From d2ec5e271c1ceb283eacb91c9c236b1ef268c873 Mon Sep 17 00:00:00 2001
From: Virgil Dupras <hsoft@hardcoded.net>
Date: Wed, 15 Apr 2020 16:09:10 -0400
Subject: [PATCH] Optimize boot binary's word execution code

Should be significantly faster now. Less push/popping and usage of
DE instead of IY.
---
 emul/forth/z80c.bin | Bin 2170 -> 2153 bytes
 forth/boot.fs       |  41 +++++++++++++++++++++--------------------
 forth/z80c.fs       |   2 ++
 3 files changed, 23 insertions(+), 20 deletions(-)

diff --git a/emul/forth/z80c.bin b/emul/forth/z80c.bin
index 2ca9802e6fec60c4f7eeca6937f6091f9fe87e74..75b5f60df14c6cefc508e1bdb60b40a17531ea0b 100644
GIT binary patch
delta 1242
zcmYjQOKTHR6h8OPWRk`@ZB4PYB92Yard1JNi&|eDM=i98PNy2F59~vzf`%%Hh(s4I
z3rh6|BpX*QiiNIqtcVaW3q?d1#Rn;|;!*^O-?@{jvq<KC=iKj{?>lFHH7+*p@*k1^
zlGNkh6#hlw*9pH_R2Iok34e#UOVe~0VLw&&b0YX3y}#?;{TCy-^<rsQmWvwsJLpR%
zmAmlFEln=GZvFi6%R&?Vm|JpRx4tW`qhc>ojq=E)SZry~D4&g=sGMC*mDpmZQ7$%7
z1wg#rD1+yX-$rjL@i-9QBq!t8L!M)aa7zg^h`VWT1!*k{rKiOwb7D^B3eu*-{RfW>
zniMiEicm(ksUv%S^6GeJvWX(Vq6$0exXwz_an{w@9&%kL<#h%*$N$F~DQxETHVs4<
zCN7U9SCV1Q;#?A&Kz@9v&udUbL=!nZs(sMfsW0U@?lIT#=(#>GriD!hd(*C$KvS=@
zx45-8Lc)ybQ)ss*c(tfnjf8C4SFa)#3F70q?AZCy(Ize_EZRV^P!<a2@h-?Fi?-2>
z<Or!qE(letHZfR7?SIo5D*=nrGg**)VcD3m<S>kx72O)&*@CiJYOR$h$Oj?L&5IdX
zf((8yiQS=D!ELOl@yrWm7F5Wgmu@YOV~|p=20HR$9+!hTBrn3x2capMh4sJTh27IY
z5`(CR>1`UDwb_IRQ#eF5{<TO_K~BM2ByV1WWJbGcF_f*vP7dbww~+>_4HVH;Ly{f#
z(kIjIp%!=;GiJpZBPeY{VurrM7LKk}BQ(*%b>owWq9*eO`B`MKRF2v-XdgS(h3vqv
z1B%T67n_4`q98N$6{N%Cm#<CUII|IAx!aQPjX;Y~6Js?cn93M0(~L|Q;8G{n*`cl{
zdvN`s@y)RRtyCk`0eY_UkhseUFipdLte>8XY8fp3EznFu+<WxVF>>X~wVvd55LvWU
zS<Ic+1_o7(amC(r+KVd}B#=C?_T`DGExrYWJ6*nkXh5x^!T3<0m%oqqK8r@U3HD}R
z*rxtIFLhe+Ta+*oENNiCIkca5n{~{{ZgAhGeoSo7D!OAPu%QX#vBuVR^G0-qc?JiJ
cK>cd0cztBKijLF_HAidOe})6~=XR3#198?Bb^rhX

delta 1254
zcmYjPO-vhS5S{Pa_1d7YKxhNRjkLiCVIWl%YN=}7Mq$}fM1o;if+O;y5Q3CMax9Wd
zb8&l!RH;$~>Ybj_a<&h-<-p2^9*`onrc$dO?8=EbIO?INNGs>fw<gIRc6a{X%zHD3
zJ+FIG{;%Z!M(Xi@6#jkTcMAVQ5gd}gApDegL_g6Zg#G!3za_%&#&&8;RJO}&1xqwn
zWn*8Xia$y9F=~wCdD`5pJ|Fn^pL(^I{(IVVpAY=Ge~R{}k!v*8?)LRHmyE`>#HC75
zg+IvA2nC75F{82HOF_MwNDLVbc!Ld52Wn9FPvRBiWU!&tgQ{<mlXdK6&q;}nXg@OC
zj2A8aES||iN!s-Fh4V8@CPhq(Vsu-#X*7Rh^Zu=|WG}_w>r}pP9M|b09k#A)d)alJ
zd2cMtS$yBuLs7G+k81$hxSd;1c9CJ0Vb{cG0AF05^|}=h(Zhls*P7Z8&CYv{yWlz=
zRrCX~D{MMHm2tgc4E4A65>;QuNSK5A4#rJ~zZSJxgAtp)I;J8P31ahle*MPHo4wpp
zSah0(BY7}5z`r1yEc%4@Bu7X^3Sp?#vWcA?H2F52eKq(x>33O@TyP}JdU6%Ql*OFJ
zHD}N^PXp~1CHW-6xka%jYoNjZH8B$@3m)U6HqL=yWI?3@W|`B9FvFBaE96lW2Ph8r
zkh}vupG0<K9@3kjXs5eLViCt6`nbkoO_<PN2S!}uw?&dlatGQXd14UB>`k|0Xxomh
zEEUe3Bn?ofX;4=KNmkg)tYqBjerPyslts=6Tic+RrB{%Jqj9xD6C?a!{4e6@$+%&D
z9$8G4qc$zs3zx@{9SXLn&<tR)I0Y`s(hGpD-pW1L{PyYy$nvx`@rMy|5o%-1rUp^j
z$H%lMQwFe<#XLt<c`}XiUyQ)8-<?zsb--28c}YCwLm=%!ew?3Ph-+CK{bRuFg4|!2
zV|4A_y$9*!SpZq|v67f4Z!Zk2*vB2GG8r$SRFLo#;Tt;In7ZP}@bIKZ6^Mq=DjKd2
z{rUK3`R?-=goj{lGEti@%zE=z6uw0%BgK?1UUa6<@oh7YJvj&L@8~Kv_Qi4f-W<k-
rri^DAOZz!rL^{elSTsWRCvf7YVn<zcq-N=Z&NlasSjhhRrzGA0H6|;&

diff --git a/forth/boot.fs b/forth/boot.fs
index 03c630e..99f832d 100644
--- a/forth/boot.fs
+++ b/forth/boot.fs
@@ -20,7 +20,7 @@ NOP,              ( 0a, unused )
 0 JPnn,           ( 0e, compiledWord )
 0 JPnn,           ( 11, pushRS )
 0 JPnn,           ( 14, popRS )
-JP(IY), NOP,      ( 17, nativeWord )
+EXDEHL, JP(HL), NOP, ( 17, nativeWord )
 0 JPnn,           ( 1a, next )
 0 JPnn,           ( 1d, chkPS )
 NOP, NOP,         ( 20, numberWord )
@@ -153,7 +153,6 @@ PC ORG @ 1 + ! ( main )
     EXDEHL,
     HL L1 @ LDddnn,
     0x03 CALLnn,        ( 03 == find )
-    DE PUSHqq,
     0x33 JPnn,          ( 33 == execute )
 
 PC ORG @ 4 + ! ( find )
@@ -261,7 +260,6 @@ L1 BSET ( abortUnderflow )
     HL PC 7 - LDddnn,
     DE RAMSTART 0x02 + LDdd(nn),   ( RAM+02 == CURRENT )
     0x03 CALLnn, ( find )
-    DE PUSHqq,
     0x33 JPnn,          ( 33 == execute )
 
 
@@ -305,17 +303,17 @@ PC ORG @ 0x1b + ! ( next )
     E (HL) LDrr,
     HL INCss,
     D (HL) LDrr,
-    DE PUSHqq,
     ( continue to execute )
 
 PC ORG @ 0x34 + ! ( execute )
-    IY POPqq,        ( is a wordref )
-    chkPS,
-    L 0 IY+ LDrIXY,
-    H 0 LDrn,
+    ( DE points to wordref )
+    EXDEHL,
+    E (HL) LDrr,
+    D 0 LDrn,
+    EXDEHL,
     ( HL points to code pointer )
-    IY INCss,
-    ( IY points to PFA )
+    DE INCss,
+    ( DE points to PFA )
     JP(HL),
 
 PC ORG @ 0x0f + ! ( compiledWord )
@@ -326,19 +324,19 @@ PC ORG @ 0x0f + ! ( compiledWord )
   3. Execute the first atom of the list. )
     RAMSTART 0x06 + LDHL(nn), ( RAMSTART+0x06 == IP )
     0x11 CALLnn,     ( 11 == pushRS )
-    IY PUSHqq, HL POPqq,
+    EXDEHL,          ( HL points to PFA )
+    ( While we increase, dereference into DE for execute call
+      later. )
+    E (HL) LDrr,
     HL INCss,
+    D (HL) LDrr,
     HL INCss,
     RAMSTART 0x06 + LD(nn)HL, ( RAMSTART+0x06 == IP )
-	( IY still is our atom reference )
-    L 0 IY+ LDrIXY,
-    H 1 IY+ LDrIXY,
-    HL PUSHqq,      ( arg for EXECUTE )
     0x33 JPnn,      ( 33 == execute )
 
 PC ORG @ 0x0c + ! ( cellWord )
 ( Pushes PFA directly )
-    IY PUSHqq,
+    DE PUSHqq,
     JPNEXT,
 
 PC ORG @ 0x2c + ! ( doesWord )
@@ -349,8 +347,11 @@ PC ORG @ 0x2c + ! ( doesWord )
   linkfrom the PFA, and then continue as a regular
   compiledWord.
 )
-    IY PUSHqq, ( like a regular cell )
-    L 2 IY+ LDrIXY,
-    H 3 IY+ LDrIXY,
-    HL PUSHqq, IY POPqq,
+    DE PUSHqq, ( like a regular cell )
+    EXDEHL,
+    HL INCss,
+    HL INCss,
+    E (HL) LDrr,
+    HL INCss,
+    D (HL) LDrr,
     0x0e JPnn, ( 0e == compiledWord )
diff --git a/forth/z80c.fs b/forth/z80c.fs
index 2a5014b..4d5fb1c 100644
--- a/forth/z80c.fs
+++ b/forth/z80c.fs
@@ -15,6 +15,8 @@
 )
 
 CODE EXECUTE
+    DE POPqq,
+    chkPS,
     0x33 JPnn,      ( 33 == execute )
 
 ( a b c -- b c a )