From 9e3d6223d2093a8903c8f570a06284453ee59944 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Fri, 9 Dec 2016 09:30:11 +0100 Subject: [PATCH] math64, timers: Fix 32bit mul_u64_u32_shr() and friends It turns out that while GCC-4.4 manages to generate 32x32->64 mult instructions for the 32bit mul_u64_u32_shr() code, any GCC after that fails horribly. Fix this by providing an explicit mul_u32_u32() function which can be architcture provided. Reported-by: Chris Metcalf Signed-off-by: Peter Zijlstra (Intel) Acked-by: Chris Metcalf [for tile] Cc: Christopher S. Hall Cc: David Gibson Cc: John Stultz Cc: Laurent Vivier Cc: Liav Rehana Cc: Linus Torvalds Cc: Parit Bhargava Cc: Peter Zijlstra Cc: Richard Cochran Cc: Thomas Gleixner Link: http://lkml.kernel.org/r/20161209083011.GD15765@worktop.programming.kicks-ass.net Signed-off-by: Ingo Molnar --- arch/tile/include/asm/Kbuild | 1 - arch/tile/include/asm/div64.h | 14 ++++++++++++++ arch/x86/include/asm/div64.h | 11 +++++++++++ include/linux/math64.h | 26 ++++++++++++++++++-------- 4 files changed, 43 insertions(+), 9 deletions(-) create mode 100644 arch/tile/include/asm/div64.h diff --git a/arch/tile/include/asm/Kbuild b/arch/tile/include/asm/Kbuild index 2d1f5638974c..20f2ba6d79be 100644 --- a/arch/tile/include/asm/Kbuild +++ b/arch/tile/include/asm/Kbuild @@ -5,7 +5,6 @@ generic-y += bug.h generic-y += bugs.h generic-y += clkdev.h generic-y += cputime.h -generic-y += div64.h generic-y += emergency-restart.h generic-y += errno.h generic-y += exec.h diff --git a/arch/tile/include/asm/div64.h b/arch/tile/include/asm/div64.h new file mode 100644 index 000000000000..bf6161966dfa --- /dev/null +++ b/arch/tile/include/asm/div64.h @@ -0,0 +1,14 @@ +#ifndef _ASM_TILE_DIV64_H +#define _ASM_TILE_DIV64_H + +#ifdef __tilegx__ +static inline u64 mul_u32_u32(u32 a, u32 b) +{ + return __insn_mul_lu_lu(a, b); +} +#define mul_u32_u32 mul_u32_u32 +#endif + +#include + +#endif /* _ASM_TILE_DIV64_H */ diff --git a/arch/x86/include/asm/div64.h b/arch/x86/include/asm/div64.h index ced283ac79df..af95c47d5c9e 100644 --- a/arch/x86/include/asm/div64.h +++ b/arch/x86/include/asm/div64.h @@ -59,6 +59,17 @@ static inline u64 div_u64_rem(u64 dividend, u32 divisor, u32 *remainder) } #define div_u64_rem div_u64_rem +static inline u64 mul_u32_u32(u32 a, u32 b) +{ + u32 high, low; + + asm ("mull %[b]" : "=a" (low), "=d" (high) + : [a] "a" (a), [b] "rm" (b) ); + + return low | ((u64)high) << 32; +} +#define mul_u32_u32 mul_u32_u32 + #else # include #endif /* CONFIG_X86_32 */ diff --git a/include/linux/math64.h b/include/linux/math64.h index 6e8b5b270ffe..80690c96c734 100644 --- a/include/linux/math64.h +++ b/include/linux/math64.h @@ -133,6 +133,16 @@ __iter_div_u64_rem(u64 dividend, u32 divisor, u64 *remainder) return ret; } +#ifndef mul_u32_u32 +/* + * Many a GCC version messes this up and generates a 64x64 mult :-( + */ +static inline u64 mul_u32_u32(u32 a, u32 b) +{ + return (u64)a * b; +} +#endif + #if defined(CONFIG_ARCH_SUPPORTS_INT128) && defined(__SIZEOF_INT128__) #ifndef mul_u64_u32_shr @@ -160,9 +170,9 @@ static inline u64 mul_u64_u32_shr(u64 a, u32 mul, unsigned int shift) al = a; ah = a >> 32; - ret = ((u64)al * mul) >> shift; + ret = mul_u32_u32(al, mul) >> shift; if (ah) - ret += ((u64)ah * mul) << (32 - shift); + ret += mul_u32_u32(ah, mul) << (32 - shift); return ret; } @@ -186,10 +196,10 @@ static inline u64 mul_u64_u64_shr(u64 a, u64 b, unsigned int shift) a0.ll = a; b0.ll = b; - rl.ll = (u64)a0.l.low * b0.l.low; - rm.ll = (u64)a0.l.low * b0.l.high; - rn.ll = (u64)a0.l.high * b0.l.low; - rh.ll = (u64)a0.l.high * b0.l.high; + rl.ll = mul_u32_u32(a0.l.low, b0.l.low); + rm.ll = mul_u32_u32(a0.l.low, b0.l.high); + rn.ll = mul_u32_u32(a0.l.high, b0.l.low); + rh.ll = mul_u32_u32(a0.l.high, b0.l.high); /* * Each of these lines computes a 64-bit intermediate result into "c", @@ -229,8 +239,8 @@ static inline u64 mul_u64_u32_div(u64 a, u32 mul, u32 divisor) } u, rl, rh; u.ll = a; - rl.ll = (u64)u.l.low * mul; - rh.ll = (u64)u.l.high * mul + rl.l.high; + rl.ll = mul_u32_u32(u.l.low, mul); + rh.ll = mul_u32_u32(u.l.high, mul) + rl.l.high; /* Bits 32-63 of the result will be in rh.l.low. */ rl.l.high = do_div(rh.ll, divisor);