This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [BENCHMARK]-mfpmath=sse should disable x387 intrinsics


Roger Sayle wrote:

Finally, you may find that if you want to use "-mfpmath=sse" effectively,
it may help to build a libm multi-lib (either sse-specific or soft-float)
that maximizes performance.  Again, most Linux distributions don't bother
with such a specialization as -fpmath=sse is so rarely a win.


I don't think its unreasonable for you to ask for this patch to be
reverted. An even better compromise is to only use this logic on
TARGET_64BIT where its a clear advantage by default. There's also
the complication that the *BSD support in the i386.c backend makes
it difficult to enable and disable 387 intrinsics independently.
Apparently, their kernel x87 emulator doesn't handle "fancy math",
so i386.c plays games with "-mfancy-math-387", such that
"-mno-fancy-math-387" no longer works on Pentium4, and the only way
to disable x87 intrinsics on the command-line it to use the corrected
"-mfpmath=sse".


I would like to present a small table of options for P4 and K8, attached as text attachment to this message, to preserve formatting.
I hope, that this table shows how various -mfpmath options should be handled for P4 and K8.


P4:
The default libm library is with i387 instructions, and the default is -mfpmath=i387. Unless a sse math library is present, there is no point to disable i387 intrinsics, even when -mfpmath=sse is specified. I suggest that -mfpmath=sse disables only "normal" i387 code (fmul,fadd, etc...), but leaves generation of intrinsics on. These intrinsics are disabled anyway without -ffast-math. -mno-fancy-math-387 should also disable these instructions.


K8:
The default libm uses sse instructions, and the default is -mfpmath=sse. Because (as Roger said) the default sse library is faster and more accurate than i387 intrinsics, even -mfpmath=i387 should _not_ generate i387 intrinsics, but should call sse math library. If someone wants to use i387 intrinsics, -mfancy-math-i387 (together with -ffast-math) should be specified.


In this case, -mno-fancy-math-387 does what it should - enables and disables use of "fancy" i387 instructions indepenently of -mfpmath choice. Actually, as it can be seen in attached table, the difference is only in default value of "x86_arch_always_fancy_math_387" variable, which should IMHO be cleared as a default choice for K8.

Uros.
P4:		-mfpmath=387	-mfpmath=sse	-mfpmath=sse,i387	-mno-fancy-math-387
		(default)
sse		-		x		x			?
x87		x		-		x			?
x87intr		x		x		x			-


K8:		-mfpmath=387	-mfpmath=sse	-mfpmath=sse,i387	-mfancy-math-387
				(default)
sse		-		x		x			?
x87		x		-		x			?
x87intr		-		-		-			x




Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]