I came across this thread from a couple years ago:
http://lists.llvm.org/pipermail/llvm-dev/2017-November/119346.html
Has there been any progress on RDF for X86? Or is there some other preferred alternative for performing reachability analysis after register allocation?
Thanks,
Scott Constable
The one blocking issue that existed in the past has been fixed. I haven’t had time to do any work on it lately, but I’m not aware of any fundamental problems that would make it not work on x86.
Do you know whether it has been fixed on the 8.0.1 release?
Scott
Revisiting this thread. I have been experimenting with the RDF module on the X86 target. For the most part, building the data-flow graph and following def-use chains seems to work fine. But I am also observing some strange behavior in the phi nodes. For example, I have one basic block which begins with the following 4 phi nodes:
b585: — %bb.25 — preds(5): %bb.24, %bb.20, %bb.22, %bb.49, %bb.52 succs(2): %bb.18, %bb.37
p1095: phi [+d1096(,d668,):, u1097(d540,b477):u555, u1098(d427,b448):u450, u1099(d427,b462):u464, u1100(d427,b1009):u1011, u1101(d427,b1027):u1029]
p1102: phi [+d1103(,u1075):, u1104(d557,b477):u563, u1105(d411,b448):, u1106(d411,b462):u459, u1107(d411,b1009):u1106, u1108(d411,b1027):u1107]
p1109: phi [+d1110(,d625,u612):, u1111(d584,b477):, u1112(d452,b448):, u1113(d466,b462):, u1114(d1013,b1009):, u1115(d1031,b1027):]
p1116: phi [+d1117<#1073741833>(,d645,u648):, u1118"<#1073741833>(d579,b477):, u1466"<#1073741833>(d578,b477):, u1467"<#1073741833>(d569,b477):u581, u1468"<#1073741833>(d546,b477):, u1469"<#1073741833>(d543,b477):, u1470"<#1073741833>(d524,b477):u530, u1119"<#1073741833>(d453,b448):, u1436"<#1073741833>(d420,b448):, u1437"<#1073741833>(d404,b448):u445, u1120"<#1073741833>(d467,b462):, u1438"<#1073741833>(d420,b462):u1436", u1439"<#1073741833>(d404,b462):u473, u1121"<#1073741833>(d1014,b1009):, u1440"<#1073741833>(d420,b1009):u1006, u1441"<#1073741833>(d404,b1009):u1439", u1122"<#1073741833>(d1032,b1027):, u1442"<#1073741833>(d420,b1027):u1440", u1443"<#1073741833>(d404,b1027):u1441"]
The first three make perfect sense to me, and seem to reflect the post-allocation MIR correctly. The fourth phi node seems entirely composed of defs and uses for some unnamed register #1073741833 (what exactly is the significance of unnamed registers?). Moreover, this phi seems to be introducing false def-use relationships into the DFG. For example, the phi introduces the following dependency chain, which as far as I can tell is not valid:
// R11D is def’ed, def ID is d524
s523: SUB32rr [d524(d519,u1470"):, d525!(d520,d533,):, u526(d519):, u527(d508):]
…
// d524 is used in the phi node to def d1117, corresponding to unnamed register #1073741833
p1116: phi [+d1117<#1073741833>(,d645,u648):, u1118"<#1073741833>(d579,b477):, u1466"<#1073741833>(d578,b477):, u1467"<#1073741833>(d569,b477):u581, u1468"<#1073741833>(d546,b477):, u1469"<#1073741833>(d543,b477):, u1470"<#1073741833>(d524,b477):u530, u1119"<#1073741833>(d453,b448):, u1436"<#1073741833>(d420,b448):, u1437"<#1073741833>(d404,b448):u445, u1120"<#1073741833>(d467,b462):, u1438"<#1073741833>(d420,b462):u1436", u1439"<#1073741833>(d404,b462):u473, u1121"<#1073741833>(d1014,b1009):, u1440"<#1073741833>(d420,b1009):u1006, u1441"<#1073741833>(d404,b1009):u1439", u1122"<#1073741833>(d1032,b1027):, u1442"<#1073741833>(d420,b1027):u1440", u1443"<#1073741833>(d404,b1027):u1441"]
…
// d1117 is used in def d645 of register R9D
s644: ADD32rr [d645(+d1117,d694,u659):d634, d646!(d633,d651,):, u647(+d1117):u637, u648(+d1117):u647]
But after examining the corresponding MIR, I do not think that R11D flows into R9D. So it looks to me as though this phi node is erroneous.
Any help wold be much appreciated!
I’m using the LLVM 8.0.1 release.
Hi Scott,
That #1073741833 is a register mask. They are treated as aggregate registers (essentially sets of registers), so if it includes R9D and R11D, it will be treated as being aliased with both.
These separate defs are there because they reach disjoint registers.
Hi Krzysztof,
Thanks for the reply, I’m starting to understand the graph structure and terminology much better now, especially with the document in RDFGraph.h. I’m still a bit confused about some of the behavior I’m seeing with the phi nodes that involve aggregate registers. Here is one example:
b1531: — %bb.36 — preds(2): %bb.35, %bb.128 succs(1): %bb.37
p3193: phi [+d3194(,u3211):, u3195(+d3170,b1526):u1437, u3196(d2141,b2486):u2324]
p3197: phi [+d3198(,u3216):, u3199(+d3138,b1526):u1465, u3200(+d3364,b2486):]
p3201: phi [+d3202(,d1541,):, u3203(+d3146,b1526):, u3204(d1878,b2486):u3148]
p3205: phi [+d3206<#1073741833>(,d1551,u1552):, u3207"<#1073741833>(d1521,b1526):u1524, u3779"<#1073741833>(d1517,b1526):, u3780"<#1073741833>(d1513,b1526):u1522, u3208"<#1073741833>(d2481,b2486):u2484, u3695"<#1073741833>(d2477,b2486):, u3696"<#1073741833>(d2473,b2486):u2482]
s1532: ADJCALLSTACKDOWN64 [d1533!(+d3206,~d3647",u1557):, d1534!(+d3206,d1540,):d1533, d1535!(+d3206,~d3646",u1558):d1534, u1536!(+d3206):, u1537!(+d3206):u1536]
s1538: MOV32r0 [d1539(+d3202,):, d1540!(d1534,d1549,):, d1541(+d3202,u3221):d1539]
s1542: MOV32ri64 [d1543(+d3206,~d3645",u1561):d1535]
s1544: COPY [d1545(+d3206,~d3644",u1559):d1543, u1546(d785):]
s1547: MOV32r0 [d1548(+d3206,~d3643",u1560):d1545, d1549!(d1540,~d3642",):]
s1550: MOV64rm [d1551(+d3206,~d1554",u1562):d1548, u1552(+d3206):u1537]
s1553: CALL64pcrel32 __foo [~d1554"<#1073741833>!(d1551,d1579,):, ~d3642"<#1073741833>!(d1549,):, ~d3643"<#1073741833>!(d1548,):, ~d3644"<#1073741833>!(d1545,):, ~d3645"<#1073741833>!(d1543,):, ~d3646"<#1073741833>!(d1535,):, ~d3647"<#1073741833>!(d1533,):, d1555!(~d1554",d1564,u1567):, d1556!(~d1554",d1566,u1568):d1555, u1557!(d1533):, u1558!(d1535):, u1559!(d1545):, u1560!(d1548):, u1561!(d1543):, u1562!(d1551):]
s1563: ADJCALLSTACKUP64 [d1564!(d1555,u3778"):, d1565!(~d1554",d1571,):d1556, d1566!(d1556,u3777"):,
u1567!(d1555):, u1568!(d1556):]
s1569: MOV32r0 [d1570(~d1554",u3776"):d1565, d1571!(d1565,d1574,):]
s1572: MOV32r0 [d1573(~d1554",u3775"):d1570, d1574!(d1571,d1577,):]
s1575: MOV32r0 [d1576(~d1554",u3774"):d1573, d1577!(d1574,u3773"):]
—> s1578: MOV64rm [d1579(~d1554",u3226"):d1576]
b1580: — %bb.37 — preds(3): %bb.36, %bb.49, %bb.64 succs(1): %bb.38
p3209: phi [+d3210(,d1731,u3212):, u3211(+d3194,b1531):, u3212(+d3210,b1710):, u3213(d1857,b1874):u3466]
p3214: phi [+d3215(,u3217):, u3216(+d3198,b1531):, u3217(+d3215,b1710):u3257, u3218(+d3290,b1874):u3470]
p3219: phi [+d3220(,d1712,u1714):, u3221(d1541,b1531):, u3222(d1712,b1710):u1717, u3223(d1878,b1874):u3204]
—> p3224: phi [+d3225<#1073741833>(,d1598,):, u3226"<#1073741833>(d1579,b1531):, u3773"<#1073741833>(d1577,b1531):, u3774"<#1073741833>(d1576,b1531):, u3775"<#1073741833>(d1573,b1531):, u3776"<#1073741833>(d1570,b1531):, u3777"<#1073741833>(d1566,b1531):, u3778"<#1073741833>(d1564,b1531):, u3227<#1073741833>(d1716,b1710):u1719, u3228"<#1073741833>(d1885,b1874):u3751", u3754"<#1073741833>(d1880,b1874):u1887, u3755"<#1073741833>(d1865,b1874):u3752", u3756"<#1073741833>(d1861,b1874):u3753"]
s1581: IMUL64rri8 [d1582(+d3225,d1596,u1592):, d1583!(+d3225,d1595,):d1582, u1584(+d3220):]
s1585: LEA64r [d1586(+d3225,u3772"):d1583, u1587(d776):, u1588(d1582):]
s1589: LEA64r [d1590(+d3225,u3771"):d1586, u1591(+d3142):u1455, u1592(d1582):u1588]
s1593: MOV32r0 [d1594(d1582,):, d1595!(d1583,d1599,):, d1596(d1582,u3770"):d1594]
s1597: MOV32r0 [d1598(+d3225,u3769"):d1590, d1599!(d1595,u3231"):]
I have used arrows to highlight two nodes. The first node, s1578, def’s d1579, which has a single reached use in phi node p3224. I am surprised that this phi node exists and has no reached uses, for two reasons. First, I built the graph without the KeepDeadPhis option. Shouldn’t this remove phi nodes without reached uses? Second, there clearly is a reached use of R11 (corresponding to the instruction represented by s1578) in another basic block:
b1721: — %bb.50 — preds(1): %bb.49 succs(1): %bb.51
s1722: COPY [d1723(+d3248,u3765"):d1713, u1724(d785):u1546]
—> s1725: COPY [d1726(+d3248,u3764"):d1723, u1727(+d3248):]
s1728: MOV32r0 [d1729(+d3210,):, d1730!(d1716,u3261"):, d1731(+d3210,u3253):d1729]
I have confirmed via manual inspection and the use of a dynamic analysis tool that the next use of the R11 def in s1578 is s1725.
Could you please clarify these two points?
Thanks,
Scott
Hi Scott,
Sorry for the late reply, I was out of office during the holidays.
Hi Krzysztof,
I have been using RDF on x86 with a lot of success, though I have encountered one weird corner case on several occasions in getLandingPadLiveIns(). I have annotated the code below to highlight the issue:
RegisterSet DataFlowGraph::getLandingPadLiveIns() const {
RegisterSet LR;
const Function &F = MF.getFunction();
const Constant *PF = F.hasPersonalityFn() ? F.getPersonalityFn() // PF may be nullptr
: nullptr;
const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
if (RegisterId R = TLI.getExceptionPointerRegister(PF))
LR.insert(RegisterRef(R));
if (RegisterId R = TLI.getExceptionSelectorRegister(PF)) // when PF is nullptr, an assertion fails in the X86 subtarget impl
LR.insert(RegisterRef(R));
return LR;
}
The assertion fail then looks like this:
llvm/lib/Target/X86/X86ISelLowering.cpp:22945: virtual unsigned int llvm::X86TargetLowering::getExceptionSelectorRegister(const llvm::Constant *) const: Assertion `!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn))’ failed.
Could you clarify what the issue may be, and how it might be fixed? (BTW I am using LLVM 8.0.1)
Thanks,
Scott Constable
Hi Scott,
This code is trying to collect the registers that are live-in on entry to the landing pad. As you probably know, the landing pads are where the exception handling runtime jumps back into the function, so they are something like additional entry points. As such, they have some set of live-in registers as defined by some calling convention (or more likely exception handling convention). All this code does is trying to find out what these registers are. What this assertion reflects is really my lack of familiarity with these conventions for x86. This function is kind of “one size fits all”, i.e. “do the same thing for all cases”, which is evidently not sufficient. The proper fix would be to identify the applicable calling convention for a particular case and set the live-in registers appropriately.
Here’s a document about exception handling in LLVM with some information about landing pads and funclets: the actual registers depend on the target OS (Linux/Windows), and also on the architecture (I’m guessing this is 32- vs 64-bit).
https://llvm.org/docs/ExceptionHandling.html
I’m happy to hear that you’re funding RDF useful.