From 2ec750ca05a87ebc20e358e73cb375a2c844ad59 Mon Sep 17 00:00:00 2001 From: Eugen Wissner Date: Tue, 8 May 2018 17:45:51 +0200 Subject: [PATCH] Fix math.nbtheory linkage to asm Don't use extern for templated functions. If the function argument is const, it gets a different mangling. So define a private function for each floatint point length and call it from template. --- arch/x64/linux/math/abs.S | 18 ++++++------ arch/x64/linux/math/log.S | 40 +++++++++++++-------------- source/tanya/math/nbtheory.d | 53 +++++++++++++++++++++--------------- 3 files changed, 60 insertions(+), 51 deletions(-) diff --git a/arch/x64/linux/math/abs.S b/arch/x64/linux/math/abs.S index edf8ac9..6635273 100644 --- a/arch/x64/linux/math/abs.S +++ b/arch/x64/linux/math/abs.S @@ -2,10 +2,10 @@ // fabsf. - .globl _D5tanya4math8nbtheory10__T3absTfZ3absFNaNbNiNffZf - .type _D5tanya4math8nbtheory10__T3absTfZ3absFNaNbNiNffZf, @function + .globl _D5tanya4math8nbtheory4fabsFNaNbNiNffZf + .type _D5tanya4math8nbtheory4fabsFNaNbNiNffZf, @function -_D5tanya4math8nbtheory10__T3absTfZ3absFNaNbNiNffZf: +_D5tanya4math8nbtheory4fabsFNaNbNiNffZf: mov $0x7fffffff, %eax movq %rax, %xmm1 andpd %xmm1, %xmm0 @@ -13,10 +13,10 @@ _D5tanya4math8nbtheory10__T3absTfZ3absFNaNbNiNffZf: // fabs. - .globl _D5tanya4math8nbtheory10__T3absTdZ3absFNaNbNiNfdZd - .type _D5tanya4math8nbtheory10__T3absTdZ3absFNaNbNiNfdZd, @function + .globl _D5tanya4math8nbtheory4fabsFNaNbNiNfdZd + .type _D5tanya4math8nbtheory4fabsFNaNbNiNfdZd, @function -_D5tanya4math8nbtheory10__T3absTdZ3absFNaNbNiNfdZd: +_D5tanya4math8nbtheory4fabsFNaNbNiNfdZd: mov $0x7fffffffffffffff, %rax movq %rax, %xmm1 andpd %xmm1, %xmm0 @@ -24,12 +24,12 @@ _D5tanya4math8nbtheory10__T3absTdZ3absFNaNbNiNfdZd: // fabsl. - .globl _D5tanya4math8nbtheory10__T3absTeZ3absFNaNbNiNfeZe - .type _D5tanya4math8nbtheory10__T3absTeZ3absFNaNbNiNfeZe, @function + .globl _D5tanya4math8nbtheory4fabsFNaNbNiNfeZe + .type _D5tanya4math8nbtheory4fabsFNaNbNiNfeZe, @function // Load the parameter from the stack onto FP stack, execute 'fabs' instruction // The result is returned in ST0. -_D5tanya4math8nbtheory10__T3absTeZ3absFNaNbNiNfeZe: +_D5tanya4math8nbtheory4fabsFNaNbNiNfeZe: fldt 0x8(%rsp) fabs ret diff --git a/arch/x64/linux/math/log.S b/arch/x64/linux/math/log.S index 941e399..9c5ed35 100644 --- a/arch/x64/linux/math/log.S +++ b/arch/x64/linux/math/log.S @@ -1,22 +1,29 @@ .text -// logl. - .globl _D5tanya4math8nbtheory9__T2lnTeZ2lnFNaNbNiNfeZe - .type _D5tanya4math8nbtheory9__T2lnTeZ2lnFNaNbNiNfeZe, @function +// logf. + .globl _D5tanya4math8nbtheory4logfFNaNbNiNffZf + .type _D5tanya4math8nbtheory4logfFNaNbNiNffZf, @function + +_D5tanya4math8nbtheory4logfFNaNbNiNffZf: + movss %xmm0, -4(%rsp) // Put the argument onto the stack -_D5tanya4math8nbtheory9__T2lnTeZ2lnFNaNbNiNfeZe: fldln2 // Put lb(e) onto the FPU stack - fldt 8(%rsp) // Put the argument onto the FPU stack + flds -4(%rsp) // Put a float onto the FPU stack fyl2x // %st1 * lb(%st0) + + // The result is on the FPU stack, but returned in %xmm0 + fstps -4(%rsp) + movss -4(%rsp), %xmm0 + ret // log. - .globl _D5tanya4math8nbtheory9__T2lnTdZ2lnFNaNbNiNfdZd - .type _D5tanya4math8nbtheory9__T2lnTdZ2lnFNaNbNiNfdZd, @function + .globl _D5tanya4math8nbtheory3logFNaNbNiNfdZd + .type _D5tanya4math8nbtheory3logFNaNbNiNfdZd, @function -_D5tanya4math8nbtheory9__T2lnTdZ2lnFNaNbNiNfdZd: +_D5tanya4math8nbtheory3logFNaNbNiNfdZd: movsd %xmm0, -8(%rsp) // Put the argument onto the stack fldln2 // Put lb(e) onto the FPU stack @@ -30,19 +37,12 @@ _D5tanya4math8nbtheory9__T2lnTdZ2lnFNaNbNiNfdZd: ret -// logf. - .globl _D5tanya4math8nbtheory9__T2lnTfZ2lnFNaNbNiNffZf - .type _D5tanya4math8nbtheory9__T2lnTfZ2lnFNaNbNiNffZf, @function - -_D5tanya4math8nbtheory9__T2lnTfZ2lnFNaNbNiNffZf: - movss %xmm0, -4(%rsp) // Put the argument onto the stack +// logl. + .globl _D5tanya4math8nbtheory4loglFNaNbNiNfeZe + .type _D5tanya4math8nbtheory4loglFNaNbNiNfeZe, @function +_D5tanya4math8nbtheory4loglFNaNbNiNfeZe: fldln2 // Put lb(e) onto the FPU stack - flds -4(%rsp) // Put a float onto the FPU stack + fldt 8(%rsp) // Put the argument onto the FPU stack fyl2x // %st1 * lb(%st0) - - // The result is on the FPU stack, but returned in %xmm0 - fstps -4(%rsp) - movss -4(%rsp), %xmm0 - ret diff --git a/source/tanya/math/nbtheory.d b/source/tanya/math/nbtheory.d index 889e10f..c20cc67 100644 --- a/source/tanya/math/nbtheory.d +++ b/source/tanya/math/nbtheory.d @@ -16,9 +16,17 @@ module tanya.math.nbtheory; import tanya.math.mp; import tanya.meta.trait; +import tanya.meta.transform; version (TanyaNative) { + private extern float fabs(float) @nogc nothrow pure @safe; + private extern double fabs(double) @nogc nothrow pure @safe; + private extern real fabs(real) @nogc nothrow pure @safe; + + private extern double log(double) @nogc nothrow pure @safe; + private extern float logf(float) @nogc nothrow pure @safe; + private extern real logl(real) @nogc nothrow pure @safe; } else { @@ -35,7 +43,7 @@ else * * Returns: Absolute value of $(D_PARAM x). */ -T abs(T)(T x) +Unqual!T abs(T)(T x) if (isIntegral!T) { static if (isSigned!T) @@ -60,24 +68,11 @@ if (isIntegral!T) static assert(is(typeof(u.abs) == uint)); } -version (D_Ddoc) +/// ditto +Unqual!T abs(T)(T x) +if (isFloatingPoint!T) { - /// ditto - T abs(T)(T x) - if (isFloatingPoint!T); -} -else version (TanyaNative) -{ - extern T abs(T)(T number) @nogc nothrow pure @safe - if (isFloatingPoint!T); -} -else -{ - T abs(T)(T x) - if (isFloatingPoint!T) - { - return fabs(cast(real) x); - } + return fabs(x); } /// @@ -122,17 +117,31 @@ version (D_Ddoc) * * Returns: Natural logarithm of $(D_PARAM x). */ - T ln(T)(T x) + Unqual!T ln(T)(T x) if (isFloatingPoint!T); } else version (TanyaNative) { - extern T ln(T)(T x) @nogc nothrow pure @safe - if (isFloatingPoint!T); + Unqual!T ln(T)(T x) @nogc nothrow pure @safe + if (isFloatingPoint!T) + { + static if (is(Unqual!T == float)) + { + return logf(x); + } + else static if (is(Unqual!T == double)) + { + return log(x); + } + else + { + return logl(x); + } + } } else { - T ln(T)(T x) + Unqual!T ln(T)(T x) if (isFloatingPoint!T) { return log(x);