cpu-z80: peephole rules

diff --git a/Kernel/cpu-z80/rules.mk b/Kernel/cpu-z80/rules.mk
index 7345cc2..b5d8740 100644 (file)
--- a/Kernel/cpu-z80/rules.mk
+++ b/Kernel/cpu-z80/rules.mk
@@ -3,7 +3,7 @@ export CROSS_LD=tools/bankld/sdldz80
 export CROSS_CC=sdcc
 #export CROSS_CCOPTS=-c --std-sdcc99 --no-std-crt0 -m$(CPU) -I$(ROOT_DIR)/cpu-$(CPU) -I$(ROOT_DIR)/platform-$(TARGET) -I$(ROOT_DIR)/include --max-allocs-per-node 1000000 --opt-code-size --Werror --stack-auto --constseg CONST
 #export CROSS_CCOPTS=-c --std-sdcc99 --no-std-crt0 -m$(CPU) -I$(ROOT_DIR)/cpu-$(CPU) -I$(ROOT_DIR)/platform-$(TARGET) -I$(ROOT_DIR)/include --max-allocs-per-node 200000 --opt-code-size --Werror --stack-auto --constseg CONST
-export CROSS_CCOPTS=-c --std-sdcc99 --no-std-crt0 -m$(CPU) -I$(ROOT_DIR)/cpu-$(CPU) -I$(ROOT_DIR)/platform-$(TARGET) -I$(ROOT_DIR)/include --max-allocs-per-node 30000 --opt-code-size --Werror --stack-auto --constseg CONST
+export CROSS_CCOPTS=-c --std-sdcc99 --no-std-crt0 -m$(CPU) -I$(ROOT_DIR)/cpu-$(CPU) -I$(ROOT_DIR)/platform-$(TARGET) -I$(ROOT_DIR)/include --max-allocs-per-node 30000 --opt-code-size --Werror --stack-auto --constseg CONST --peep-file $(ROOT_DIR)/cpu-z80/switch.peep --fverbose-asm
 #export CROSS_CCOPTS+=--nostdlib --nostdinc -Isdcclib/include 
 export CROSS_CC_SEG2=--codeseg CODE2
 # For now but we are overspilling in a lot of configs so will need a real CODE3

diff --git a/Kernel/cpu-z80/switch.peep b/Kernel/cpu-z80/switch.peep
new file mode 100644 (file)
index 0000000..da61796
--- /dev/null
+++ b/Kernel/cpu-z80/switch.peep
@@ -0,0 +1,489 @@
+// Some slightly daft reloads
+
+replace restart {
+       ld      hl,#%1 + 0
+       ld      (hl), #<(%2)
+       ld      hl,#%1 + 1
+       ld      (hl), #>(%2)
+} by {
+       ld      hl,#%1 + 0
+       ld      (hl), #<(%2)
+       inc     hl
+       ld      (hl), #>(%2)
+       ; AC use inc for half loads
+}
+
+replace restart {
+       ld      hl,#%1+0
+       inc     (hl)
+       jr      NZ,%2
+       ld      hl,#%1+1
+       inc     (hl)
+} by {
+       ld      hl, #%1+0
+       inc     (hl)
+       jr      NZ,%2
+       inc     hl
+       inc     (hl)
+       ; AC suppress hl reload on ++x
+}
+
+replace restart {
+       ld      a,(#%1+1)
+       ld      hl,#%1+0
+} by {
+       ld      hl,#%1+1
+       ld      a,(hl)
+       dec     hl
+       ; AC suppress hl reload on word zero check
+}
+
+replace restart {
+       ld      a,#0x00
+       ld      (%1),a
+       inc     %1
+       ld      a,#0x00
+       ld      (%1),a
+} by {
+       xor     a,a
+       ld      (%1),a
+       inc     %1
+       ld      (%1),a
+       ; AC zero indirect store cleanup
+}
+
+replace restart {
+       ld      a,%1
+       ld      (%2),a
+       inc     %2
+       ld      a,%1
+} by {
+       ld      a,%1
+       ld      (%2),a
+       inc     %2
+       ; AC matching halves indirect store cleanup
+} if notVolatile(%1)
+
+//
+//     Should add rules for ix/iy switches with 0 or 1 base!!
+//
+replace restart {
+       ld      a,%1(%5)
+       sub     a,%2
+       jr      %3,%4
+       ld      a,%1(%5)
+       sub     a,%6
+       jr      %7,%8
+       ld      a,%1(%5)
+       sub     a,%9
+       jr      %10,%11
+       ld      a,%1(%5)
+       sub     a,%12
+       jr      %13,%14
+       ld      a,%1(%5)
+       sub     a,%15
+       jr      %16,%17
+       ld      a,%1(%5)
+       sub     a,%18
+       jr      %19,%20
+} by {
+       ld      a,%1(%5)
+       cp      a,%2
+       jr      %3,%4
+       cp      a,%6
+       jr      %7,%8
+       cp      a,%9
+       jr      %10,%11
+       cp      a,%12
+       jr      %13,%14
+       cp      a,%15
+       jr      %16,%17
+       cp      a,%18
+       jr      %19,%20
+       ; AC peephole 82ixiy6 compacted switch
+} if notUsed('a'), notUsedFrom(%4 'a'), notUsedFrom(%8 'a'), notUsedFrom(%11 'a'), notUsedFrom(%14 'a'), notUsedFrom(%17 'a'), notUsedFrom(%20 'a')
+
+replace restart {
+       ld      a,%1(%5)
+       sub     a,%2
+       jr      %3,%4
+       ld      a,%1(%5)
+       sub     a,%6
+       jr      %7,%8
+       ld      a,%1(%5)
+       sub     a,%9
+       jr      %10,%11
+       ld      a,%1(%5)
+       sub     a,%12
+       jr      %13,%14
+       ld      a,%1(%5)
+       sub     a,%15
+       jr      %16,%17
+} by {
+       ld      a,%1(%5)
+       cp      a,%2
+       jr      %3,%4
+       cp      a,%6
+       jr      %7,%8
+       cp      a,%9
+       jr      %10,%11
+       cp      a,%12
+       jr      %13,%14
+       cp      a,%15
+       jr      %16,%17
+       ; AC peephole 82ixiy5 compacted switch
+} if notUsed('a'), notUsedFrom(%4 'a'), notUsedFrom(%8 'a'), notUsedFrom(%11 'a'), notUsedFrom(%14 'a'), notUsedFrom(%17 'a')
+
+replace restart {
+       ld      a,%1(%5)
+       sub     a,%2
+       jr      %3,%4
+       ld      a,%1(%5)
+       sub     a,%6
+       jr      %7,%8
+       ld      a,%1(%5)
+       sub     a,%9
+       jr      %10,%11
+       ld      a,%1(%5)
+       sub     a,%12
+       jr      %13,%14
+} by {
+       ld      a,%1(%5)
+       cp      a,%2
+       jr      %3,%4
+       cp      a,%6
+       jr      %7,%8
+       cp      a,%9
+       jr      %10,%11
+       cp      a,%12
+       jr      %13,%14
+       ; AC peephole 82ixiy4 compacted switch
+} if notUsed('a'), notUsedFrom(%4 'a'), notUsedFrom(%8 'a'), notUsedFrom(%11 'a'), notUsedFrom(%14 'a')
+
+replace restart {
+       ld      a,%1(%5)
+       sub     a,%2
+       jr      %3,%4
+       ld      a,%1(%5)
+       sub     a,%6
+       jr      %7,%8
+       ld      a,%1(%5)
+       sub     a,%9
+       jr      %10,%11
+} by {
+       ld      a,%1(%5)
+       cp      a,%2
+       jr      %3,%4
+       cp      a,%6
+       jr      %7,%8
+       cp      a,%9
+       jr      %10,%11
+       ; AC peephole 82ixiy3 compacted switch
+} if notUsed('a'), notUsedFrom(%4 'a'), notUsedFrom(%8 'a'), notUsedFrom(%11 'a')
+
+replace restart {
+       ld      a,%1(%5)
+       sub     a,%2
+       jr      %3,%4
+       ld      a,%1(%5)
+       sub     a,%6
+       jr      %7,%8
+} by {
+       ld      a,%1(%5)
+       cp      a,%2
+       jr      %3,%4
+       cp      a,%6
+       jr      %7,%8
+       ; AC peephole 82ixiy2 compacted switch
+} if notUsed('a'), notUsedFrom(%4 'a'), notUsedFrom(%8 'a')
+
+replace restart {
+       ld      a,%1(%5)
+       sub     a,%2
+       jr      %3,%4
+       ld      a,%1(%5)
+} by {
+       ld      a,%1(%5)
+       cp      a,%2
+       jr      %3,%4
+       ; AC peephole 82ixiy compacted switch
+} if notUsed('a'), notUsedFrom(%4 'a')
+
+replace restart {
+       ld      a,%1(%2)
+       or      a, a
+       jr      %3,%4
+       ld      a,%1(%2)
+} by {
+       ld      a,%1(%2)
+       or      a, a
+       jr      %3,%4
+       ; AC remove or/reload
+} if notUsed('a'), notUsedFrom(%4 'a')
+
+replace restart {
+       ld      a,%1
+       or      a,a
+       jr      %2,%3
+       ld      a,%1
+       dec     a
+       jr      %4,%5
+       ld      a,%1
+       sub     a,#%6
+       jr      %7,%8
+} by {
+       ld      a,%1
+       or      a,a
+       jr      %2,%3
+       dec     a
+       jr      %4,%5
+       cp      a,#(%6-1)
+       jr      %7,%8
+       ; AC Fold together 3 way switch compares from zero
+} if notUsed('a'), notUsedFrom(%3 'a'), notUsedFrom(%5 'a'), notUsedFrom(%8 'a')
+// Switch things but for no n (ix/iy) cases
+// Both sets needed expanding to cover each others type
+
+replace restart {
+       ld      a,%1
+       or      a,a
+       jr      %2,%3
+       ld      a,%1
+       dec     a
+       jr      %4,%5
+} by {
+       ld      a,%1
+       or      a,a
+       jr      %2,%3
+       dec     a
+       jr      %4,%5
+       ; AC Fold together 2 way switch compares from zero
+} if notUsed('a'), notUsedFrom(%3 'a'), notUsedFrom(%5 'a')
+
+replace restart {
+       ld      a,%1
+       dec     a
+       jr      %2,%3
+       ld      a,%1
+       sub     a,#%4
+       jr      Z,%6
+       ld      a,%1
+       sub     a, #%7
+       jr      Z,%9
+} by {
+       ld      a,%1
+       dec     a
+       jr      %2,%3
+       cp      a,#%(4-1)
+       jr      Z,%6
+       cp      a,#(%7-1)
+       jr      Z,%9
+       ; AC Fold together 3 way switch compares from one base
+} if notUsed('a'), notUsedFrom(%3 'a'), notUsedFrom(%6 'a'), notUsedFrom(%9 'a')
+
+replace restart {
+       ld      a,%1
+       dec     a
+       jr      %2,%3
+       ld      a,%1
+       sub     a,#%4
+       jr      Z,%6
+} by {
+       ld      a,%1
+       dec     a
+       jr      %2,%3
+       cp      a,#(%4-1)
+       jr      Z,%6
+       ; AC Fold together 2 way switch compares from one base
+} if notUsed('a'), notUsedFrom(%3 'a'), notUsedFrom(%6 'a')
+
+
+replace restart {
+       ld      a,%1
+       sub     a,%2
+       jr      %3,%4
+       ld      a,%1
+       sub     a,%6
+       jr      %7,%8
+       ld      a,%1
+       sub     a,%9
+       jr      %10,%11
+} by {
+       ld      a,%1
+       cp      a,%2
+       jr      %3,%4
+       cp      a,%6
+       jr      %7,%8
+       cp      a,%9
+       jr      %10,%11
+       ; AC Fold together 3 way switch compares
+} if notUsed('a'), notUsedFrom(%4 'a'), notUsedFrom(%8 'a'), notUsedFrom(%11 'a')
+
+replace restart {
+       ld      a,%1
+       sub     a,%2
+       jr      %3,%4
+       ld      a,%1
+       sub     a,%6
+       jr      %7,%8
+} by {
+       ld      a,%1
+       cp      a,%2
+       jr      %3,%4
+       cp      a,%6
+       jr      %7,%8
+       ; AC Fold together 2 way switch compares
+} if notUsed('a'), notUsedFrom(%4 'a'), notUsedFrom(%8 'a')
+
+
+replace restart {
+       ld      hl, #%1+0
+       inc     (hl)
+       jr      NZ,%2
+       ld      hl, #%1+1
+       inc     (hl)
+%2:
+} by {
+       ld      hl, #%1+0
+       inc     (hl)
+       jr      NZ,%2
+       inc     hl
+       inc     (hl)
+       ; AC: collapse 16bit increment of static/global
+%2:
+} if notUsed('hl')
+
+replace restart {
+       ld      hl, #%1+0
+       inc     (hl)
+       jr      NZ,%2
+       ld      hl, #%1+1
+       inc     (hl)
+       jr      NZ,%2
+       ld      hl, #%1+2
+       inc     (hl)
+       jr      NZ,%2
+       ld      hl, #%1+3
+       inc     (hl)
+%2:
+} by {
+       ld      hl, #%1+0
+       inc     (hl)
+       jr      NZ,%2
+       inc     hl
+       inc     (hl)
+       jr      NZ,%2
+       inc     hl
+       inc     (hl)
+       jr      NZ,%2
+       inc     hl
+       inc     (hl)
+       ; AC: collapse 32bit increment of static/global
+%2:
+} if notUsed('hl')
+
+replace restart {
+       ld      iy,#%1
+       ld      a,%2(iy)
+} by {
+       ld      a,(%1 + %2)
+       ; AC    remove iy indirection
+} if notUsed('iy')
+
+replace restart {
+       push    hl
+       ld      %1,(hl)
+       inc     hl
+       ld      %2,(hl)
+       inc     hl
+       ld      %3,(hl)
+       inc     hl
+       ld      %4,(hl)
+       pop     hl
+       ld      a,%1
+       and     a,%5(%6)
+       ld      %1,a
+       ld      a,%2
+       and     a,%7(%6)
+       ld      %2,a
+       ld      a,%3
+       and     a,%8(%6)
+       ld      %3,a
+       ld      a,%4
+       and     a,%9(%6)
+       ld      %4,a
+       ld      (hl),%1
+       inc     hl
+       ld      (hl),%2
+       inc     hl
+       ld      (hl),%3
+       inc     hl
+       ld      (hl),%4
+} by {
+       ld      a,%5(%6)
+       and     a, (hl)
+       ld      (hl),a
+       inc hl
+       ld      a,%7(%6)
+       and     a, (hl)
+       ld      (hl),a
+       inc hl
+       ld      a,%8(%6)
+       and     a, (hl)
+       ld      (hl),a
+       inc hl
+       ld      a,%9(%6)
+       and     a, (hl)
+       ld      (hl),a
+       ;       AC fix &= L
+} if notUsed(%1), notUsed(%2), notUsed(%3), notUsed(%4)
+// Clean up the 32bit &= operator
+
+replace restart {
+       push    hl
+       ld      %1,(hl)
+       inc     hl
+       ld      %2,(hl)
+       inc     hl
+       ld      %3,(hl)
+       inc     hl
+       ld      %4,(hl)
+       pop     hl
+       ld      a,%1
+       or      a,%5(%6)
+       ld      %1,a
+       ld      a,%2
+       or      a,%7(%6)
+       ld      %2,a
+       ld      a,%3
+       or      a,%8(%6)
+       ld      %3,a
+       ld      a,%4
+       or      a,%9(%6)
+       ld      %4,a
+       ld      (hl),%1
+       inc     hl
+       ld      (hl),%2
+       inc     hl
+       ld      (hl),%3
+       inc     hl
+       ld      (hl),%4
+} by {
+       ld      a,%5(%6)
+       or      a, (hl)
+       ld      (hl),a
+       inc hl
+       ld      a,%7(%6)
+       or      a, (hl)
+       ld      (hl),a
+       inc hl
+       ld      a,%8(%6)
+       or      a, (hl)
+       ld      (hl),a
+       inc hl
+       ld      a,%9(%6)
+       or      a, (hl)
+       ld      (hl),a
+       ;       AC fix |= L
+} if notUsed(%1), notUsed(%2), notUsed(%3), notUsed(%4)
+// Clean up the 32bit |= operator