https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64342
Jeffrey A. Law <law at redhat dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |law at redhat dot com
Assignee|unassigned at gcc dot gnu.org |vmakarov at redhat dot
com
--- Comment #13 from Jeffrey A. Law <law at redhat dot com> ---
AFACIT only one issue is left, namely avx512f-kandnw-1.c when compiled with
-mavx512f -m32 does not generate a kandnw instruction.
This appears to be some kind of register allocation issue.
In the IRA dump we have:
ssigning 69 to a1r105
Assigning 70 to a5r103
Assigning 69 to a6r101
Disposition:
10:r87 l0 0 3:r91 l0 22 4:r92 l0 23 2:r93 l0 21
9:r100 l0 21 6:r101 l0 69 8:r102 l0 0 5:r103 l0 70
7:r104 l0 1 1:r105 l0 69 0:r110 l0 21
Which is basically what we want for insn 17 to generate the proper insn:
(insn 17 11 19 2 (parallel [
(set (reg:HI 105)
(and:HI (not:HI (reg:HI 101 [ k1 ]))
(reg:HI 103 [ k2 ])))
(clobber (reg:CC 17 flags))
]) ./include/avx512fintrin.h:9995 388 {kandnhi}
(expr_list:REG_DEAD (reg:HI 103 [ k2 ])
(expr_list:REG_DEAD (reg:HI 101 [ k1 ])
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil)))))
LRA comes along and mucks things up. It seems to initially do the right thing:
Choosing alt 2 in insn 17: (0) !k (1) k (2) k {kandnhi}
Then later:
********** Assignment #1: **********
Spill r101 after risky transformations
Assigning to 117 (cl=MASK_EVEX_REGS, orig=105, freq=2000, tfirst=117,
tfreq=2000)...
Assign 71 to reload r117 (freq=2000)
Reassigning non-reload pseudos
Assign 0 to r101 (freq=2000)
Note the Assign 0 to r101, at this point we're going to lose badly.
Choosing alt 1 in insn 17: (0) &r (1) 0 (2) r {kandnhi}
Creating newreg=118 from oldreg=105, assigning class GENERAL_REGS to r118
Creating newreg=119 from oldreg=103, assigning class GENERAL_REGS to r119
With insn 17 looking like:
(insn 17 40 39 2 (parallel [
(set (reg:HI 0 ax [105])
(and:HI (not:HI (reg:HI 0 ax [105]))
(reg:HI 2 cx [orig:103 k2 ] [103])))
(clobber (reg:CC 17 flags))
]) ./include/avx512fintrin.h:9995 388 {kandnhi}
(nil))
with input & output reloads to get things into the k registers and of course no
kandnw instruction since we end up doing the logical operation in the general
purpose registers.
