Make use of lswx/stswx where applicable

extremscorner · Jun 25, 2024 · bf67b31 · bf67b31
1 parent cb23a1c
commit bf67b31
Show file tree

Hide file tree

Showing 3 changed files with 29 additions and 19 deletions.
diff --git a/gc/ogc/machine/processor.h b/gc/ogc/machine/processor.h
@@ -90,12 +90,20 @@
 	__asm__ __volatile__ ("lwbrx	%0,%1,%2" : "=r"(res) : "b%"(index), "r"(base) : "memory"); \
 	res; })
 
+#define __lswx(base,bytes)			\
+({	register u32 res;				\
+	__asm__ __volatile__ ("mtxer %2; lswx %0,%y1" : "=&r"(res) : "Z"(*(u32*)(base)), "r"(bytes) : "xer"); \
+	res; })
+
 #define __sthbrx(base,index,value)	\
 	__asm__ __volatile__ ("sthbrx	%0,%1,%2" : : "r"(value), "b%"(index), "r"(base) : "memory")
 
 #define __stwbrx(base,index,value)	\
 	__asm__ __volatile__ ("stwbrx	%0,%1,%2" : : "r"(value), "b%"(index), "r"(base) : "memory")
 
+#define __stswx(base,bytes,value)	\
+	__asm__ __volatile__ ("mtxer %2; stswx %1,%y0" : "=Z"(*(u32*)(base)) : "r"(value), "r"(bytes) : "xer");
+
 #define cntlzw(_val) ({register u32 _rval; \
 					  __asm__ __volatile__("cntlzw %0, %1" : "=r"((_rval)) : "r"((_val))); _rval;})
 

diff --git a/libogc/exi.c b/libogc/exi.c
@@ -435,7 +435,7 @@ s32 EXI_Sync(s32 nChn)
 {
 	u8 *buf;
 	s32 ret;
-	u32 level,i,cnt,val;
+	u32 level,len,val;
 	exibus_priv *exi = &eximap[nChn];
 #ifdef _EXI_DEBUG
 	printf("EXI_Sync(%d)\n",nChn);
@@ -447,11 +447,11 @@ s32 EXI_Sync(s32 nChn)
 	ret = 0;
 	if(exi->flags&EXI_FLAG_SELECT && exi->flags&(EXI_FLAG_DMA|EXI_FLAG_IMM)) {
 		if(exi->flags&EXI_FLAG_IMM) {
-			cnt = exi->imm_len;
 			buf = exi->imm_buff;
-			if(buf && cnt>0) {
+			len = exi->imm_len;
+			if(len>0) {
 				val = _exiReg[nChn*5+4];
-				for(i=0;i<cnt;i++) ((u8*)buf)[i] = (val>>((3-i)*8))&0xFF;
+				__stswx(buf,len,val);
 			}
 		}
 		exi->flags &= ~(EXI_FLAG_DMA|EXI_FLAG_IMM);
@@ -463,8 +463,8 @@ s32 EXI_Sync(s32 nChn)
 
 s32 EXI_Imm(s32 nChn,void *pData,u32 nLen,u32 nMode,EXICallback tc_cb)
 {
+	u32 val;
 	u32 level;
-	u32 value,i;
 	exibus_priv *exi = &eximap[nChn];
 #ifdef _EXI_DEBUG
 	printf("EXI_Imm(%d,%p,%d,%d,%p)\n",nChn,pData,nLen,nMode,tc_cb);
@@ -486,8 +486,8 @@ s32 EXI_Imm(s32 nChn,void *pData,u32 nLen,u32 nMode,EXICallback tc_cb)
 	exi->imm_buff = pData;
 	exi->imm_len = nLen;
 	if(nMode!=EXI_READ) {
-		for(i=0,value=0;i<nLen;i++) value |= (((u8*)pData)[i])<<((3-i)*8);
-		_exiReg[nChn*5+4] = value;
+		val = __lswx(pData,nLen);
+		_exiReg[nChn*5+4] = val;
 	}
 	if(nMode==EXI_WRITE) exi->imm_len = 0;
 
@@ -934,7 +934,7 @@ void __exi_irq_handler(u32 nIrq,frame_context *pCtx)
 
 void __tc_irq_handler(u32 nIrq,frame_context *pCtx)
 {
-	u32 cnt,len,d,chan,dev;
+	u32 len,val,chan,dev;
 	EXICallback tccb;
 	void *buf = NULL;
 	exibus_priv *exi = NULL;
@@ -956,13 +956,11 @@ void __tc_irq_handler(u32 nIrq,frame_context *pCtx)
 	exi->CallbackTC = NULL;
 	if(exi->flags&(EXI_FLAG_DMA|EXI_FLAG_IMM)) {
 		if(exi->flags&EXI_FLAG_IMM) {
-			len = exi->imm_len;
 			buf = exi->imm_buff;
-			if(len>0 && buf) {
-				d = _exiReg[chan*5+4];
-				if(d>0) {
-					for(cnt=0;cnt<len;cnt++) ((u8*)buf)[cnt] = (d>>((3-cnt)*8))&0xFF;
-				}
+			len = exi->imm_len;
+			if(len>0) {
+				val = _exiReg[chan*5+4];
+				__stswx(buf,len,val);
 			}
 		}
 		exi->flags &= ~(EXI_FLAG_DMA|EXI_FLAG_IMM);

diff --git a/libogc/si.c b/libogc/si.c
@@ -165,7 +165,7 @@ static void __si_alarmhandler(syswd_t thealarm,void *cbarg)
 
 static u32 __si_completetransfer(void)
 {
-	u32 val,sisr,cnt,i;
+	u32 sisr,val,cnt,i;
 	u32 *in;
 
 #ifdef _SI_DEBUG
@@ -183,7 +183,7 @@ static u32 __si_completetransfer(void)
 	for(i=0;i<cnt;i++) in[i] = _siReg[32+i];
 	if(sicntrl.in_bytes&0x03) {
 		val = _siReg[32+cnt];
-		for(i=0;i<(sicntrl.in_bytes&0x03);i++) ((u8*)in)[(cnt*4)+i] = (val>>((3-i)*8))&0xff;
+		__stswx(in+cnt,(sicntrl.in_bytes&0x03),val);
 	}
 #ifdef _SI_DEBUG
 	printf("__si_completetransfer(csr = %08x)\n",_siReg[13]);
@@ -203,7 +203,7 @@ static u32 __si_completetransfer(void)
 
 static u32 __si_transfer(s32 chan,void *out,u32 out_len,void *in,u32 in_len,SICallback cb)
 {
-	u32 level,cnt,i;
+	u32 level,val,cnt,i;
 	sicomcsr csr;
 #ifdef _SI_DEBUG
 	printf("__si_transfer(%d,%p,%d,%p,%d,%p)\n",chan,out,out_len,in,in_len,cb);
@@ -225,8 +225,12 @@ static u32 __si_transfer(s32 chan,void *out,u32 out_len,void *in,u32 in_len,SICa
 #ifdef _SI_DEBUG
 	printf("__si_transfer(csr = %08x,sr = %08x)\n",_siReg[13],_siReg[14]);
 #endif
-	cnt = ((out_len+3)/4);
-	for(i=0;i<cnt;i++)  _siReg[32+i] = ((u32*)out)[i];
+	cnt = (out_len/4);
+	for(i=0;i<cnt;i++) _siReg[32+i] = ((u32*)out)[i];
+	if(out_len&0x03) {
+		val = __lswx((u32*)out+cnt,(out_len&0x03));
+		_siReg[32+cnt] = val;
+	}
 
 	csr.val = _siReg[13];
 	csr.csrmap.tcint = 1;