and_b32_e32 v[[AND:[0-9]+]], 0x3fffffff, v[[SHRLO]]{{$}}
 ; GCN-DAG: v_mov_b32_e32 v[[ZERO1:[0-9]+]], v[[ZERO]]
 ; GCN: buffer_store_dwordx2 v{{\[}}[[AND]]:[[ZERO1]]{{\]}}
@@ -223,10 +224,9 @@ define amdgpu_kernel void @v_uextract_bit_33_63_i64(i64 addrspace(1)* %out, i64
 
 ; GCN-LABEL: {{^}}v_uextract_bit_31_63_i64:
 ; GCN: v_mov_b32_e32 v[[ZERO:[0-9]+]], 0{{$}}
-; GCN: buffer_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]]
-; GCN: v_lshr_b64 v{{\[}}[[SHRLO:[0-9]+]]:[[SHRHI:[0-9]+]]{{\]}}, [[VAL]], 31
-; GCN-NEXT: v_mov_b32_e32 v[[SHRHI]], v[[ZERO]]
-; GCN: buffer_store_dwordx2 v{{\[}}[[SHRLO]]:[[SHRHI]]{{\]}}
+; GCN: buffer_load_dwordx2 v{{\[}}[[VALLO:[0-9]+]]:[[VALHI:[0-9]+]]{{\]}}
+; GCN: v_alignbit_b32 v[[SHRLO:[0-9]+]], v[[VALHI]], v[[VALLO]], 31
+; GCN: buffer_store_dwordx2 v{{\[}}[[SHRLO]]:[[ZERO]]{{\]}}
 define amdgpu_kernel void @v_uextract_bit_31_63_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in) #1 {
   %id.x = tail call i32 @llvm.amdgcn.workitem.id.x()
   %in.gep = getelementptr i64, i64 addrspace(1)* %in, i32 %id.x
@@ -288,8 +288,8 @@ define amdgpu_kernel void @v_uextract_bit_33_i64_trunc_i32(i32 addrspace(1)* %ou
 }
 
 ; GCN-LABEL: {{^}}v_uextract_bit_31_32_i64_trunc_i32:
-; GCN: buffer_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]]
-; GCN: v_lshr_b64 v{{\[}}[[SHRLO:[0-9]+]]:[[SHRHI:[0-9]+]]{{\]}}, [[VAL]], 31
+; GCN: buffer_load_dwordx2 v{{\[}}[[VALLO:[0-9]+]]:[[VALHI:[0-9]+]]{{\]}}
+; GCN: v_alignbit_b32 v[[SHRLO:[0-9]+]], v[[VALHI]], v[[VALLO]], 31
 ; GCN-NEXT: v_and_b32_e32 v[[SHRLO]], 3, v[[SHRLO]]
 ; GCN-NOT: v[[SHRLO]]
 ; GCN: buffer_store_dword v[[SHRLO]]