https://bugs.kde.org/show_bug.cgi?id=385055
--- Comment #2 from Carl Love <c...@us.ibm.com> ---
The second workload that hits the same issue is a GCC instruction test. The
test runs a whole bunch of float 128-bit instructions to test the code
generation.
In this case, there are several instructions: xsmulqp, xsmaddqp, xssubqp,
xsdivqp, xsaddqp which all call the function generate_store_FPRF( Ity_F128, vT
); to set the condition code for the instruction.
The function expands into:
t125 =
Or32(8Uto32(GET:I8(1342)),Shl32(8Uto32(GET:I8(1344)),0xC:I8))
t124 = And32(t125,0x3:I32)
t122 = GET:F128(896)
t121 = GET:F128(912)
t123 =
AddF128(Xor32(t124,And32(Shl32(t124,0x1:I8),0x2:I32)),t121,t122)
t134 =
64HLtoV128(ReinterpF64asI64(F128HItoF64(t123)),ReinterpF64asI64(F128LOtoF64(t123)))
t133 = 64to1(And64(Shr64(V128HIto64(t134),0x3F:I8),0x1:I64))
t136 = 0x7FFF000000000000:I64
t135 = 0xFFFFFFFFFFFF:I64
t137 = 0x0:I64
t126 =
32to1(And32(1Uto32(CmpEQ64(And64(V128HIto64(t134),t136),t136)),1Uto32(Not1(CmpEQ64(Or64(And64(V128HIto64(t134),t135),V128to64(t134)),t137)))))
t139 = 0x7FFF000000000000:I64
t138 = 0xFFFFFFFFFFFF:I64
t140 = 0x0:I64
t127 =
32to1(And32(1Uto32(CmpEQ64(And64(V128HIto64(t134),t139),t139)),1Uto32(CmpEQ64(Or64(And64(V128HIto64(t134),t138),V128to64(t134)),t140))))
t142 = 0x7FFF000000000000:I64
t141 = 0xFFFFFFFFFFFF:I64
t143 = 0x0:I64
t132 =
32to1(And32(1Uto32(CmpEQ64(And64(V128HIto64(t134),t142),t143)),1Uto32(CmpEQ64(Or64(And64(V128HIto64(t134),t141),V128to64(t134)),t143))))
t144 = 0x7FFF000000000000:I64
t145 = 0x0:I64
t129 =
32to1(And32(1Uto32(Not1(CmpEQ64(And64(V128HIto64(t134),t144),t145))),1Uto32(Not1(CmpEQ64(And64(V128HIto64(t134),t144),t144)))))
t147 = 0x7FFF000000000000:I64
t146 = 0xFFFFFFFFFFFF:I64
t148 = 0x0:I64
t128 =
32to1(And32(1Uto32(CmpEQ64(And64(V128HIto64(t134),t147),t148)),1Uto32(Not1(CmpEQ64(Or64(And64(V128HIto64(t134),t146),V128to64(t134)),t148)))))
t130 =
32to1(And32(1Uto32(32to1(Not32(1Uto32(t133)))),1Uto32(1:I1)))
t131 = 32to1(And32(1Uto32(t133),1Uto32(1:I1)))
PUT(1344) =
32to8(Or32(And32(0xF:I32,8Uto32(GET:I8(1344))),Shl32(And32(0x1:I32,1Uto32(32to1(Or32(1Uto32(32to1(Or32(1Uto32(t126),1Uto32(32to1(And32(1Uto32(t131)\
,1Uto32(t128))))))),1Uto32(32to1(Or32(1Uto32(32to1(And32(1Uto32(t131),1Uto32(t132)))),1Uto32(32to1(And32(1Uto32(t130),1Uto32(t128))))))))))),0x4:I8)))
PUT(1344) =
32to8(Or32(And32(0x10:I32,8Uto32(GET:I8(1344))),And32(0xF:I32,Or32(Or32(1Uto32(32to1(Or32(1Uto32(t126),1Uto32(t127)))),Shl32(1Uto32(32to1(And32(1Ut\
o32(32to1(Not32(1Uto32(t126)))),1Uto32(t132)))),0x1:I8)),Or32(Shl32(1Uto32(32to1(And32(1Uto32(32to1(Not32(1Uto32(t126)))),1Uto32(32to1(And32(1Uto32(32to1(Or32(1Uto32(32to1(O\
r32(1Uto32(32to1(And32(1Uto32(t130),1Uto32(t128)))),1Uto32(32to1(And32(1Uto32(t130),1Uto32(t129))))))),1Uto32(32to1(And32(1Uto32(t130),1Uto32(t127))))))),1Uto32(32to1(And32(\
1Uto32(32to1(Not32(1Uto32(t132)))),1Uto32(32to1(Not32(1Uto32(t126))))))))))))),0x2:I8),Shl32(1Uto32(32to1(And32(1Uto32(32to1(Not32(1Uto32(t126)))),1Uto32(32to1(And32(1Uto32(\
32to1(Or32(1Uto32(32to1(Or32(1Uto32(32to1(And32(1Uto32(t131),1Uto32(t128)))),1Uto32(32to1(And32(1Uto32(t131),1Uto32(t129))))))),1Uto32(32to1(And32(1Uto32(t131),1Uto32(t127))\
))))),1Uto32(32to1(And32(1Uto32(32to1(Not32(1Uto32(t132)))),1Uto32(32to1(Not32(1Uto32(t126))))))))))))),0x3:I8))))))
PUT(784) = t123
PUT(1296) = 0x4157DB0:I64
The basic block again seems to have about 30 instructions, with 6 of the
instructions having the above expansion for generate_store_FPRF(). This with
the added dres->hint = Dis_HintVerbose on each of these instructions.
The generate_store_FPRF() stores a condition code, which for this application
is not used so I commented out the body of the function to avoid calculating
the code and storing it. Once it is removed, the workload runs normally. So,
we either need to get the dres->hint to limit the BB more, perhaps end the BB
once it sees an instruction with the hint or perhaps use a C-code handler in
place of the generate_store_FPRF() function. Other thoughts?
--
You are receiving this mail because:
You are watching all bug changes.
