15228 – [3.4/4.0 Regression] useless copies of floating point operands (original) (raw)
Description Sylvain Pion 2004-04-30 15:41:53 UTC
Consider the following function :
// Forces a double to memory double force2memdouble(double x) { asm("" : "=m"(x) : "m"(x)); return x; }
This is used to force a double value in memory for platforms like x86 where the FP registers have too much precision.
It gets compiled by GCC (3.4 and 3.3.3, with -O2 -fomit-frame-pointer) to :
force2memdouble: subl $12, %esp fldl 16(%esp) fstpl (%esp) fldl (%esp) addl $12, %esp ret
I believe that the store/load is useless, and that the function could be compiled as simply :
force2memdouble: fldl 4(%esp) ret
It is interesting to note that when using "int", "float" or "long double" instead of "double", then there is no such useless store/load.
So, is this something that could be improved in GCC ?
Maybe the code responsible for asm() adds too many requirements on the operands for doubles ?
Comment 1 Sylvain Pion 2004-04-30 15:51:07 UTC
I still have gcc 2.95 around, so I have tested with it as well. It produces the better code that I expected :
force2memdouble: #APP #NO_APP fldl 4(%esp) ret
So, if the code is correct, then this problem is a regression on code quality compared to 2.95.
I want to mention that this asm() is critical when implementing interval arithmetic. Also, GMP uses that kind of asm() in mpz_get_d_2exp.
Comment 2 Drea Pinski 2004-04-30 15:57:08 UTC
Note before 3.3 and after 2.95.3, the following sequence was used to copy the FP: subl $12, %esp movl 16(%esp), %edx movl 20(%esp), %ecx movl %edx, (%esp) movl %ecx, 4(%esp)
Confirmed as a regression.
Comment 3 Gabriel Dos Reis 2004-05-16 22:11:39 UTC
Will not fix in 3.3.x.
-- Gaby
Comment 4 Mark Mitchell 2004-05-31 22:10:04 UTC
Richard --
Would you please take a look at this?
Thanks,
-- Mark
Comment 5 Richard Henderson 2004-06-07 19:27:33 UTC
Testing a patch.
Comment 8 Richard Henderson 2004-06-09 23:52:49 UTC
Fixed.