Alex Susu via llvm-dev
2016-Jul-30 02:14 UTC
[llvm-dev] Instruction selection bug for vector store with FixedStack
Hello. Could you please help me solve the following LLC bug happening at instruction selection time: ISEL: Starting pattern match on root node: t172: ch = store<ST64[FixedStack6]> t0, t6, FrameIndex:i64<6>, undef:i64 Initial Opcode index to 157 Skipped scope entry (due to false predicate) at index 162, continuing at 236 Match failed at index 241 Continuing at 263 LLVM ERROR: Cannot select: t172: ch = store<ST64[FixedStack6]> t0, t6, FrameIndex:i64<6>, undef:i64 t6: v8i64,ch = CopyFromReg t0, Register:v8i64 %vreg19 t5: v8i64 = Register %vreg19 t171: i64 = FrameIndex<6> t73: i64 = undef In function: foo I actually have difficulties in understanding in the Selection DAG what the FixedStack operand means. Also, it is not really clear for me how to handle it in my back end. Below follows and excerpt of the Selection DAG with the problem: Optimized legalized selection DAG: BB#16 'foo:vector.body34' SelectionDAG has 71 nodes: t0: ch = EntryToken t6: v8i64,ch = CopyFromReg t0, Register:v8i64 %vreg19 t142: i64,ch = CopyFromReg t0, Register:i64 %vreg18 t143: i64 = add t142, Constant:i64<8> t172: ch = store<ST64[FixedStack6]> t0, t6, FrameIndex:i64<6>, undef:i64 t173: i64,ch = load<LD8[FixedStack6](align=64)> t172, FrameIndex:i64<6>, undef:i64 ... t127: ch = store<ST8[%aPtr.addr74]> t105:1, t109, Constant:i64<0>, undef:i64 t106: v8i64 = add t105, t79 t129: ch = store<ST64[inttoptr (i64 3 to <8 x i64>*)](align=8)(alias.scope=<0x20c8ea0>)(noalias=<0x20c8a80>,<0x20cd700>)> t127, t106, Constant:i64<3>, undef:i64 ... ... Thank you, Alex
Tim Northover via llvm-dev
2016-Aug-01 17:20 UTC
[llvm-dev] Instruction selection bug for vector store with FixedStack
Hi Alex, On 29 July 2016 at 19:14, Alex Susu via llvm-dev <llvm-dev at lists.llvm.org> wrote:> I actually have difficulties in understanding in the Selection DAG what > the FixedStack operand means. Also, it is not really clear for me how to > handle it in my back end.FixedStack refers to objects at a known location, created by MachineFrameInfo::CreateFixedObject rather than allocated an arbitrary stack slot. They're usually used for things like function parameters: you know exactly where they are on the stack. What you're seeing is the MachineMemOperand being printed. They typically don't take part in instruction selection, the FrameIndex is the most likely problem, though that's a bigger than average vector so maybe it should have been split?> Initial Opcode index to 157 > Skipped scope entry (due to false predicate) at index 162, continuing at 236 > Match failed at index 241 > Continuing at 263Assuming you've written a pattern that you think should be able to store a v8i64 (which ought to be enough, really), what I'd do now is look at build/lib/Target/XYZ/XYZGenDAGISel.inc. You should be able to correlate the debug info as above with your pattern and see just which bit of it is failing. Hopefully that'll tell you how to fix it. If I saw this on one of the in-tree targets I'd be suspicious that the C++ code in SelectWhateverAddrMode (from XYZISelDAGToDAG.cpp) was refusing to handle FrameIndex operands for the pointer (or maybe that an Operand<i32> had been used for something that's really a pointer). But you'll know more after looking at the .inc file. Cheers. Tim.
Maybe Matching Threads
- Instruction selection problems due to SelectionDAGBuilder
- llc gives Segmentation fault at instruction selection [was Re: Instruction selection gives "LLVM ERROR: Cannot select"]
- Spill hoisting on RAL: looking for some debugging ideas
- Instruction selection problem with type i64 - mistaken as v8i64?
- [LLVMdev] Help with definition of subregisters; spill, rematerialization and implicit uses