From 056b49e1d1bc12fa9a4958be7d373e8fe6f23cd2 Mon Sep 17 00:00:00 2001
From: drmortalwombat <90205530+drmortalwombat@users.noreply.github.com>
Date: Fri, 4 Oct 2024 20:06:56 +0200
Subject: [PATCH] Optimize 8 by 16 bit multiply

---
 include/crt.c                   | 50 +++++++++++++++++++++++++++++++++
 oscar64/NativeCodeGenerator.cpp |  8 ++++--
 2 files changed, 56 insertions(+), 2 deletions(-)

diff --git a/include/crt.c b/include/crt.c
index 8750d5c..08a6535 100644
--- a/include/crt.c
+++ b/include/crt.c
@@ -839,6 +839,55 @@ W1:		asl	accu
 #endif
 }
 
+#if 1
+__asm mul16by8
+{
+		lsr
+		beq	zero
+	more:
+		ldx #0
+		ldy #0
+		bcc	skip
+	odd:
+		ldy accu
+		ldx accu + 1
+		bcs skip
+		
+	loop:
+		sta tmpy
+		
+		clc
+		tya
+		adc	accu
+		tay
+		txa
+		adc	accu + 1		
+		tax
+		
+		lda tmpy
+	skip:	
+		asl accu
+		rol accu + 1
+		lsr
+		bcc skip
+		bne loop
+	done:
+		clc
+		tya
+		adc accu
+		sta accu
+		txa
+		adc accu + 1
+		sta accu + 1
+		rts
+	zero:
+		bcs one
+		sta accu
+		sta accu + 1
+	one:
+		rts		
+}
+#else
 __asm mul16by8
 {
 		ldy	#0
@@ -866,6 +915,7 @@ L2:
 		sty tmp + 2
 		rts
 }
+#endif
 
 __asm divs16
 {
diff --git a/oscar64/NativeCodeGenerator.cpp b/oscar64/NativeCodeGenerator.cpp
index 512863d..e966b63 100644
--- a/oscar64/NativeCodeGenerator.cpp
+++ b/oscar64/NativeCodeGenerator.cpp
@@ -9390,6 +9390,8 @@ int NativeCodeBasicBlock::ShortMultiply(InterCodeProcedure* proc, NativeCodeProc
 
 				NativeCodeGenerator::Runtime& rt(nproc->mGenerator->ResolveRuntime(Ident::Unique("mul16by8")));
 				mIns.Push(NativeCodeInstruction(ins, ASMIT_JSR, ASMIM_ABSOLUTE, rt.mOffset, rt.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER | NCIF_USE_CPU_REG_A));
+
+				return BC_REG_ACCU;
 			}
 			else
 			{
@@ -9425,6 +9427,8 @@ int NativeCodeBasicBlock::ShortMultiply(InterCodeProcedure* proc, NativeCodeProc
 
 			NativeCodeGenerator::Runtime& rt(nproc->mGenerator->ResolveRuntime(Ident::Unique("mul16by8")));
 			mIns.Push(NativeCodeInstruction(ins, ASMIT_JSR, ASMIM_ABSOLUTE, rt.mOffset, rt.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER | NCIF_USE_CPU_REG_A));
+
+			return BC_REG_ACCU;
 		}
 
 		return BC_REG_WORK + 2;
@@ -10995,7 +10999,7 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 
 				NativeCodeGenerator::Runtime& frt(nproc->mGenerator->ResolveRuntime(Ident::Unique("mul16by8")));
 				mIns.Push(NativeCodeInstruction(ins, ASMIT_JSR, ASMIM_ABSOLUTE, frt.mOffset, frt.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER | NCIF_USE_CPU_REG_A));
-				reg = BC_REG_WORK + 2;
+				reg = BC_REG_ACCU;
 			}
 			else if (ins->mOperator == IA_MUL && ins->mSrc[1].IsUByte())
 			{
@@ -11019,7 +11023,7 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 
 				NativeCodeGenerator::Runtime& frt(nproc->mGenerator->ResolveRuntime(Ident::Unique("mul16by8")));
 				mIns.Push(NativeCodeInstruction(ins, ASMIT_JSR, ASMIM_ABSOLUTE, frt.mOffset, frt.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER | NCIF_USE_CPU_REG_A));
-				reg = BC_REG_WORK + 2;
+				reg = BC_REG_ACCU;
 			}
 #if 1
 			else if (ins->mOperator == IA_DIVS && ins->mSrc[0].mTemp < 0 && (ins->mSrc[0].mIntConst == 2 || ins->mSrc[0].mIntConst == 4 || ins->mSrc[0].mIntConst == 8))