diff --git a/compiler/aarch64/cgcpu.pas b/compiler/aarch64/cgcpu.pas
index ad18f5c660..1f8183f4e1 100644
--- a/compiler/aarch64/cgcpu.pas
+++ b/compiler/aarch64/cgcpu.pas
@@ -84,7 +84,7 @@ interface
 
         procedure a_opmm_reg_reg(list: TAsmList; Op: TOpCG; size: tcgsize; src, dst: tregister; shuffle: pmmshuffle); override;
 
-        procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
+        procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
         { comparison operations }
         procedure a_cmp_const_reg_label(list: TAsmList; size: tcgsize; cmp_op: topcmp; a: tcgint; reg: tregister; l: tasmlabel);override;
         procedure a_cmp_reg_reg_label(list: TAsmList; size: tcgsize; cmp_op: topcmp; reg1, reg2: tregister; l: tasmlabel);override;
@@ -1292,7 +1292,7 @@ implementation
       end;
 
 
-    procedure tcgaarch64.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
+    procedure tcgaarch64.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
       var
         bitsize: longint;
       begin
@@ -1304,8 +1304,9 @@ implementation
           begin
             bitsize:=32;
           end;
-        { source is 0 -> dst will have to become 255 }
-        list.concat(taicpu.op_reg_const(A_CMP,src,0));
+        if not(not_zero) then
+          { source is 0 -> dst will have to become 255 }
+          list.concat(taicpu.op_reg_const(A_CMP,src,0));
         if reverse then
           begin
             list.Concat(taicpu.op_reg_reg(A_CLZ,makeregsize(dst,srcsize),src));
@@ -1319,10 +1320,13 @@ implementation
             list.Concat(taicpu.op_reg_reg(A_CLZ,dst,dst));
           end;
         { set dst to -1 if src was 0 }
-        list.Concat(taicpu.op_reg_reg_reg_cond(A_CSINV,dst,dst,makeregsize(NR_XZR,dstsize),C_NE));
-        { mask the -1 to 255 if src was 0 (anyone find a two-instruction
-          branch-free version? All of mine are 3...) }
-        list.Concat(taicpu.op_reg_reg(A_UXTB,makeregsize(dst,OS_32),makeregsize(dst,OS_32)));
+        if not(not_zero) then
+          begin
+            list.Concat(taicpu.op_reg_reg_reg_cond(A_CSINV,dst,dst,makeregsize(NR_XZR,dstsize),C_NE));
+            { mask the -1 to 255 if src was 0 (anyone find a two-instruction
+              branch-free version? All of mine are 3...) }
+            list.Concat(taicpu.op_reg_reg(A_UXTB,makeregsize(dst,OS_32),makeregsize(dst,OS_32)));
+          end;
       end;
 
 
diff --git a/compiler/arm/cgcpu.pas b/compiler/arm/cgcpu.pas
index 878ffebcd8..7161e9c00b 100644
--- a/compiler/arm/cgcpu.pas
+++ b/compiler/arm/cgcpu.pas
@@ -106,7 +106,7 @@ unit cgcpu;
 
         procedure a_opmm_reg_reg(list: TAsmList; Op: TOpCG; size : tcgsize;src,dst: tregister;shuffle : pmmshuffle); override;
         { Transform unsupported methods into Internal errors }
-        procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister); override;
+        procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister); override;
 
         { try to generate optimized 32 Bit multiplication, returns true if successful generated }
         function try_optimized_mul32_const_reg_reg(list: TAsmList; a: tcgint; src, dst: tregister) : boolean;
@@ -1827,7 +1827,7 @@ unit cgcpu;
       end;
 
 
-    procedure tbasecgarm.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister);
+    procedure tbasecgarm.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister);
       begin
         if reverse then
           begin
diff --git a/compiler/cghlcpu.pas b/compiler/cghlcpu.pas
index 50774ea1b2..0a5b8f06c1 100644
--- a/compiler/cghlcpu.pas
+++ b/compiler/cghlcpu.pas
@@ -65,14 +65,14 @@ uses
       procedure a_cmp_reg_reg_label(list: TAsmList; size: tcgsize; cmp_op: topcmp; reg1, reg2: tregister; l: tasmlabel); override;
       procedure a_call_reg(list: TAsmList; reg: tregister); override;
       procedure a_call_name(list: TAsmList; const s: string; weak: boolean); override;
-      procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
+      procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
     end;
 
 implementation
 
    { thlbasecgcpu }
 
-    procedure thlbasecgcpu.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
+    procedure thlbasecgcpu.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
       begin
         internalerror(2012042801);
       end;
diff --git a/compiler/cgobj.pas b/compiler/cgobj.pas
index aaf0d4e3ca..ac82da79ef 100644
--- a/compiler/cgobj.pas
+++ b/compiler/cgobj.pas
@@ -267,7 +267,7 @@ unit cgobj;
           procedure a_loadaddr_ref_reg(list : TAsmList;const ref : treference;r : tregister);virtual; abstract;
 
           { bit scan instructions }
-          procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); virtual;
+          procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); virtual;
 
           { Multiplication with doubling result size.
             dstlo or dsthi may be NR_NO, in which case corresponding half of result is discarded. }
@@ -3026,7 +3026,7 @@ implementation
       end;
 
 
-    procedure tcg.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
+    procedure tcg.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
       begin
         internalerror(2014070601);
       end;
diff --git a/compiler/hlcg2ll.pas b/compiler/hlcg2ll.pas
index 7a0c264162..3c1df0e16b 100644
--- a/compiler/hlcg2ll.pas
+++ b/compiler/hlcg2ll.pas
@@ -171,7 +171,7 @@ unit hlcg2ll;
           procedure a_loadaddr_ref_reg(list : TAsmList;fromsize, tosize : tdef;const ref : treference;r : tregister);override;
 
           { bit scan instructions }
-          procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tdef; src, dst: tregister); override;
+          procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tdef; src, dst: tregister); override;
 
           { fpu move instructions }
           procedure a_loadfpu_reg_reg(list: TAsmList; fromsize, tosize: tdef; reg1, reg2: tregister); override;
@@ -586,9 +586,9 @@ implementation
       cg.a_loadaddr_ref_reg(list,ref,r);
     end;
 
-  procedure thlcg2ll.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tdef; src, dst: tregister);
+  procedure thlcg2ll.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tdef; src, dst: tregister);
     begin
-      cg.a_bit_scan_reg_reg(list,reverse,def_cgsize(srcsize),def_cgsize(dstsize),src,dst);
+      cg.a_bit_scan_reg_reg(list,reverse,not_zero,def_cgsize(srcsize),def_cgsize(dstsize),src,dst);
     end;
 
   procedure thlcg2ll.a_loadfpu_reg_reg(list: TAsmList; fromsize, tosize: tdef; reg1, reg2: tregister);
diff --git a/compiler/hlcgobj.pas b/compiler/hlcgobj.pas
index ba6b3a935e..a27870759d 100644
--- a/compiler/hlcgobj.pas
+++ b/compiler/hlcgobj.pas
@@ -310,7 +310,7 @@ unit hlcgobj;
          public
 
           { bit scan instructions (still need transformation to thlcgobj) }
-          procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tdef; src, dst: tregister); virtual; abstract;
+          procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tdef; src, dst: tregister); virtual; abstract;
 
           { fpu move instructions }
           procedure a_loadfpu_reg_reg(list: TAsmList; fromsize, tosize: tdef; reg1, reg2: tregister); virtual; abstract;
diff --git a/compiler/jvm/hlcgcpu.pas b/compiler/jvm/hlcgcpu.pas
index 956e404d56..673e3d942e 100644
--- a/compiler/jvm/hlcgcpu.pas
+++ b/compiler/jvm/hlcgcpu.pas
@@ -114,7 +114,7 @@ uses
       procedure gen_exit_code(list: TAsmList); override;
 
       { unimplemented/unnecessary routines }
-      procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tdef; src, dst: tregister); override;
+      procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tdef; src, dst: tregister); override;
       procedure a_loadmm_loc_reg(list: TAsmList; fromsize, tosize: tdef; const loc: tlocation; const reg: tregister; shuffle: pmmshuffle); override;
       procedure a_loadmm_reg_reg(list: TAsmList; fromsize, tosize: tdef; reg1, reg2: tregister; shuffle: pmmshuffle); override;
       procedure a_loadmm_ref_reg(list: TAsmList; fromsize, tosize: tdef; const ref: treference; reg: tregister; shuffle: pmmshuffle); override;
@@ -1898,7 +1898,7 @@ implementation
       { nothing }
     end;
 
-  procedure thlcgjvm.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tdef; src, dst: tregister);
+  procedure thlcgjvm.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tdef; src, dst: tregister);
     begin
       internalerror(2012090201);
     end;
diff --git a/compiler/llvm/hlcgllvm.pas b/compiler/llvm/hlcgllvm.pas
index 11434345ac..3919a3f709 100644
--- a/compiler/llvm/hlcgllvm.pas
+++ b/compiler/llvm/hlcgllvm.pas
@@ -145,7 +145,7 @@ uses
 {$endif cpuflags}
 
       { unimplemented or unnecessary routines }
-      procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tdef; src, dst: tregister); override;
+      procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tdef; src, dst: tregister); override;
       procedure g_stackpointer_alloc(list: TAsmList; size: longint); override;
       procedure g_intf_wrapper(list: TAsmList; procdef: tprocdef; const labelname: string; ioffset: longint); override;
       procedure g_adjust_self_value(list: TAsmList; procdef: tprocdef; ioffset: aint); override;
@@ -2019,7 +2019,7 @@ implementation
     end;
 
 
-  procedure thlcgllvm.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tdef; src, dst: tregister);
+  procedure thlcgllvm.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tdef; src, dst: tregister);
     begin
       internalerror(2012090201);
     end;
diff --git a/compiler/loongarch64/cgcpu.pas b/compiler/loongarch64/cgcpu.pas
index fb6b526f59..de1295d180 100644
--- a/compiler/loongarch64/cgcpu.pas
+++ b/compiler/loongarch64/cgcpu.pas
@@ -57,7 +57,7 @@ unit cgcpu;
         procedure a_loadaddr_ref_reg(list : TAsmList;const ref : treference;r : tregister);override;
 
         { bit scan instructions }
-        procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
+        procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
 
         { fpu move instructions }
         procedure a_loadfpu_reg_reg(list: TAsmList; fromsize, tosize: tcgsize; reg1, reg2: tregister); override;
@@ -512,7 +512,7 @@ implementation
       end;
 
 
-    procedure tcgloongarch64.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
+    procedure tcgloongarch64.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
       begin
         internalerror(2022111943);
       end;
diff --git a/compiler/ncginl.pas b/compiler/ncginl.pas
index 3981ff45c9..14ddb11ad2 100644
--- a/compiler/ncginl.pas
+++ b/compiler/ncginl.pas
@@ -81,7 +81,7 @@ implementation
       aasmbase,aasmdata,
       cgbase,pass_2,
       cpubase,procinfo,
-      ncon,ncal,
+      nadd,ncon,ncal,
       tgobj,ncgutil,
       cgutils,cgobj,hlcgobj,
       defcmp
@@ -994,10 +994,13 @@ implementation
 
     procedure tcginlinenode.second_BsfBsr;
     var
+      not_zero,
       reverse: boolean;
       opsize: tcgsize;
     begin
       reverse:=(inlinenumber = in_bsr_x);
+      not_zero:=(left.nodetype=orn) and (((is_constintnode(taddnode(left).left) and (tordconstnode(taddnode(left).left).value<>0))) or
+        ((is_constintnode(taddnode(left).right) and (tordconstnode(taddnode(left).right).value<>0))));
       secondpass(left);
 
       opsize:=tcgsize2unsigned[left.location.size];
@@ -1006,7 +1009,7 @@ implementation
 
       location_reset(location,LOC_REGISTER,def_cgsize(resultdef));
       location.register:=cg.getintregister(current_asmdata.CurrAsmList,location.size);
-      cg.a_bit_scan_reg_reg(current_asmdata.CurrAsmList,reverse,opsize,location.size,left.location.register,location.register);
+      cg.a_bit_scan_reg_reg(current_asmdata.CurrAsmList,reverse,not_zero,opsize,location.size,left.location.register,location.register);
     end;
 
 
diff --git a/compiler/ppcgen/cgppc.pas b/compiler/ppcgen/cgppc.pas
index 37d191471d..fe3c7b73fc 100644
--- a/compiler/ppcgen/cgppc.pas
+++ b/compiler/ppcgen/cgppc.pas
@@ -35,7 +35,7 @@ unit cgppc;
       tcgppcgen = class(tcg)
         procedure a_loadaddr_ref_cgpara(list : TAsmList;const r : treference;const paraloc : tcgpara); override;
 
-        procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
+        procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
 
         procedure a_call_reg(list : TAsmList;reg: tregister); override;
 
@@ -205,7 +205,7 @@ unit cgppc;
       end;
 
 
-    procedure tcgppcgen.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
+    procedure tcgppcgen.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
       var
         tmpreg: tregister;
         cntlzop: tasmop;
diff --git a/compiler/riscv/cgrv.pas b/compiler/riscv/cgrv.pas
index 8f313fc1ad..74732e5504 100644
--- a/compiler/riscv/cgrv.pas
+++ b/compiler/riscv/cgrv.pas
@@ -35,7 +35,7 @@ unit cgrv;
       tcgrv = class(tcg)
         procedure a_loadaddr_ref_cgpara(list : TAsmList;const r : treference;const paraloc : tcgpara); override;
 
-        procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
+        procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister); override;
 
         procedure a_call_reg(list : TAsmList;reg: tregister); override;
         procedure a_call_name(list : TAsmList;const s : string; weak: boolean); override;
@@ -196,7 +196,7 @@ unit cgrv;
       end;
 
 
-    procedure tcgrv.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
+    procedure tcgrv.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tcgsize; src, dst: TRegister);
       begin
         internalerror(2016060401);
       end;       
diff --git a/compiler/wasm32/hlcgcpu.pas b/compiler/wasm32/hlcgcpu.pas
index 2a0cef59ed..59f4517aa9 100644
--- a/compiler/wasm32/hlcgcpu.pas
+++ b/compiler/wasm32/hlcgcpu.pas
@@ -133,7 +133,7 @@ uses
       procedure gen_stack_check_call(list: TAsmList); override;
 
       { unimplemented/unnecessary routines }
-      procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tdef; src, dst: tregister); override;
+      procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tdef; src, dst: tregister); override;
       procedure a_loadmm_loc_reg(list: TAsmList; fromsize, tosize: tdef; const loc: tlocation; const reg: tregister; shuffle: pmmshuffle); override;
       procedure a_loadmm_reg_reg(list: TAsmList; fromsize, tosize: tdef; reg1, reg2: tregister; shuffle: pmmshuffle); override;
       procedure a_loadmm_ref_reg(list: TAsmList; fromsize, tosize: tdef; const ref: treference; reg: tregister; shuffle: pmmshuffle); override;
@@ -2446,7 +2446,7 @@ implementation
       inherited;
     end;
 
-  procedure thlcgwasm.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: tdef; src, dst: tregister);
+  procedure thlcgwasm.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: tdef; src, dst: tregister);
     begin
       internalerror(2012090201);
     end;
diff --git a/compiler/x86/cgx86.pas b/compiler/x86/cgx86.pas
index 5229fd2e79..8625dced7d 100644
--- a/compiler/x86/cgx86.pas
+++ b/compiler/x86/cgx86.pas
@@ -84,7 +84,7 @@ unit cgx86;
         procedure a_loadaddr_ref_reg(list : TAsmList;const ref : treference;r : tregister);override;
 
         { bit scan instructions }
-        procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister); override;
+        procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister); override;
 
         { fpu move instructions }
         procedure a_loadfpu_reg_reg(list: TAsmList; fromsize, tosize: tcgsize; reg1, reg2: tregister); override;
@@ -2473,7 +2473,7 @@ unit cgx86;
         list.concat(taicpu.op_ref(TOpCG2AsmOp[op],tcgsize2opsize[size],tmpref));
       end;
 
-     procedure tcgx86.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister);
+     procedure tcgx86.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister);
      var
        tmpreg: tregister;
        opsize: topsize;
diff --git a/compiler/xtensa/cgcpu.pas b/compiler/xtensa/cgcpu.pas
index e933b8a300..5268fe0dd7 100644
--- a/compiler/xtensa/cgcpu.pas
+++ b/compiler/xtensa/cgcpu.pas
@@ -68,7 +68,7 @@ interface
         procedure a_cmp_reg_reg_label(list: TAsmList; size: tcgsize; cmp_op: topcmp; reg1, reg2: tregister; l: tasmlabel);override;
         procedure a_jmp_always(list: TAsmList; l: TAsmLabel);override;
 
-        procedure a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister);override;
+        procedure a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister);override;
 
         procedure g_flags2reg(list: TAsmList; size: TCgSize; const f: tresflags; reg: TRegister);override;
 
@@ -1320,7 +1320,7 @@ implementation
       end;
 
 
-    procedure tcgcpu.a_bit_scan_reg_reg(list: TAsmList; reverse: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister);
+    procedure tcgcpu.a_bit_scan_reg_reg(list: TAsmList; reverse,not_zero: boolean; srcsize, dstsize: TCGSize; src, dst: TRegister);
       var
         ai: taicpu;
         tmpreg: TRegister;