From c0727315e06d459832f0485e085ac22b2037e277 Mon Sep 17 00:00:00 2001
From: Stephen Heumann <stephenheumann@gmail.com>
Date: Fri, 5 Mar 2021 21:59:18 -0600
Subject: [PATCH] Recognize byte swapping and generate an xba instruction for
 it.

Specifically, this recognizes the pattern "(exp << 8) | (exp >> 8)", where exp has an unsigned 16-bit type and does not have side effects.
---
 CGI.Comments |  8 ++++++++
 CGI.Debug    |  1 +
 CGI.pas      |  2 +-
 DAG.pas      | 21 +++++++++++++++++----
 Gen.pas      | 13 ++++++++++++-
 5 files changed, 39 insertions(+), 6 deletions(-)

diff --git a/CGI.Comments b/CGI.Comments
index 7142d91..b2dd33f 100644
--- a/CGI.Comments
+++ b/CGI.Comments
@@ -551,6 +551,14 @@
 {  source address onto the stack.                               }
 {                                                               }
 {                                                               }
+{  pc_rbo - reverse byte order                                  }
+{                                                               }
+{  Gen0(pc_rbo)         cgWord,cgUWord                          }
+{                                                               }
+{  The value on the top of the evaluation stack is removed, has }
+{  the order of its constituent bytes reversed, and is replaced.}
+{                                                               }
+{                                                               }
 {  pc_sbf - save bit field                                      }
 {                                                               }
 {  Gen2t(pc_sbf, disp, size, type)                              }
diff --git a/CGI.Debug b/CGI.Debug
index fc52fe6..d8a31b7 100644
--- a/CGI.Debug
+++ b/CGI.Debug
@@ -127,6 +127,7 @@ opt[pc_dvq] := 'dvq';
 opt[pc_udq] := 'udq';
 opt[pc_mdq] := 'mdq';
 opt[pc_uqm] := 'uqm';
+opt[pc_rbo] := 'rbo';
 end; {InitWriteCode}
 
 
diff --git a/CGI.pas b/CGI.pas
index 4ca07ef..912d4f7 100644
--- a/CGI.pas
+++ b/CGI.pas
@@ -242,7 +242,7 @@ type
        pc_gli,pc_gdl,pc_gld,pc_cpi,pc_tri,pc_lbu,pc_lbf,pc_sbf,pc_cbf,dc_cns,
        dc_prm,pc_nat,pc_bno,pc_nop,pc_psh,pc_ili,pc_iil,pc_ild,pc_idl,
        pc_bqr,pc_bqx,pc_baq,pc_bnq,pc_ngq,pc_adq,pc_sbq,pc_mpq,pc_umq,pc_dvq,
-       pc_udq,pc_mdq,pc_uqm,pc_slq,pc_sqr,pc_wsr);
+       pc_udq,pc_mdq,pc_uqm,pc_slq,pc_sqr,pc_wsr,pc_rbo);
 
                                         {intermediate code}
                                         {-----------------}
diff --git a/DAG.pas b/DAG.pas
index 7613c4a..bfe0458 100644
--- a/DAG.pas
+++ b/DAG.pas
@@ -1100,7 +1100,20 @@ case op^.opcode of			{check for optimizations of this node}
             end {if}
          else if op^.right^.q = 0 then
             opv := op^.left;
-         end; {else if}
+         end {else if}
+      else if ((op^.left^.opcode = pc_shl) and (op^.right^.opcode = pc_usr))
+         or ((op^.left^.opcode = pc_usr) and (op^.right^.opcode = pc_shl)) then
+         if op^.left^.right^.opcode = pc_ldc then
+            if op^.right^.right^.opcode = pc_ldc then
+               if op^.left^.right^.q = 8 then
+                  if op^.right^.right^.q = 8 then
+                     if CodesMatch(op^.left^.left, op^.right^.left, false) then
+                        if not SideEffects(op^.left^.left) then begin
+                           op^.opcode := pc_rbo;
+                           op^.left := op^.left^.left;
+                           op^.right := nil;
+                           end; {if}
+         
       end; {case pc_bor}
 
    pc_bxr: begin			{pc_bxr}
@@ -2733,7 +2746,7 @@ case op^.opcode of
    pc_neq, pc_ior, pc_lor, pc_mod, pc_mpi, pc_sbi, pc_shl, pc_shr:
       TypeOf := cgWord;
 
-   pc_udi, pc_uim, pc_umi, pc_usr:
+   pc_udi, pc_uim, pc_umi, pc_usr, pc_rbo:
       TypeOf := cgUWord;
                          
    pc_bnl, pc_ngl, pc_adl, pc_bal, pc_blr, pc_blx, pc_dvl, pc_mdl,
@@ -4544,7 +4557,7 @@ var
                    pc_ngl,pc_ngr,pc_not,pc_pop,pc_sbi,pc_sbl,pc_sbr,
                    pc_shl,pc_sll,pc_shr,pc_usr,pc_slr,pc_vsr,pc_tri,
                    pc_bqr,pc_bqx,pc_baq,pc_bnq,pc_ngq,pc_adq,pc_sbq,
-                   pc_mpq,pc_umq,pc_dvq,pc_udq,pc_mdq,pc_uqm]
+                   pc_mpq,pc_umq,pc_dvq,pc_udq,pc_mdq,pc_uqm,pc_rbo]
                   then begin
         	  op^.parents := icount;
                   icount := icount+1;
@@ -5412,7 +5425,7 @@ case code^.opcode of
    pc_bnt, pc_bnl, pc_cnv, pc_dec, pc_inc, pc_ind, pc_lbf, pc_lbu,
    pc_ngi, pc_ngl, pc_ngr, pc_not, pc_stk, pc_cop, pc_cpo, pc_tl1,
    pc_sro, pc_str, pc_fjp, pc_tjp, pc_xjp, pc_cup, pc_pop, pc_iil,
-   pc_ili, pc_idl, pc_ild, pc_bnq, pc_ngq:
+   pc_ili, pc_idl, pc_ild, pc_bnq, pc_ngq, pc_rbo:
       begin
       code^.left := Pop;
       Push(code);
diff --git a/Gen.pas b/Gen.pas
index c027e31..32ac1c7 100644
--- a/Gen.pas
+++ b/Gen.pas
@@ -621,7 +621,7 @@ NeedsCondition := opcode in
    [pc_and,pc_ior,pc_cui,pc_cup,pc_lor,pc_lnd,pc_ldl,pc_lil,pc_lld,
     pc_lli,pc_gil,pc_gli,pc_gdl,pc_gld,pc_iil,pc_ili,pc_idl,pc_ild,
     pc_cop,pc_cpo,pc_cpi,pc_dvi,pc_mpi,pc_adi,pc_sbi,pc_mod,pc_bno,
-    pc_udi,pc_uim,pc_umi,pc_cnv];
+    pc_udi,pc_uim,pc_umi,pc_cnv,pc_rbo];
 end; {NeedsCondition}
 
 
@@ -6589,6 +6589,16 @@ procedure GenTree {op: icptr};
    end; {GenPsh}
 
 
+   procedure GenRbo (op: icptr);
+
+   { Generate code for a pc_rbo					}
+
+   begin {GenRbo}
+   GenTree(op^.left);
+   GenImplied(m_xba);
+   end; {GenRbo}
+
+
    procedure GenRealBinOp (op: icptr);
 
    { Generate code for a pc_adr, pc_dvr, pc_mpr or pc_sbr	}
@@ -7168,6 +7178,7 @@ case op^.opcode of
    pc_nop: ;
    pc_pop: GenPop(op);                
    pc_psh: GenPsh(op);
+   pc_rbo: GenRbo(op);
    pc_ret: GenRet(op);
    pc_sbf,pc_cbf: GenSbfCbf(op);
    pc_sbi: GenSbi(op);