internal/bytealg: add regABI support in bytealg functions on loong64

Update #40724

Co-authored-by: Xiaolin Zhao <[email protected]>
Change-Id: I4a7392afd7238d44e7d09aaca7e0d733649926ac
Reviewed-on: https://go-review.googlesource.com/c/go/+/521785
TryBot-Result: Gopher Robot <[email protected]>
Reviewed-by: Cherry Mui <[email protected]>
Run-TryBot: David Chase <[email protected]>
Reviewed-by: Meidan Li <[email protected]>
Reviewed-by: David Chase <[email protected]>
Auto-Submit: David Chase <[email protected]>

Author: 2 people

src/internal/bytealg/compare_loong64.s

@@ -5,83 +5,102 @@
 #include "go_asm.h"
 #include "textflag.h"
 
-TEXT ·Compare(SB),NOSPLIT,$0-56
-	MOVV	a_base+0(FP), R6
-	MOVV	b_base+24(FP), R7
-	MOVV	a_len+8(FP), R4
-	MOVV	b_len+32(FP), R5
+TEXT ·Compare<ABIInternal>(SB),NOSPLIT,$0-56
+#ifndef GOEXPERIMENT_regabiargs
+	MOVV	a_base+0(FP), R4
+	MOVV	a_len+8(FP), R5
+	MOVV	b_base+24(FP), R6
+	MOVV	b_len+32(FP), R7
 	MOVV	$ret+48(FP), R13
+#else
+	// R4 = a_base
+	// R5 = a_len
+	// R6 = a_cap (unused)
+	// R7 = b_base (want in R6)
+	// R8 = b_len (want in R7)
+	// R9 = b_cap (unused)
+	MOVV	R7, R6
+	MOVV	R8, R7
+#endif
 	JMP	cmpbody<>(SB)
 
-TEXT runtime·cmpstring(SB),NOSPLIT,$0-40
-	MOVV	a_base+0(FP), R6
-	MOVV	b_base+16(FP), R7
-	MOVV	a_len+8(FP), R4
-	MOVV	b_len+24(FP), R5
+TEXT runtime·cmpstring<ABIInternal>(SB),NOSPLIT,$0-40
+#ifndef GOEXPERIMENT_regabiargs
+	MOVV	a_base+0(FP), R4
+	MOVV	b_base+16(FP), R6
+	MOVV	a_len+8(FP), R5
+	MOVV	b_len+24(FP), R7
 	MOVV	$ret+32(FP), R13
+#endif
+	// R4 = a_base
+	// R5 = a_len
+	// R6 = b_base
+	// R7 = b_len
 	JMP	cmpbody<>(SB)
 
 // On entry:
-// R4 length of a
-// R5 length of b
-// R6 points to the start of a
-// R7 points to the start of b
+// R5 length of a
+// R7 length of b
+// R4 points to the start of a
+// R6 points to the start of b
 // R13 points to the return value (-1/0/1)
 TEXT cmpbody<>(SB),NOSPLIT|NOFRAME,$0
-	BEQ	R6, R7, samebytes // same start of a and b
+	BEQ	R4, R6, samebytes // same start of a and b
 
-	SGTU	R4, R5, R9
+	SGTU	R5, R7, R9
 	BNE	R0, R9, r2_lt_r1
-	MOVV	R4, R14
+	MOVV	R5, R14
 	JMP	entry
 r2_lt_r1:
-	MOVV	R5, R14	// R14 is min(R4, R5)
+	MOVV	R7, R14	// R14 is min(R4, R5)
 entry:
-	ADDV	R6, R14, R12	// R6 start of a, R14 end of a
-	BEQ	R6, R12, samebytes // length is 0
+	ADDV	R4, R14, R12	// R6 start of a, R14 end of a
+	BEQ	R4, R12, samebytes // length is 0
 
 	SRLV	$4, R14		// R14 is number of chunks
 	BEQ	R0, R14, byte_loop
 
 	// make sure both a and b are aligned.
-	OR	R6, R7, R15
+	OR	R4, R6, R15
 	AND	$7, R15
 	BNE	R0, R15, byte_loop
 
 	PCALIGN	$16
 chunk16_loop:
 	BEQ	R0, R14, byte_loop
-	MOVV	(R6), R8
-	MOVV	(R7), R9
+	MOVV	(R4), R8
+	MOVV	(R6), R9
 	BNE	R8, R9, byte_loop
-	MOVV	8(R6), R16
-	MOVV	8(R7), R17
+	MOVV	8(R4), R16
+	MOVV	8(R6), R17
+	ADDV	$16, R4
 	ADDV	$16, R6
-	ADDV	$16, R7
 	SUBVU	$1, R14
 	BEQ	R16, R17, chunk16_loop
+	SUBV	$8, R4
 	SUBV	$8, R6
-	SUBV	$8, R7
 
 byte_loop:
-	BEQ	R6, R12, samebytes
-	MOVBU	(R6), R8
+	BEQ	R4, R12, samebytes
+	MOVBU	(R4), R8
+	ADDVU	$1, R4
+	MOVBU	(R6), R9
 	ADDVU	$1, R6
-	MOVBU	(R7), R9
-	ADDVU	$1, R7
 	BEQ	R8, R9, byte_loop
 
 byte_cmp:
-	SGTU	R8, R9, R12 // R12 = 1 if (R8 > R9)
-	BNE	R0, R12, ret
-	MOVV	$-1, R12
+	SGTU	R8, R9, R4 // R12 = 1 if (R8 > R9)
+	BNE	R0, R4, ret
+	MOVV	$-1, R4
 	JMP	ret
 
 samebytes:
-	SGTU	R4, R5, R8
-	SGTU	R5, R4, R9
-	SUBV	R9, R8, R12
+	SGTU	R5, R7, R8
+	SGTU	R7, R5, R9
+	SUBV	R9, R8, R4
 
 ret:
-	MOVV	R12, (R13)
+#ifndef GOEXPERIMENT_regabiargs
+	MOVV	R4, (R13)
+#endif
 	RET

src/internal/bytealg/equal_loong64.s

@@ -8,17 +8,21 @@
 #define	REGCTXT	R29
 
 // memequal(a, b unsafe.Pointer, size uintptr) bool
-TEXT runtime·memequal(SB),NOSPLIT|NOFRAME,$0-25
+TEXT runtime·memequal<ABIInternal>(SB),NOSPLIT|NOFRAME,$0-25
+#ifndef GOEXPERIMENT_regabiargs
 	MOVV	a+0(FP), R4
 	MOVV	b+8(FP), R5
-	BEQ	R4, R5, eq
 	MOVV	size+16(FP), R6
+#endif
+	BEQ	R4, R5, eq
 	ADDV	R4, R6, R7
 	PCALIGN	$16
 loop:
 	BNE	R4, R7, test
 	MOVV	$1, R4
+#ifndef GOEXPERIMENT_regabiargs
 	MOVB	R4, ret+24(FP)
+#endif
 	RET
 test:
 	MOVBU	(R4), R9
@@ -27,27 +31,38 @@ test:
 	ADDV	$1, R5
 	BEQ	R9, R10, loop
 
+	MOVB    R0, R4
+#ifndef GOEXPERIMENT_regabiargs
 	MOVB	R0, ret+24(FP)
+#endif
 	RET
 eq:
 	MOVV	$1, R4
+#ifndef GOEXPERIMENT_regabiargs
 	MOVB	R4, ret+24(FP)
+#endif
 	RET
 
 // memequal_varlen(a, b unsafe.Pointer) bool
-TEXT runtime·memequal_varlen(SB),NOSPLIT,$40-17
+TEXT runtime·memequal_varlen<ABIInternal>(SB),NOSPLIT,$40-17
+#ifndef GOEXPERIMENT_regabiargs
 	MOVV	a+0(FP), R4
 	MOVV	b+8(FP), R5
+#endif
 	BEQ	R4, R5, eq
 	MOVV	8(REGCTXT), R6    // compiler stores size at offset 8 in the closure
 	MOVV	R4, 8(R3)
 	MOVV	R5, 16(R3)
 	MOVV	R6, 24(R3)
 	JAL	runtime·memequal(SB)
 	MOVBU	32(R3), R4
+#ifndef GOEXPERIMENT_regabiargs
 	MOVB	R4, ret+16(FP)
+#endif
 	RET
 eq:
 	MOVV	$1, R4
+#ifndef GOEXPERIMENT_regabiargs
 	MOVB	R4, ret+16(FP)
+#endif
 	RET

src/internal/bytealg/indexbyte_loong64.s

@@ -5,11 +5,18 @@
 #include "go_asm.h"
 #include "textflag.h"
 
-TEXT ·IndexByte(SB),NOSPLIT,$0-40
+TEXT ·IndexByte<ABIInternal>(SB),NOSPLIT,$0-40
+#ifndef GOEXPERIMENT_regabiargs
 	MOVV	b_base+0(FP), R4
 	MOVV	b_len+8(FP), R5
-	MOVBU	c+24(FP), R6	// byte to find
-	MOVV	R4, R7		// store base for later
+	MOVBU	c+24(FP), R7	// byte to find
+#endif
+	// R4 = b_base
+	// R5 = b_len
+	// R6 = b_cap (unused)
+	// R7 = byte to find
+	AND	$0xff, R7
+	MOVV	R4, R6		// store base for later
 	ADDV	R4, R5		// end
 	ADDV	$-1, R4
 
@@ -18,21 +25,30 @@ loop:
 	ADDV	$1, R4
 	BEQ	R4, R5, notfound
 	MOVBU	(R4), R8
-	BNE	R6, R8, loop
+	BNE	R7, R8, loop
 
-	SUBV	R7, R4		// remove base
+	SUBV	R6, R4		// remove base
+#ifndef GOEXPERIMENT_regabiargs
 	MOVV	R4, ret+32(FP)
+#endif
 	RET
 
 notfound:
 	MOVV	$-1, R4
+#ifndef GOEXPERIMENT_regabiargs
 	MOVV	R4, ret+32(FP)
+#endif
 	RET
 
-TEXT ·IndexByteString(SB),NOSPLIT,$0-32
+TEXT ·IndexByteString<ABIInternal>(SB),NOSPLIT,$0-32
+#ifndef GOEXPERIMENT_regabiargs
 	MOVV	s_base+0(FP), R4
 	MOVV	s_len+8(FP), R5
 	MOVBU	c+16(FP), R6	// byte to find
+#endif
+	// R4 = s_base
+	// R5 = s_len
+	// R6 = byte to find
 	MOVV	R4, R7		// store base for later
 	ADDV	R4, R5		// end
 	ADDV	$-1, R4
@@ -45,10 +61,14 @@ loop:
 	BNE	R6, R8, loop
 
 	SUBV	R7, R4		// remove base
+#ifndef GOEXPERIMENT_regabiargs
 	MOVV	R4, ret+24(FP)
+#endif
 	RET
 
 notfound:
 	MOVV	$-1, R4
+#ifndef GOEXPERIMENT_regabiargs
 	MOVV	R4, ret+24(FP)
+#endif
 	RET