[v2,10/13] libc: mips: Improve performance of strcmp implementation

Message ID 20250402141228.1973965-11-jovan.dmitrovic@htecgroup.com
State New
Headers
Series A series of updates related to MIPS |

Commit Message

Jovan Dmitrovic April 2, 2025, 2:14 p.m. UTC
  From: Faraz Shahbazker <fshahbazker@wavecomp.com>

Improve `strcmp` by using `ext` instruction, if available.
---
 newlib/libc/machine/mips/strcmp.S | 202 ++++++++++++++++++------------
 1 file changed, 123 insertions(+), 79 deletions(-)
  

Patch

diff --git a/newlib/libc/machine/mips/strcmp.S b/newlib/libc/machine/mips/strcmp.S
index a59f73e52..126344d6c 100644
--- a/newlib/libc/machine/mips/strcmp.S
+++ b/newlib/libc/machine/mips/strcmp.S
@@ -41,6 +41,10 @@ 
    performance loss, so we are not turning it on by default.  */
 #if defined(ENABLE_CLZ) && (__mips_isa_rev > 1)
 # define USE_CLZ
+#elif (__mips_isa_rev >= 2)
+# define USE_EXT 1
+#else
+# define USE_EXT 0
 #endif
 
 /* Some asm.h files do not have the L macro definition.  */
@@ -61,6 +65,10 @@ 
 # endif
 #endif
 
+/* Haven't yet found a configuration where DSP code outperforms
+   normal assembly.  */
+#define __mips_using_dsp 0
+
 /* Allow the routine to be named something else if desired.  */
 #ifndef STRCMP_NAME
 # define STRCMP_NAME strcmp
@@ -72,25 +80,36 @@  LEAF(STRCMP_NAME, 0)
 LEAF(STRCMP_NAME)
 #endif
 	.set	nomips16
-
 	or	t0, a0, a1
-	andi	t0,0x3
+	andi	t0, t0, 0x3
 	bne	t0, zero, L(byteloop)
 
 /* Both strings are 4 byte aligned at this point.  */
 
 	li	t8, 0x01010101
+#if !__mips_using_dsp
 	li	t9, 0x7f7f7f7f
+#endif
 
-#define STRCMP32(OFFSET) \
-	lw	vt0, OFFSET(a0); \
-	lw	vt1, OFFSET(a1); \
-	subu	t0, vt0, t8; \
-	nor	t1, vt0, t9; \
-	bne	vt0, vt1, L(worddiff); \
-	and	t0, t0, t1; \
+#if __mips_using_dsp
+# define STRCMP32(OFFSET)	\
+	lw	a2, OFFSET(a0); \
+	lw	a3, OFFSET(a1); \
+	subu_s.qb t0, t8, a2;	\
+	bne	a2, a3, L(worddiff); \
 	bne	t0, zero, L(returnzero)
+#else  /* !__mips_using_dsp */
+# define STRCMP32(OFFSET)	\
+	lw	a2, OFFSET(a0); \
+	lw	a3, OFFSET(a1); \
+	subu	t0, a2, t8;	\
+	nor	t1, a2, t9;	\
+	bne	a2, a3, L(worddiff); \
+	and	t1, t0, t1;	\
+	bne	t1, zero, L(returnzero)
+#endif /* __mips_using_dsp */
 
+	.align 2
 L(wordloop):
 	STRCMP32(0)
 	STRCMP32(4)
@@ -99,118 +118,143 @@  L(wordloop):
 	STRCMP32(16)
 	STRCMP32(20)
 	STRCMP32(24)
-	lw	vt0, 28(a0)
-	lw	vt1, 28(a1)
-	subu	t0, vt0, t8
-	nor	t1, vt0, t9
-	bne	vt0, vt1, L(worddiff)
-	and	t0, t0, t1
+	lw	a2, 28(a0)
+	lw	a3, 28(a1)
+#if __mips_using_dsp
+	subu_s.qb t0, t8, a2
+#else
+	subu	t0, a2, t8
+	nor	t1, a2, t9
+	and	t1, t0, t1
+#endif
+
 	PTR_ADDIU a0, a0, 32
-	bne	t0, zero, L(returnzero)
+	bne	a2, a3, L(worddiff)
 	PTR_ADDIU a1, a1, 32
-	b	L(wordloop)
+	beq	t1, zero, L(wordloop)
 
 L(returnzero):
 	move	va0, zero
 	jr	ra
 
+	.align 2
 L(worddiff):
 #ifdef USE_CLZ
-	subu	t0, vt0, t8
-	nor	t1, vt0, t9
-	and	t1, t0, t1
-	xor	t0, vt0, vt1
+	xor	t0, a2, a3
 	or	t0, t0, t1
 # if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 	wsbh	t0, t0
 	rotr	t0, t0, 16
-# endif
+# endif /* LITTLE_ENDIAN */
 	clz	t1, t0
-	and	t1, 0xf8
-# if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
-	neg	t1
-	addu	t1, 24
+	or	t0, t1, 24	/* Only care about multiples of 8.  */
+	xor	t1, t1, t0	/* {0,8,16,24} => {24,16,8,0}  */
+# if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+	sllv	a2,a2,t1
+	sllv	a3,a3,t1
+# else
+	srlv	a2,a2,t1
+	srlv	a3,a3,t1
 # endif
-	rotrv	vt0, vt0, t1
-	rotrv	vt1, vt1, t1
-	and	vt0, vt0, 0xff
-	and	vt1, vt1, 0xff
-	subu	va0, vt0, vt1
+	subu	va0, a2, a3
 	jr	ra
 #else /* USE_CLZ */
 # if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
-	andi	t0, vt0, 0xff
-	andi	t1, vt1, 0xff
-	beq	t0, zero, L(wexit01)
-	srl	t8, vt0, 8
-	bne	t0, t1, L(wexit01)
-
-	srl	t9, vt1, 8
+	andi	a0, a2, 0xff	/* abcd => d */
+	andi	a1, a3, 0xff
+	beq	a0, zero, L(wexit01)
+#  if USE_EXT
+	ext	t8, a2, 8, 8
+	bne	a0, a1, L(wexit01)
+	ext	t9, a3, 8, 8
+	beq	t8, zero, L(wexit89)
+	ext	a0, a2, 16, 8
+	bne	t8, t9, L(wexit89)
+	ext	a1, a3, 16, 8
+#  else  /* !USE_EXT */
+	srl	t8, a2, 8
+	bne	a0, a1, L(wexit01)
+	srl	t9, a3, 8
 	andi	t8, t8, 0xff
 	andi	t9, t9, 0xff
 	beq	t8, zero, L(wexit89)
-	srl	t0, vt0, 16
+	srl	a0, a2, 16
 	bne	t8, t9, L(wexit89)
+	srl	a1, a3, 16
+	andi	a0, a0, 0xff
+	andi	a1, a1, 0xff
+#  endif  /* !USE_EXT */
 
-	srl	t1, vt1, 16
-	andi	t0, t0, 0xff
-	andi	t1, t1, 0xff
-	beq	t0, zero, L(wexit01)
-	srl	t8, vt0, 24
-	bne	t0, t1, L(wexit01)
-
-	srl	t9, vt1, 24
 # else /* __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ */
-	srl	t0, vt0, 24
-	srl	t1, vt1, 24
-	beq	t0, zero, L(wexit01)
-	srl	t8, vt0, 16
-	bne	t0, t1, L(wexit01)
+	srl	a0, a2, 24	/* abcd => a */
+	srl	a1, a3, 24
+	beq	a0, zero, L(wexit01)
 
-	srl	t9, vt1, 16
+#  if USE_EXT
+	ext	t8, a2, 16, 8
+	bne	a0, a1, L(wexit01)
+	ext	t9, a3, 16, 8
+	beq	t8, zero, L(wexit89)
+	ext	a0, a2, 8, 8
+	bne	t8, t9, L(wexit89)
+	ext	a1, a3, 8, 8
+#  else /* ! USE_EXT */
+	srl	t8, a2, 16
+	bne	a0, a1, L(wexit01)
+	srl	t9, a3, 16
 	andi	t8, t8, 0xff
 	andi	t9, t9, 0xff
 	beq	t8, zero, L(wexit89)
-	srl	t0, vt0, 8
+	srl	a0, a2, 8
 	bne	t8, t9, L(wexit89)
+	srl	a1, a3, 8
+	andi	a0, a0, 0xff
+	andi	a1, a1, 0xff
+#  endif  /* USE_EXT */
 
-	srl	t1, vt1, 8
-	andi	t0, t0, 0xff
-	andi	t1, t1, 0xff
-	beq	t0, zero, L(wexit01)
-	andi	t8, vt0, 0xff
-	bne	t0, t1, L(wexit01)
-
-	andi	t9, vt1, 0xff
 # endif /* __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ */
 
+	beq	a0, zero, L(wexit01)
+	bne	a0, a1, L(wexit01)
+
+	/* The other bytes are identical, so just subract the 2 words
+	  and return the difference.  */
+	move a0, a2
+	move a1, a3
+
+L(wexit01):
+	subu	va0, a0, a1
+	jr	ra
+
 L(wexit89):
 	subu	va0, t8, t9
 	jr	ra
-L(wexit01):
-	subu	va0, t0, t1
-	jr	ra
+
 #endif /* USE_CLZ */
 
+#define DELAY_NOP nop
+
 /* It might seem better to do the 'beq' instruction between the two 'lbu'
    instructions so that the nop is not needed but testing showed that this
    code is actually faster (based on glibc strcmp test).  */
+
 #define BYTECMP01(OFFSET) \
-	lbu	vt1, OFFSET(a1); \
-	nop; \
-	beq	vt0, zero, L(bexit01); \
+	lbu	a3, OFFSET(a1); \
+	DELAY_NOP; \
+	beq	a2, zero, L(bexit01); \
 	lbu	t8, OFFSET+1(a0); \
-	bne	vt0, vt1, L(bexit01)
+	bne	a2, a3, L(bexit01)
 
 #define BYTECMP89(OFFSET) \
 	lbu	t9, OFFSET(a1); \
-	nop;	\
+	DELAY_NOP; \
 	beq	t8, zero, L(bexit89); \
-	lbu	vt0, OFFSET+1(a0); \
+	lbu	a2, OFFSET+1(a0); \
 	bne	t8, t9, L(bexit89)
 
+	.align 2
 L(byteloop):
-	lbu	vt0, 0(a0)
+	lbu	a2, 0(a0)
 	BYTECMP01(0)
 	BYTECMP89(1)
 	BYTECMP01(2)
@@ -219,20 +263,20 @@  L(byteloop):
 	BYTECMP89(5)
 	BYTECMP01(6)
 	lbu	t9, 7(a1)
-	nop
-	beq	t8, zero, L(bexit89)
+
 	PTR_ADDIU a0, a0, 8
-	bne	t8, t9, L(bexit89)
+	beq	t8, zero, L(bexit89)
 	PTR_ADDIU a1, a1, 8
-	b	L(byteloop)
+	beq	t8, t9, L(byteloop)
 
-L(bexit01):
-	subu	va0, vt0, vt1
-	jr	ra
 L(bexit89):
 	subu	va0, t8, t9
 	jr	ra
 
+L(bexit01):
+	subu	va0, a2, a3
+	jr	ra
+
 	.set	at
 
 END(STRCMP_NAME)