https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89721
Bug ID: 89721
Summary: __builtin_mffs sometimes optimized away
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: pc at gcc dot gnu.org
Target Milestone: ---
Created attachment 45969
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=45969&action=edit
test case where 3rd call to __builtin_mffs() is optimized away
I have a testcase that reports arguably incorrect results under any
optimization, correct results without optimization. This is with GCC 7.3, 8.2,
and trunk.
Testcase attached.
Without optimization:
--
$ ./mffs-bug
fpscr = 0x0000000000000002
fpscr = 0x0000000000000003
fpscr = 0x0000000000000001
--
With -O:
--
$ ./mffs-bug-O
fpscr = 0x0000000000000002
fpscr = 0x0000000000000003
fpscr = 0x0000000000000003
If I remove interesting pieces of the code above the last stanza, the problem
disappears.
The generated assembly is missing the final (of 3) "mffs" instructions. Below,
the instruction is expected roughly between 76c and 770:
--
...
10000718: li r3,2
1000071c: bl 10000560 <00000018.plt_call.fesetround@@GLIBC_2.17>
10000720: ld r2,24(r1)
10000724: mffs f0
10000728: addis r31,r2,-2
1000072c: addi r31,r31,-30144
10000730: mfvsrd r4,vs0
10000734: mr r3,r31
10000738: bl 10000540 <00000018.plt_call.printf@@GLIBC_2.17>
1000073c: ld r2,24(r1)
10000740: li r3,3
10000744: bl 10000560 <00000018.plt_call.fesetround@@GLIBC_2.17>
10000748: ld r2,24(r1)
1000074c: mffs f0
10000750: fmr f31,f0
10000754: mfvsrd r4,vs0
10000758: mr r3,r31
1000075c: bl 10000540 <00000018.plt_call.printf@@GLIBC_2.17>
10000760: ld r2,24(r1)
10000764: li r3,1
10000768: bl 10000560 <00000018.plt_call.fesetround@@GLIBC_2.17>
1000076c: ld r2,24(r1)
10000770: mfvsrd r4,vs31
10000774: mr r3,r31
10000778: bl 10000540 <00000018.plt_call.printf@@GLIBC_2.17>
--
