Adjust expected result of test to match ARM and others

[ibuclaw]

From the conversation and comparing of results in #3483 - the following was found for the result of exp(-0x1.64p+9L):

Win64	ARM	ARM libm (expl)	x86 (double)	x86 libm (exp)	LLVM intrinsic	GCC intrinsic
0x0.06f84920bb2d3p-1022	0x0.06f84920bb2d4p-1022	0x0.06f84920bb2d4p-1022	0x0.06f84920bb2d4p-1022	0x0.06f84920bb2d4p-1022	0x0.06f84920bb2d4p-1022	0x0.06f84920bb2d4p-1022

With feqrel giving the following number of mantissa bits in common (53 means identical to 0x0.06f84920bb2d3p-1022L).

Win64	ARM	ARM libm (expl)	x86 (double)	x86 libm (exp)	LLVM intrinsic	GCC intrinsic
53	51	51	51*	51*	51	36*

[*] UPDATE: Whoops, I forgot to set rounding mode as per unittest. These are now correct, except for the GCC intrinsic, which const-folded the result. :-)

Changing the test to the mode result returned yields (53 means identical to 0x0.06f84920bb2d4p-1022L

Win64	ARM	ARM libm (expl)	x86 (double)	x86 libm (exp)	LLVM intrinsic	GCC intrinsic
51	53	53	53	53	53	53

So Win64 is the odd one out (but we are not sure why) and the rest produce the exact same result.

[kinke]

Without rechecking the LDC results with the new expected value, but given how feqrel() works for subnormals, Win64 should yield 51 and the LLVM intrinsic (at least on Win64) 53. I.e., the tests still pass. I'm 99.999% sure about that. ;)

[ibuclaw]

Thanks @kinke - updated.

[dnadlinger]

Auto-merge toggled on

[ibuclaw]

Thanks.