Instead of using parameterised rsts for stack access, add a huge swathe of

automatically built helper tools. Star Trek goes up from 40243 to 40779 bytes,
but should be a lot faster.

mach/i80/libem/build.lua

@@ -1,8 +1,68 @@
-for _, plat in ipairs(vars.plats) do
+local generated = {}
-	acklibrary {
+
-		name = "lib_"..plat,
+definerule("generate",
-		srcs = { "./*.s" },
+	{
-		vars = { plat = plat },
+		body = { type="string" },
+		offset = { type="object" },
+	},
+	function(e)
+		return normalrule {
+			name = e.name,
+			ins = { "./generate.sh", "./"..e.body },
+			outleaves = { e.name..".s" },
+			commands = {
+				"%{ins[1]} "..e.body.." "..e.offset.." > %{outs}"
+			}
+		}
+	end
+)
+		
+for i = 1, 128 do
+	generated[#generated+1] = generate {
+		name = "fload"..i,
+		body = "fload.h",
+		offset = i
+	}
+
+	generated[#generated+1] = generate {
+		name = "floadn"..i,
+		body = "floadn.h",
+		offset = i
+	}
+
+	generated[#generated+1] = generate {
+		name = "fstore"..i,
+		body = "fstore.h",
+		offset = i
+	}
+
+	generated[#generated+1] = generate {
+		name = "fstoren"..i,
+		body = "fstoren.h",
+		offset = i
+	}
+
+	generated[#generated+1] = generate {
+		name = "faddr"..i,
+		body = "faddr.h",
+		offset = i
+	}
+
+	generated[#generated+1] = generate {
+		name = "faddrn"..i,
+		body = "faddrn.h",
+		offset = i
+	}
+end
+
+for _, plat in ipairs(vars.plats) do
+	acklibrary {
+		name = "lib_"..plat,
+		srcs = concat("./*.s", generated),
+		vars = {
+			plat = plat,
+			["+ackcflags"] = {"-Imach/i80/libem"}
+		},
 	}
 end
 

mach/i80/libem/faddr.h

@@ -0,0 +1,30 @@
+.sect .text
+.sect .rom
+.sect .data
+.sect .bss
+.sect .text
+
+! Fetches the word at positive stack offset OFFSET into de.
+
+#define PASTE(a, b) a ## b
+#define LABEL(prefix, offset) PASTE(prefix, offset)
+
+.define LABEL(.faddr, OFFSET)
+LABEL(.faddr, OFFSET):
+    #if OFFSET == 0
+        mov l, c
+        mov h, b
+    #elif OFFSET == 1
+        mov l, c
+        mov h, b
+        inx h
+    #elif OFFSET == 2
+        mov l, c
+        mov h, b
+        inx h
+        inx h
+    #else
+        lxi h, OFFSET
+        dad b
+    #endif
+    ret

mach/i80/libem/faddrn.h

@@ -0,0 +1,30 @@
+.sect .text
+.sect .rom
+.sect .data
+.sect .bss
+.sect .text
+
+! Fetches the word at negative stack offset OFFSET into de.
+
+#define PASTE(a, b) a ## b
+#define LABEL(prefix, offset) PASTE(prefix, offset)
+
+.define LABEL(.faddrn, OFFSET)
+LABEL(.faddrn, OFFSET):
+    #if OFFSET == 0
+        mov l, c
+        mov h, b
+    #elif OFFSET == 1
+        mov l, c
+        mov h, b
+        dcx h
+    #elif OFFSET == 2
+        mov l, c
+        mov h, b
+        dcx h
+        dcx h
+    #else
+        lxi h, -OFFSET
+        dad b
+    #endif
+    ret

mach/i80/libem/fload.h

@@ -0,0 +1,32 @@
+.sect .text
+.sect .rom
+.sect .data
+.sect .bss
+.sect .text
+
+! Fetches the word at positive stack offset OFFSET into de.
+
+#define PASTE(a, b) a ## b
+#define LABEL(prefix, offset) PASTE(prefix, offset)
+
+.define LABEL(.fload, OFFSET)
+LABEL(.fload, OFFSET):
+    #if OFFSET == 0
+        #error "0 shouldn't happen"
+    #elif OFFSET == 1
+        mov l, c
+        mov h, b
+        inx h
+    #elif OFFSET == 2
+        mov l, c
+        mov h, b
+        inx h
+        inx h
+    #else
+        lxi h, OFFSET
+        dad b
+    #endif
+    mov e, m
+    inx h
+    mov d, m
+    ret

mach/i80/libem/floadn.h

@@ -0,0 +1,38 @@
+.sect .text
+.sect .rom
+.sect .data
+.sect .bss
+.sect .text
+
+! Fetches the word at negative stack offset OFFSET into de.
+
+#define PASTE(a, b) a ## b
+#define LABEL(prefix, offset) PASTE(prefix, offset)
+
+.define LABEL(.floadn, OFFSET)
+LABEL(.floadn, OFFSET):
+    #if OFFSET == 0
+        #error "0 shouldn't happen"
+    #elif OFFSET == 1
+        mov l, c
+        mov h, b
+        mov d, m
+        dcx h
+        mov e, m
+        ret
+    #elif OFFSET == 2
+        mov l, c
+        mov h, b
+        dcx h
+        mov d, m
+        dcx h
+        mov e, m
+        ret
+    #else
+        lxi h, -OFFSET
+        dad b
+        mov e, m
+        inx h
+        mov d, m
+        ret
+    #endif

mach/i80/libem/fstore.h

@@ -0,0 +1,32 @@
+.sect .text
+.sect .rom
+.sect .data
+.sect .bss
+.sect .text
+
+! Stores the word in de to positive stack offset OFFSET.
+
+#define PASTE(a, b) a ## b
+#define LABEL(prefix, offset) PASTE(prefix, offset)
+
+.define LABEL(.fstore, OFFSET)
+LABEL(.fstore, OFFSET):
+    #if OFFSET == 0
+        #error "0 shouldn't happen"
+    #elif OFFSET == 1
+        mov l, c
+        mov h, b
+        inx h
+    #elif OFFSET == 2
+        mov l, c
+        mov h, b
+        inx h
+        inx h
+    #else
+        lxi h, OFFSET
+        dad b
+    #endif
+    mov m, e
+    inx h
+    mov m, d
+    ret

mach/i80/libem/fstoren.h

@@ -0,0 +1,39 @@
+.sect .text
+.sect .rom
+.sect .data
+.sect .bss
+.sect .text
+
+! Stores the word in de to negative stack offset OFFSET.
+
+#define PASTE(a, b) a ## b
+#define LABEL(prefix, offset) PASTE(prefix, offset)
+
+.define LABEL(.fstoren, OFFSET)
+LABEL(.fstoren, OFFSET):
+    #if OFFSET == 0
+        #error "0 shouldn't happen"
+    #elif OFFSET == 1
+        mov l, c
+        mov h, b
+        mov m, d
+        dcx h
+        mov m, e
+        ret
+    #elif OFFSET == 2
+        mov l, c
+        mov h, b
+        dcx h
+        mov m, d
+        dcx h
+        mov m, e
+        ret
+    #else
+        lxi h, -OFFSET
+        dad b
+        mov m, e
+        inx h
+        mov m, d
+        ret
+    #endif
+

mach/i80/libem/generate.sh

@@ -0,0 +1,4 @@
+#!/bin/sh
+echo "#"
+echo "#define OFFSET $2"
+echo "#include \"$1\""

mach/i80/libem/rst.s

@@ -5,68 +5,36 @@
 .sect .bss
 .sect .text
 
+! Which resets we install are determined by statistical analysis of Star
+! Trek. When changing these, make sure to update the i80 table to match.
+!     97 call .floadn2
+!     41 call .floadn4
+!     34 call .fload4
+!     28 call .fstoren2
+
 .define .rst_init
 .rst_init:
-    mvi a, 0xc3     ! jmp <a16>
+    lxi h, .floadn2
-    sta 0x08
+    lxi d, 0x0008
-    sta 0x10
+    call copy
-    sta 0x18
+    lxi h, .floadn4
-    lxi h, rst1
+    lxi d, 0x0010
-    shld 0x09
+    call copy
-    lxi h, rst2
+    lxi h, .fload4
-    shld 0x11
+    lxi d, 0x0018
-    lxi h, rst3
+    call copy
-    shld 0x19
+    lxi h, .fstoren2
-    ret
+    lxi d, 0x0020
+    jmp copy
 
-    ! de = [bc+const1] (remember bc is the frame pointer)
+! Copies eight bytes from HL to DE.
-rst1:
+copy:
-    pop h
+    mvi c, 8
+.1:
     mov a, m
+    stax d
     inx h
-    push h
+    inx d
-
+    dcr c
-	mov l, a
+    jnz .1
-	ral
-	sbb a
-	mov h, a
-
-	dad b
-    mov e, m
-    inx h
-    mov d, m
     ret
-
-    ! [bc+const1] = de (remember bc is the frame pointer)
-rst2:
-    pop h
-    mov a, m
-    inx h
-    push h
-
-	mov l, a
-	ral
-	sbb a
-	mov h, a
-
-	dad b
-    mov m, e
-    inx h
-    mov m, d
-    ret
-
-    ! hl = bc+const1
-rst3:
-    pop h
-    mov a, m
-    inx h
-    push h
-
-    mov l, a
-    ral
-    sbb a
-    mov h, a
-
-    dad b
-    ret
-

mach/i80/ncg/table

@@ -15,6 +15,8 @@ EM_BSIZE = 4
 
 SL=4
 
+STACKHELPERS=128
+
 PROPERTIES
 
 areg			/* the a-register */
@@ -48,6 +50,7 @@ const1			= { INT num; } 1 num.
 const2			= { INT num; } 2 num.
 smallconst2    = { INT num; } 2 num.
 label			= { ADDR off; } 2 off.
+plabel      = { ADDR off; INT param; } 2 off param.
 m			= { } 2 cost(0,3) "m".
 
 SETS
@@ -69,6 +72,7 @@ INSTRUCTIONS
    ana	reg1:ro 	kills a:cc	cost(1, 4).
    ani	const1:ro 	kills a:cc	cost(2, 7).
    Call	"call" label:ro			cost(3,17).
+   Call	"call" plabel:ro			cost(3,17).
 			   /* 'call' is a reserved word */
 /* cc	label:ro			cost(3,14). 	*/
 /* cm	label:ro			cost(3,14).	*/
@@ -253,22 +257,44 @@ pat ldc
    yields {const2, highw($1)} {const2, loww($1)}
 
 #ifdef USE_I80_RSTS
-   pat lol sfit($1, 8)
+   pat lol $1==0-2
-      uses hlreg, areg, dereg
+      uses dereg, hlreg
       gen
          rst {const1, 1}
-         data1 {const1, $1}
+      yields de
+      
+   pat lol $1==0-4
+      uses dereg, hlreg
+      gen
+         rst {const1, 2}
+      yields de
+      
+   pat lol $1==4
+      uses dereg, hlreg
+      gen
+         rst {const1, 3}
       yields de
 #endif
 
+pat lol ($1>0) && ($1<=STACKHELPERS)
+   uses dereg, hlreg
+   gen
+      Call {plabel, ".fload", $1}
+   yields de
+
+pat lol ($1<0) && ($1>=0-STACKHELPERS)
+   uses dereg, hlreg
+   gen
+      Call {plabel, ".floadn", 0-$1}
+   yields de
+
 pat lol
-  uses hlreg={const2, $1}, dereg
+   uses dereg
-  gen
+   gen
-	 dad lb
+      dad lb
-	 mov e,{m}
+      mov e, {m}
-	 inx hl
+      inx hl
-	 mov d,{m}
+      mov d, {m}
-  yields de
 
 pat loe
    uses hlreg
@@ -286,14 +312,17 @@ pat lof
       adp $1
       loi 2
 
-#ifdef USE_I80_RSTS
+pat lal ($1>0) && ($1<=STACKHELPERS)
-   pat lal sfit($1, 8)
+   uses hlreg
-      uses dereg, hlreg, areg
+   gen
-      gen
+      Call {plabel, ".faddr", $1}
-         rst {const1, 3}
+   yields hl
-         data1 {const1, $1}
+
-      yields hl
+pat lal ($1<0) && ($1>=0-STACKHELPERS)
-#endif
+   uses hlreg
+   gen
+      Call {plabel, ".faddrn", 0-$1}
+   yields hl
 
 pat lal
    uses hlreg={const2,$1}
@@ -457,22 +486,33 @@ pat stl lol $1==$2
 with dereg				yields de de leaving stl $1
 
 #ifdef USE_I80_RSTS
-   pat stl sfit($1, 8)
+   pat stl $1==0-2
-	   with dereg
+      with dereg
-      uses hlreg, areg
+         uses hlreg
-      gen
+         gen
-         rst {const1, 2}
+            rst {const1, 4}
-         data1 {const1, $1}
 #endif
 
+pat stl ($1>0) && ($1<=STACKHELPERS)
+   with dereg
+      uses hlreg
+      gen
+         Call {plabel, ".fstore", $1}
+
+pat stl ($1<0) && ($1>=0-STACKHELPERS)
+   with dereg
+      uses hlreg
+      gen
+         Call {plabel, ".fstoren", 0-$1}
+
 pat stl
 	with dereg
-	uses hlreg={const2, $1}
+      uses hlreg={const2, $1}
-	gen
+      gen
-		dad lb
+         dad lb
-		mov {m}, e
+         mov {m}, e
-		inx hl
+         inx hl
-		mov {m}, d
+         mov {m}, d
 
 pat ste loe $1==$2
 with hlreg				yields hl hl leaving ste $1
@@ -906,18 +946,25 @@ pat inc
          inx %1
       yields %1
 
-#ifdef USE_I80_RSTS
+pat inl ($1>0) && ($1<STACKHELPERS)
-   pat inl sfit($1, 8)
+   uses hlreg
-      uses hlreg, areg
       gen
-         rst {const1, 3}
+         Call {plabel, ".faddr", $1}
-         data1 {const1, $1}
          inr {m}
-         jnz {label, 1f}
+         jnz {label,1f}
+         inx hl
+         inr {m}
+         1:
+
+pat inl ($1<0) && ($1<0-STACKHELPERS)
+   uses hlreg
+      gen
+         Call {plabel, ".faddrn", $1}
+         inr {m}
+         jnz {label,1f}
          inx hl
          inr {m}
          1:
-#endif
 
 pat inl
    uses hlreg={const2,$1}
@@ -941,12 +988,10 @@ pat dec
 with hl_or_de
 gen dcx %1				yields %1
 
-#ifdef USE_I80_RSTS
+pat del
-   pat del sfit($1, 8)
+   uses hlreg={const2,$1}
-      uses hlreg, areg, dereg
       gen
-         rst {const1, 3}
+         dad lb
-         data1 {const1, $1}
          mov e, {m}
          inx hl
          mov d, {m}
@@ -954,18 +999,6 @@ gen dcx %1				yields %1
          mov {m}, d
          dcx hl
          mov {m}, e
-#endif
-
-pat del
-   uses hlreg={const2,$1}, dereg
-   gen dad lb
-      mov e,{m}
-      inx hl
-      mov d,{m}
-      dcx de
-      mov {m},d
-      dcx hl
-      mov {m},e
 
 pat dee
 uses hlreg
@@ -973,18 +1006,6 @@ gen lhld {label,$1}
     dcx hl
     shld {label,$1}
 
-#ifdef USE_I80_RSTS
-   pat zrl sfit($1, 8)
-      uses hlreg, areg
-      gen
-         rst {const1, 3}
-         data1 {const1, $1}
-         xra a
-         mov {m}, a
-         inx hl
-         mov {m}, a
-#endif
-
 pat zrl
    uses hlreg={const2,$1}, areg
       gen
@@ -2046,19 +2067,6 @@ gen xra a
     jnz {label,$1}
     1:
 
-#ifdef USE_I80_RSTS
-   pat lol zeq sfit($1, 8)
-      with STACK
-         uses hlreg, areg
-         gen
-            rst {const1, 3}
-            data1 {const1, $1}
-            mov a, {m}
-            inx hl
-            ora {m}
-            jz {label, $2}
-#endif
-
 pat lol zeq
    with STACK
       uses hlreg={const2,$1}, areg
@@ -2069,19 +2077,6 @@ pat lol zeq
          ora {m}
          jz {label,$2}
 
-#ifdef USE_I80_RSTS
-   pat lol zne sfit($1, 8)
-      with STACK
-         uses hlreg, areg
-         gen
-            rst {const1, 3}
-            data1 {const1, $1}
-            mov a, {m}
-            inx hl
-            ora {m}
-            jnz {label, $2}
-#endif
-
 pat lol zne
    with STACK
       uses hlreg={const2,$1}, areg

plat/cpm/descr

@@ -76,6 +76,7 @@ name led
 		{FLOATS?} \
 		(.e:{TAIL}={PLATFORMDIR}/libem.a \
 		           {PLATFORMDIR}/libsys.a \
+				   {PLATFORMDIR}/libem.a \
 		           {PLATFORMDIR}/libend.a)
 	linker
 end