[PATCH] D37401: [MIParser] Make sure that getHexUint doesn't produce APInts with a bitwidth of 0
Jessica Paquette via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Fri Sep 1 14:23:26 PDT 2017
paquette created this revision.
Herald added subscribers: kristof.beyls, aemerson.
If getHexUint reads in a 0 in hex, it will create an APInt A with a value of 0. It will then use that APInt to populate Result. It calculates the bitwidth of Result using A.getActiveBits(). This function returns the bitwidth of A - the number of leading zeroes in A.
A is 0, so all of its bits are leading 0s. Thus, we try to construct Result with a bitwidth of 0, hitting an assert.
This appeared originally by compiling test-suite/SingleSource/Benchmarks/Misc/mandel.c like so:
clang -Oz -mno-red-zone -S -emit-llvm test-suite/SingleSource/Benchmarks/Misc/mandel.c -o /tmp/mandel.ll
llc -march=aarch64 -stop-after=patchable-function /tmp/mandel.ll -o /tmp/mandel.mir
llc -march=aarch64 /tmp/mandel.mir
If you have asserts enabled, this blows up because of a successor basic block being tagged with a probability of 0.
https://reviews.llvm.org/D37401
Files:
lib/CodeGen/MIRParser/MIParser.cpp
test/CodeGen/Generic/zero-probability.mir
Index: test/CodeGen/Generic/zero-probability.mir
===================================================================
--- /dev/null
+++ test/CodeGen/Generic/zero-probability.mir
@@ -0,0 +1,39 @@
+# RUN: llc -o /dev/null %s
+# REQUIRES: asserts
+# Makes sure that having a probability of 0x00000000 to branch to a successor
+# doesn't hit an APInt assert in the MIParser.
+
+--- |
+ define i32 @main() local_unnamed_addr #0 {
+ entry:
+ ret i32 0
+
+ other:
+ ret i32 0
+ }
+
+ attributes #0 = { nounwind }
+
+ !llvm.module.flags = !{!0, !1}
+ !llvm.ident = !{!2}
+
+ !0 = !{i32 1, !"wchar_size", i32 4}
+ !1 = !{i32 7, !"PIC Level", i32 2}
+ !2 = !{!"clang version 6.0.0"}
+ !3 = !{!"branch_weights", i32 0, i32 -1}
+
+...
+---
+name: main
+alignment: 2
+exposesReturnsTwice: false
+legalized: false
+regBankSelected: false
+selected: false
+tracksRegLiveness: true
+body: |
+ bb.0.entry:
+ successors: %bb.1.other(0x00000000)
+ bb.1.other:
+
+...
Index: lib/CodeGen/MIRParser/MIParser.cpp
===================================================================
--- lib/CodeGen/MIRParser/MIParser.cpp
+++ lib/CodeGen/MIRParser/MIParser.cpp
@@ -2083,8 +2083,11 @@
return true;
StringRef V = S.substr(2);
APInt A(V.size()*4, V, 16);
- Result = APInt(A.getActiveBits(),
- ArrayRef<uint64_t>(A.getRawData(), A.getNumWords()));
+
+ // If A is 0, then A.getActiveBits() is 0. This isn't a valid bitwidth. Make
+ // sure it isn't the case before constructing result.
+ unsigned NumBits = (A == 0) ? 32 : A.getActiveBits();
+ Result = APInt(NumBits, ArrayRef<uint64_t>(A.getRawData(), A.getNumWords()));
return false;
}
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D37401.113594.patch
Type: text/x-patch
Size: 1721 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170901/178a4c4b/attachment.bin>
More information about the llvm-commits
mailing list