[AMDGPU] Add pattern for v_alignbit_b32 with immediate

[rampitec]

Summary

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

Diff is based on D34723.

I'm also removing alignbit pattern for the left shift as it is
covered by the better optimization in D34723.

[arsenm]

LGTM

[rampitec]

Rebased to master.

lib/Target/AMDGPU/SIInstrInfo.td

 def NegSubInlineConst32 : ImmLeaf<i32, [{
   return Imm < -16 && Imm >= -64;
 }], NegateImm>;
 
 def NegSubInlineConst16 : ImmLeaf<i16, [{
   return Imm < -16 && Imm >= -64;
 }], NegateImm>;
 
+def ShiftAmt32Imm : PatLeaf <(imm), [{
+  return N->getZExtValue() < 32;
+}]>;
+
 //===----------------------------------------------------------------------===//
 // Custom Operands
 //===----------------------------------------------------------------------===//
 
 def SoppBrTarget : AsmOperandClass {
   let Name = "SoppBrTarget";
   let ParserMethod = "parseSOppBrTarget";
 }

lib/Target/AMDGPU/SIInstructions.td

 
 defm : BFIPatterns <V_BFI_B32, S_MOV_B32, SReg_64>;
 def : ROTRPattern <V_ALIGNBIT_B32>;
 
 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  **********/
 /********** ====================== **********/
 
 multiclass SI_INDIRECT_Pattern <ValueType vt, ValueType eltvt, string VecSize> {
   // Extract with offset
   def : Pat<

test/CodeGen/AMDGPU/alignbit-pat.ll

bb:
   %tmp3 = and i32 %arg2, 63
   %tmp4 = zext i32 %tmp3 to i64
   %tmp5 = lshr i64 %tmp, %tmp4
   %tmp6 = trunc i64 %tmp5 to i32
   store i32 %tmp6, i32 addrspace(1)* %arg1, align 4
   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

 ; OPT: %1 = lshr i64 %arg1, 30
 ; OPT-NEXT: %val1 = and i64 %1, 127
 
 ; OPT: ret:
 ; OPT: store
 ; 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
 define amdgpu_kernel void @sink_ubfe_i64_span_midpoint(i64 addrspace(1)* %out, i64 %arg1) #0 {
 entry:
   %shr = lshr i64 %arg1, 30
   br i1 undef, label %bb0, label %bb1
 

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

define amdgpu_kernel void @v_uextract_bit_1_31_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in) #1 {
   %out.gep = getelementptr i64, i64 addrspace(1)* %out, i32 %id.x
   %ld.64 = load i64, i64 addrspace(1)* %in.gep
   %srl = lshr i64 %ld.64, 1
   %bit = and i64 %srl, 2147483647
   store i64 %bit, i64 addrspace(1)* %out.gep
   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{{\[}}[[AND]]:[[ZERO]]{{\]}}
 define amdgpu_kernel void @v_uextract_bit_31_32_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in) #1 {
   %id.x = tail call i32 @llvm.amdgcn.workgroup.id.x()
   %in.gep = getelementptr i64, i64 addrspace(1)* %in, i32 %id.x
   %out.gep = getelementptr i64, i64 addrspace(1)* %out, i32 %id.x
   %ld.64 = load i64, i64 addrspace(1)* %in.gep
define amdgpu_kernel void @v_uextract_bit_32_33_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in) #1 {
   %srl = lshr i64 %ld.64, 33
   %bit = and i64 %srl, 3
   store i64 %bit, i64 addrspace(1)* %out.gep
   ret void
 }
 
 ; 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{{\[}}[[AND]]:[[ZERO1]]{{\]}}
 define amdgpu_kernel void @v_uextract_bit_30_60_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
   %out.gep = getelementptr i64, i64 addrspace(1)* %out, i32 %id.x
   %ld.64 = load i64, i64 addrspace(1)* %in.gep
define amdgpu_kernel void @v_uextract_bit_33_63_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in) #1 {
   %srl = lshr i64 %ld.64, 33
   %bit = and i64 %srl, 1073741823
   store i64 %bit, i64 addrspace(1)* %out.gep
   ret void
 }
 
 ; 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
   %out.gep = getelementptr i64, i64 addrspace(1)* %out, i32 %id.x
   %ld.64 = load i64, i64 addrspace(1)* %in.gep
   %srl = lshr i64 %ld.64, 31
   %and = and i64 %srl, 4294967295
   store i64 %and, i64 addrspace(1)* %out
define amdgpu_kernel void @v_uextract_bit_33_i64_trunc_i32(i32 addrspace(1)* %out, i64 addrspace(1)* %in) #1 {
   %srl = lshr i64 %ld.64, 33
   %trunc = trunc i64 %srl to i32
   %bit = and i32 %trunc, 1
   store i32 %bit, i32 addrspace(1)* %out.gep
   ret void
 }
 
 ; 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]]
 define amdgpu_kernel void @v_uextract_bit_31_32_i64_trunc_i32(i32 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
   %out.gep = getelementptr i32, i32 addrspace(1)* %out, i32 %id.x
   %ld.64 = load i64, i64 addrspace(1)* %in.gep