The license of glibc requires referenced code to be open-source, but amath is closed-source. Are there any license issues with using it?， exsample expf， While the register names differ, the computational logic, conditions, and the arm_math_exp_data remain entirely identical.
arm_math_exp_data
__expf_finite@GLIBC_2.17: fmov w0, s0 fcvt d2, s0 ubfx x1, x0, #20, #11 cmp w1, #0x42a b.hi 3fd78 <__expf_finite@GLIBC_2.17+0x68> // b.pmore adrp x0, 78000 <f64xsubf128@@GLIBC_2.28+0x29250> add x0, x0, #0x830 fmov d1, #1.000000000000000000e+00 ldp d0, d4, [x0, #296] ldr d3, [x0, #312] fmul d2, d2, d0 ldr d0, [x0, #320] frinta d5, d2 fcvtas x1, d2 fsub d2, d2, d5 and x2, x1, #0x1f fmadd d3, d4, d2, d3 fmadd d0, d0, d2, d1 fmul d2, d2, d2 ldr x0, [x0, x2, lsl #3] add x1, x0, x1, lsl #47 fmov d1, x1 fmadd d0, d3, d2, d0 fmul d0, d0, d1 fcvt s0, d0 ret cmn w0, #0x800, lsl #12 b.eq 3fde0 <__expf_finite@GLIBC_2.17+0xd0> // b.none cmp w1, #0x7f7 b.hi 3fdd8 <__expf_finite@GLIBC_2.17+0xc8> // b.pmore mov w0, #0x7217 // #29207 movk w0, #0x42b1, lsl #16 fmov s1, w0 fcmpe s0, s1 b.gt 3fdc8 <__expf_finite@GLIBC_2.17+0xb8> mov w0, #0xf1b4 // #61876 movk w0, #0xc2cf, lsl #16 fmov s1, w0 fcmpe s0, s1 b.mi 3fde8 <__expf_finite@GLIBC_2.17+0xd8> // b.first mov w0, #0x8ecf // #36559 movk w0, #0xc2ce, lsl #16 fmov s1, w0 fcmpe s0, s1 b.mi 3fdd0 <__expf_finite@GLIBC_2.17+0xc0> // b.first b 3fd24 <__expf_finite@GLIBC_2.17+0x14> mov w0, #0x0 // #0 b 4cc00 <__exp2f_finite@GLIBC_2.17+0x1e90> mov w0, #0x0 // #0 b 4cbf0 <__exp2f_finite@GLIBC_2.17+0x1e80> fadd s0, s0, s0 ret movi v0.2s, #0x0 ret mov w0, #0x0 // #0 b 4cbe0 <__exp2f_finite@GLIBC_2.17+0x1e70>
armpl_exp_f32: fmov w0, s0 fcvt d31, s0 ubfx x1, x0, #20, #11 cmp w1, #0x42a b.hi 159e4 <armpl_exp_f32+0x64> // b.pmore adrp x0, 5f000 <arm_math_exp_data+0x570> add x0, x0, #0x7b0 fmov d0, #1.000000000000000000e+00 ldp d30, d28, [x0, #288] fmul d31, d31, d30 ldp d30, d29, [x0, #304] frinta d27, d31 fcvtas x1, d31 fsub d31, d31, d27 and w2, w1, #0x1f fmadd d30, d28, d31, d30 fmadd d0, d29, d31, d0 fmul d31, d31, d31 ldr x0, [x0, w2, sxtw #3] add x1, x0, x1, lsl #47 fmadd d0, d30, d31, d0 fmov d31, x1 fmul d0, d0, d31 fcvt s0, d0 ret cmn w0, #0x800, lsl #12 b.eq 15a4c <armpl_exp_f32+0xcc> // b.none cmp w1, #0x7f7 b.hi 15a44 <armpl_exp_f32+0xc4> // b.pmore mov w0, #0x7217 // #29207 movk w0, #0x42b1, lsl #16 fmov s30, w0 fcmpe s0, s30 b.gt 15a34 <armpl_exp_f32+0xb4> mov w0, #0xf1b4 // #61876 movk w0, #0xc2cf, lsl #16 fmov s30, w0 fcmpe s0, s30 b.mi 15a54 <armpl_exp_f32+0xd4> // b.first mov w0, #0x8ecf // #36559 movk w0, #0xc2ce, lsl #16 fmov s30, w0 fcmpe s0, s30 b.mi 15a3c <armpl_exp_f32+0xbc> // b.first b 15994 <armpl_exp_f32+0x14> mov w0, #0x0 // #0 b 54a00 <arm_math_oflowf> mov w0, #0x0 // #0 b 549e8 <arm_math_may_uflowf> fadd s0, s0, s0 ret movi v0.2s, #0x0 ret mov w0, #0x0 // #0 b 549e0 <arm_math_uflowf> nop
Hi.
Thanks for gettting in touch. Yes, the code is very similar and that is done on purpose, precisely to avoid licence issues. We actually upstream the code first into Arm Optimized Routines, and that purposefully has a very permissive licence. From there we can then use it in our own products, such as libamath, and also upstream into other projects such as glibc. As you identified GNU's GPL licence is more restrictive, hence why it lands first in AOR.
If you are looking to take source code directly for use in a non-GPL project you can use our AOR version, or link against libamath from ArmPL.
Hope this helps.
Chris
Thanks