diff options
Diffstat (limited to 'sysdeps/ia64/fpu/libm_scalblnf.S')
| -rw-r--r-- | sysdeps/ia64/fpu/libm_scalblnf.S | 384 |
1 files changed, 226 insertions, 158 deletions
diff --git a/sysdeps/ia64/fpu/libm_scalblnf.S b/sysdeps/ia64/fpu/libm_scalblnf.S index 362e68b641..af620d45f8 100644 --- a/sysdeps/ia64/fpu/libm_scalblnf.S +++ b/sysdeps/ia64/fpu/libm_scalblnf.S @@ -21,20 +21,20 @@ // products derived from this software without specific prior written // permission. -// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS -// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR -// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, -// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, -// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR -// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY +// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, +// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR +// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING -// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS -// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. -// +// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS +// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +// // Intel Corporation is the author of this code, and requests that all -// problem reports or change requests be submitted to it directly at +// problem reports or change requests be submitted to it directly at // http://www.intel.com/software/products/opensource/libraries/num.htm. // // History @@ -44,38 +44,51 @@ // 02/06/02 Corrected to handle 32- or 64-bit integers // 05/20/02 Cleaned up namespace and sf0 syntax // 02/10/03 Reordered header: .section, .global, .proc, .align +// 08/25/03 Improved performance // // API //============================================================== -// float = __libm_scalblnf (float x, long int n, int long_int_type) +// float __libm_scalblnf (float x, long int n, int long_int_type) // input floating point f8 and long int n (r33) // input long_int_type = 0 if long int defined as 32 bits, = 1 if 64 bits -// // output floating point f8 // - // Returns x* 2**n using an fma and detects overflow -// and underflow. +// and underflow. // // +// Strategy: +// Compute biased exponent of result exp_Result = N + exp_X +// Break into ranges: +// exp_Result > 0x1007e -> Certain overflow +// exp_Result = 0x1007e -> Possible overflow +// 0x0ff81 <= exp_Result < 0x1007e -> No over/underflow (main path) +// 0x0ff81 - 23 <= exp_Result < 0x0ff81 -> Possible underflow +// exp_Result < 0x0ff81 - 23 -> Certain underflow FR_Big = f6 FR_NBig = f7 FR_Floating_X = f8 FR_Result = f8 FR_Result2 = f9 -FR_Result3 = f11 -FR_Norm_X = f12 -FR_Two_N = f14 -FR_Two_to_Big = f15 +FR_Result3 = f10 +FR_Norm_X = f11 +FR_Two_N = f12 +GR_neg_ov_limit= r14 GR_N_Biased = r15 GR_Big = r16 GR_NBig = r17 -GR_Scratch = r18 -GR_Scratch1 = r19 +GR_exp_Result = r18 +GR_pos_ov_limit= r19 GR_Bias = r20 GR_N_as_int = r21 +GR_signexp_X = r22 +GR_exp_X = r23 +GR_exp_mask = r24 +GR_max_exp = r25 +GR_min_exp = r26 +GR_min_den_exp = r27 GR_SAVE_B0 = r32 GR_SAVE_GP = r33 @@ -93,243 +106,298 @@ GLOBAL_LIBM_ENTRY(__libm_scalblnf) // Build the exponent Bias // { .mfi - alloc r32=ar.pfs,3,0,4,0 - fclass.m.unc p7,p0 = FR_Floating_X, 0xe7 //@snan | @qnan | @inf | @zero - addl GR_Bias = 0x0FFFF,r0 + getf.exp GR_signexp_X = FR_Floating_X // Get signexp of x + fclass.m p6,p0 = FR_Floating_X, 0xe7 // @snan | @qnan | @inf | @zero + mov GR_Bias = 0x0ffff } - - // -// Is N zero? // Normalize x -// Do we need to sign extend input (long_int_type = 0)? +// Is long integer type 32 bits? // { .mfi - cmp.eq.unc p6,p0 = r33,r0 - fnorm.s1 FR_Norm_X = FR_Floating_X - cmp.eq.unc p8,p9 = r34,r0 + mov GR_Big = 35000 // If N this big then certain overflow + fnorm.s1 FR_Norm_X = FR_Floating_X + cmp.eq p8,p9 = r34,r0 } ;; -{ .mii -(p9) mov GR_N_as_int = r33 // Get n directly if long int 64 bits -(p8) sxt4 GR_N_as_int = r33 // Sign extend n if long int 32 bits - nop.i 0 +// Sign extend N if long int is 32 bits +{ .mfi +(p9) mov GR_N_as_int = r33 // Copy N if long int is 64 bits + fclass.m p9,p0 = FR_Floating_X, 0x0b // Test for x=unorm +(p8) sxt4 GR_N_as_int = r33 // Sign extend N if long int is 32 bits +} +{ .mfi + mov GR_NBig = -35000 // If N this small then certain underflow + nop.f 0 + mov GR_max_exp = 0x1007e // Exponent of maximum float +} +;; + +// Create biased exponent for 2**N +{ .mfi + add GR_N_Biased = GR_Bias,GR_N_as_int + nop.f 0 + cmp.ge p7, p0 = GR_N_as_int, GR_Big // Certain overflow? +} +{ .mib + cmp.le p8, p0 = GR_N_as_int, GR_NBig // Certain underflow? + mov GR_min_exp = 0x0ff81 // Exponent of minimum float +(p9) br.cond.spnt SCALBNF_UNORM // Branch if x=unorm +} +;; + +SCALBNF_COMMON: +// Main path continues. Also return here from x=unorm path. +// Create 2**N +.pred.rel "mutex",p7,p8 +{ .mfi + setf.exp FR_Two_N = GR_N_Biased + nop.f 0 +(p7) mov GR_N_as_int = GR_Big // Limit max N +} +{ .mfi +(p8) mov GR_N_as_int = GR_NBig // Limit min N + nop.f 0 +(p8) cmp.eq p7,p0 = r0,r0 // Set p7 if |N| big } ;; // -// Normalize x -// Branch and return special values. -// Create -35000 -// Create 35000 +// Create biased exponent for 2**N for N big +// Is N zero? // { .mfi - addl GR_Big = 35000,r0 +(p7) add GR_N_Biased = GR_Bias,GR_N_as_int nop.f 0 - add GR_N_Biased = GR_Bias,GR_N_as_int + cmp.eq.or p6,p0 = r33,r0 } -{ .mfb - addl GR_NBig = -35000,r0 -(p7) fma.s.s0 FR_Result = FR_Floating_X,f1, f0 -(p7) br.ret.spnt b0 -};; +{ .mfi + mov GR_pos_ov_limit = 0x1007f // Exponent for positive overflow + nop.f 0 + mov GR_exp_mask = 0x1ffff // Exponent mask +} +;; // -// Build the exponent Bias -// Return x when N = 0 +// Create 2**N for N big +// Return x when N = 0 or X = Nan, Inf, Zero // { .mfi - setf.exp FR_Two_N = GR_N_Biased +(p7) setf.exp FR_Two_N = GR_N_Biased nop.f 0 - addl GR_Scratch1 = 0x063BF,r0 + mov GR_min_den_exp = 0x0ff81 - 23 // Exponent of min denorm float } { .mfb - addl GR_Scratch = 0x019C3F,r0 -(p6) fma.s.s0 FR_Result = FR_Floating_X,f1, f0 -(p6) br.ret.spnt b0 -};; + and GR_exp_X = GR_exp_mask, GR_signexp_X +(p6) fma.s.s0 FR_Result = FR_Floating_X, f1, f0 +(p6) br.ret.spnt b0 +} +;; // -// Create 2*big -// Create 2**-big -// Is N > 35000 -// Is N < -35000 // Raise Denormal operand flag with compare -// Main path, create 2**N +// Compute biased result exponent // { .mfi - setf.exp FR_NBig = GR_Scratch1 - nop.f 0 - cmp.ge.unc p6, p0 = GR_N_as_int, GR_Big -} -{ .mfi - setf.exp FR_Big = GR_Scratch + add GR_exp_Result = GR_exp_X, GR_N_as_int fcmp.ge.s0 p0,p11 = FR_Floating_X,f0 - cmp.le.unc p8, p0 = GR_N_as_int, GR_NBig -};; + mov GR_neg_ov_limit = 0x3007f // Exponent for negative overflow +} +;; // -// Adjust 2**N if N was very small or very large +// Do final operation // { .mfi - nop.m 0 -(p6) fma.s1 FR_Two_N = FR_Big,f1,f0 - nop.i 0 + cmp.lt p7,p6 = GR_exp_Result, GR_max_exp // Test no overflow + fma.s.s0 FR_Result = FR_Two_N,FR_Norm_X,f0 + cmp.lt p9,p0 = GR_exp_Result, GR_min_den_exp // Test sure underflow } -{ .mlx - nop.m 999 - movl GR_Scratch = 0x000000000003007F -};; +{ .mfb + nop.m 0 + nop.f 0 +(p9) br.cond.spnt SCALBNF_UNDERFLOW // Branch if certain underflow +} +;; +{ .mib +(p6) cmp.gt.unc p6,p8 = GR_exp_Result, GR_max_exp // Test sure overflow +(p7) cmp.ge.unc p7,p9 = GR_exp_Result, GR_min_exp // Test no over/underflow +(p7) br.ret.sptk b0 // Return from main path +} +;; -{ .mfi - nop.m 0 -(p8) fma.s1 FR_Two_N = FR_NBig,f1,f0 - nop.i 0 +{ .bbb +(p6) br.cond.spnt SCALBNF_OVERFLOW // Branch if certain overflow +(p8) br.cond.spnt SCALBNF_POSSIBLE_OVERFLOW // Branch if possible overflow +(p9) br.cond.spnt SCALBNF_POSSIBLE_UNDERFLOW // Branch if possible underflow } -{ .mlx - nop.m 999 - movl GR_Scratch1= 0x000000000001007F -};; +;; + +// Here if possible underflow. +// Resulting exponent: 0x0ff81-23 <= exp_Result < 0x0ff81 +SCALBNF_POSSIBLE_UNDERFLOW: +// +// Here if possible overflow. +// Resulting exponent: 0x1007e = exp_Result +SCALBNF_POSSIBLE_OVERFLOW: -// Set up necessary status fields +// Set up necessary status fields // // S0 user supplied status // S2 user supplied status + WRE + TD (Overflows) // S3 user supplied status + FZ + TD (Underflows) // { .mfi - nop.m 999 + nop.m 0 fsetc.s3 0x7F,0x41 - nop.i 999 + nop.i 0 } { .mfi - nop.m 999 + nop.m 0 fsetc.s2 0x7F,0x42 - nop.i 999 -};; + nop.i 0 +} +;; // -// Do final operation +// Do final operation with s2 and s3 // { .mfi - setf.exp FR_NBig = GR_Scratch - fma.s.s0 FR_Result = FR_Two_N,FR_Norm_X,f0 - nop.i 999 + setf.exp FR_NBig = GR_neg_ov_limit + fma.s.s3 FR_Result3 = FR_Two_N,FR_Norm_X,f0 + nop.i 0 } { .mfi - nop.m 999 - fma.s.s3 FR_Result3 = FR_Two_N,FR_Norm_X,f0 - nop.i 999 -};; -{ .mfi - setf.exp FR_Big = GR_Scratch1 - fma.s.s2 FR_Result2 = FR_Two_N,FR_Norm_X,f0 - nop.i 999 -};; + setf.exp FR_Big = GR_pos_ov_limit + fma.s.s2 FR_Result2 = FR_Two_N,FR_Norm_X,f0 + nop.i 0 +} +;; // Check for overflow or underflow. // Restore s3 // Restore s2 // { .mfi - nop.m 0 + nop.m 0 fsetc.s3 0x7F,0x40 - nop.i 999 + nop.i 0 } { .mfi - nop.m 0 + nop.m 0 fsetc.s2 0x7F,0x40 - nop.i 999 -};; + nop.i 0 +} +;; // // Is the result zero? // { .mfi - nop.m 999 - fclass.m.unc p6, p0 = FR_Result3, 0x007 - nop.i 999 -} + nop.m 0 + fclass.m p6, p0 = FR_Result3, 0x007 + nop.i 0 +} { .mfi - addl GR_Tag = 205, r0 - fcmp.ge.unc.s1 p7, p8 = FR_Result2 , FR_Big - nop.i 0 -};; + nop.m 0 + fcmp.ge.s1 p7, p8 = FR_Result2 , FR_Big + nop.i 0 +} +;; // // Detect masked underflow - Tiny + Inexact Only // { .mfi - nop.m 999 + nop.m 0 (p6) fcmp.neq.unc.s1 p6, p0 = FR_Result , FR_Result2 - nop.i 999 -};; + nop.i 0 +} +;; // // Is result bigger the allowed range? // Branch out for underflow // { .mfb -(p6) addl GR_Tag = 206, r0 + nop.m 0 (p8) fcmp.le.unc.s1 p9, p10 = FR_Result2 , FR_NBig -(p6) br.cond.spnt scalbnf_UNDERFLOW -};; +(p6) br.cond.spnt SCALBNF_UNDERFLOW +} +;; // // Branch out for overflow // -{ .mbb - nop.m 0 -(p7) br.cond.spnt scalbnf_OVERFLOW -(p9) br.cond.spnt scalbnf_OVERFLOW -};; +{ .bbb +(p7) br.cond.spnt SCALBNF_OVERFLOW +(p9) br.cond.spnt SCALBNF_OVERFLOW + br.ret.sptk b0 // Return from main path. +} +;; -// -// Return from main path. -// -{ .mfb - nop.m 999 - nop.f 0 - br.ret.sptk b0;; +// Here if result overflows +SCALBNF_OVERFLOW: +{ .mib + alloc r32=ar.pfs,3,0,4,0 + addl GR_Tag = 205, r0 // Set error tag for overflow + br.cond.sptk __libm_error_region // Call error support for overflow } +;; -GLOBAL_LIBM_END(__libm_scalblnf) -__libm_error_region: +// Here if result underflows +SCALBNF_UNDERFLOW: +{ .mib + alloc r32=ar.pfs,3,0,4,0 + addl GR_Tag = 206, r0 // Set error tag for underflow + br.cond.sptk __libm_error_region // Call error support for underflow +} +;; + +// Here if x=unorm +SCALBNF_UNORM: +{ .mib + getf.exp GR_signexp_X = FR_Norm_X // Get signexp of normalized x + nop.i 0 + br.cond.sptk SCALBNF_COMMON // Return to main path +} +;; -scalbnf_OVERFLOW: -scalbnf_UNDERFLOW: + +GLOBAL_LIBM_END(__libm_scalblnf) +LOCAL_LIBM_ENTRY(__libm_error_region) // // Get stack address of N // .prologue { .mfi - add GR_Parameter_Y=-32,sp + add GR_Parameter_Y=-32,sp nop.f 0 .save ar.pfs,GR_SAVE_PFS - mov GR_SAVE_PFS=ar.pfs + mov GR_SAVE_PFS=ar.pfs } // -// Adjust sp +// Adjust sp // { .mfi .fframe 64 - add sp=-64,sp + add sp=-64,sp nop.f 0 - mov GR_SAVE_GP=gp + mov GR_SAVE_GP=gp };; // -// Store N on stack in correct position +// Store N on stack in correct position // Locate the address of x on stack // { .mmi - st8 [GR_Parameter_Y] = GR_N_as_int,16 - add GR_Parameter_X = 16,sp + st8 [GR_Parameter_Y] = GR_N_as_int,16 + add GR_Parameter_X = 16,sp .save b0, GR_SAVE_B0 - mov GR_SAVE_B0=b0 + mov GR_SAVE_B0=b0 };; // @@ -338,42 +406,42 @@ scalbnf_UNDERFLOW: // .body { .mib - stfs [GR_Parameter_X] = FR_Norm_X - add GR_Parameter_RESULT = 0,GR_Parameter_Y + stfs [GR_Parameter_X] = FR_Norm_X + add GR_Parameter_RESULT = 0,GR_Parameter_Y nop.b 0 } { .mib - stfs [GR_Parameter_Y] = FR_Result + stfs [GR_Parameter_Y] = FR_Result add GR_Parameter_Y = -16,GR_Parameter_Y - br.call.sptk b0=__libm_error_support# + br.call.sptk b0=__libm_error_support# };; // // Get location of result on stack // { .mmi + add GR_Parameter_RESULT = 48,sp nop.m 0 - nop.m 0 - add GR_Parameter_RESULT = 48,sp + nop.i 0 };; // -// Get the new result +// Get the new result // { .mmi - ldfs FR_Result = [GR_Parameter_RESULT] + ldfs FR_Result = [GR_Parameter_RESULT] .restore sp - add sp = 64,sp - mov b0 = GR_SAVE_B0 + add sp = 64,sp + mov b0 = GR_SAVE_B0 };; // // Restore gp, ar.pfs and return // { .mib - mov gp = GR_SAVE_GP - mov ar.pfs = GR_SAVE_PFS - br.ret.sptk b0 + mov gp = GR_SAVE_GP + mov ar.pfs = GR_SAVE_PFS + br.ret.sptk b0 };; LOCAL_LIBM_END(__libm_error_region) |