diff mbox series

[2/2] MIPS: Avoid handcoded DIVU in `__div64_32' altogether

Message ID alpine.DEB.2.21.2104222203130.44318@angie.orcam.me.uk (mailing list archive)
State Accepted
Commit 25ab14cbe9d1b66fda44c71a2db7582a31b6f5cd
Headers show
Series MIPSr6 `do_div' support and typo fixes | expand

Commit Message

Maciej W. Rozycki April 22, 2021, 8:36 p.m. UTC
Remove the inline asm with a DIVU instruction from `__div64_32' and use 
plain C code for the intended DIVMOD calculation instead.  GCC is smart 
enough to know that both the quotient and the remainder are calculated 
with single DIVU, so with ISAs up to R5 the same instruction is actually 
produced with overall similar code.

For R6 compiled code will work, but separate DIVU and MODU instructions 
will be produced, which are also interlocked, so scalar implementations 
will likely not perform as well as older ISAs with their asynchronous MD 
unit.  Likely still faster then the generic algorithm though.

This removes a compilation error for R6 however where the original DIVU 
instruction is not supported anymore and the MDU accumulator registers 
have been removed and consequently GCC complains as to a constraint it 
cannot find a register for:

In file included from ./include/linux/math.h:5,
                 from ./include/linux/kernel.h:13,
                 from mm/page-writeback.c:15:
./include/linux/math64.h: In function 'div_u64_rem':
./arch/mips/include/asm/div64.h:76:17: error: inconsistent operand constraints in an 'asm'
   76 |                 __asm__("divu   $0, %z1, %z2"                           \
      |                 ^~~~~~~
./include/asm-generic/div64.h:245:25: note: in expansion of macro '__div64_32'
  245 |                 __rem = __div64_32(&(n), __base);       \
      |                         ^~~~~~~~~~
./include/linux/math64.h:91:22: note: in expansion of macro 'do_div'
   91 |         *remainder = do_div(dividend, divisor);
      |                      ^~~~~~

This has passed correctness verification with test_div64 and reduced the 
module's average execution time down to 1.0404s from 1.0445s with R3400 
@40MHz.  The module's MIPS I machine code has also shrunk by 12 bytes or 
3 instructions.

Signed-off-by: Maciej W. Rozycki <macro@orcam.me.uk>
---
 arch/mips/include/asm/div64.h |    8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
diff mbox series

Patch

Index: linux-3maxp-div64/arch/mips/include/asm/div64.h
===================================================================
--- linux-3maxp-div64.orig/arch/mips/include/asm/div64.h
+++ linux-3maxp-div64/arch/mips/include/asm/div64.h
@@ -58,7 +58,6 @@ 
 
 #define __div64_32(n, base) ({						\
 	unsigned long __upper, __low, __high, __radix;			\
-	unsigned long long __modquot;					\
 	unsigned long long __quot;					\
 	unsigned long long __div;					\
 	unsigned long __mod;						\
@@ -73,11 +72,8 @@ 
 		__upper = __high;					\
 		__high = 0;						\
 	} else {							\
-		__asm__("divu	$0, %z1, %z2"				\
-		: "=x" (__modquot)					\
-		: "Jr" (__high), "Jr" (__radix));			\
-		__upper = __modquot >> 32;				\
-		__high = __modquot;					\
+		__upper = __high % __radix;				\
+		__high /= __radix;					\
 	}								\
 									\
 	__mod = do_div64_32(__low, __upper, __low, __radix);		\