Improve LIR dumping #10140
Improve LIR dumping #10140
Conversation
1. Use the LIR node dumper to display nodes to be generated by codegen, since we're in LIR form at that point. Add a new "prefix message" argument to allow "Generating: " to prefix all such lines. 2. Fix off-by-one error in LIR dump due to `#ifdef` versus `#if`. 3. Remove extra trailing line for each LIR node. This interfered with #1. But I always thought it was unnecessarily verbose; I don't believe there is any ambiguity without that extra space. 4. Add dTreeLIR()/cTreeLIR() functions for use in the debugger.
|
@dotnet/jit-contrib PTAL |
| @@ -501,11 +501,13 @@ void CodeGen::genCodeForTreeNode(GenTreePtr treeNode) | |||
|
|
|||
| #ifdef DEBUG | |||
| lastConsumedNode = nullptr; | |||
| if (compiler->verbose) | |||
| { | |||
| unsigned seqNum = treeNode->gtSeqNum; // Useful for setting a conditional break in Visual Studio | |||
There was a problem hiding this comment.
Can you not set a breakpoint on treeNode->gtSeqNum itself?
There was a problem hiding this comment.
Seems like it. I just copy-and-pasted that from the other codegen files. Since it doesn't hurt, it made sense to keep it the same.
src/jit/gentree.cpp
Outdated
| assert(operand != nullptr); | ||
| assert(message != nullptr); | ||
|
|
||
| if (prefixMsg != nullptr) | ||
| { | ||
| printf("%s", prefixMsg); |
There was a problem hiding this comment.
Minor nit: you could capture the length of prefixMsg and use it to adjust the alignment below. I'd be interested to see how the codegen dumps looked in both cases.
There was a problem hiding this comment.
The issue is that the prefixMsg is output for both the operand (here), and the operator (below), and the operator doesn't have any leading space. And, I want the operator and operands to be indented by the same amount, when a prefixMsg is given.
Here's an example from codegen:
Generating: N037 ( 1, 4) [000034] ------------ t34 = const(h) int 0x8B6604 ftn REG NA
Generating: /--* t34 int
Generating: N039 ( 3, 6) [000035] ------------ t35 = * indir int REG NA
Generating: /--* t35 int control expr
Generating: N041 ( 69, 33) [000008] --C-G------- * call void System.Console.WriteLine $VN.Void
A "full dump" looks like:
/--* t32 int control expr
N029 ( 45, 22) [000005] --C-G------- t5 = * call float Program.f2 $c3
/--* t5 float
N031 ( 49, 25) [000021] DAC-G-----L- * st.lclVar float V02 tmp1 d:2 REG NA
N033 ( 3, 2) [000022] ------------ t22 = lclVar float V02 tmp1 u:2 (last use) REG NA $c3
/--* t22 float
N035 (???,???) [000033] ------------ * putarg_stk [+0x00] float REG NA
N037 ( 1, 4) [000034] ------------ t34 = const(h) int 0x8B6604 ftn REG NA
/--* t34 int
N039 ( 3, 6) [000035] ------------ t35 = * indir int REG NA
/--* t35 int control expr
N041 ( 69, 33) [000008] --C-G------- * call void System.Console.WriteLine $VN.Void
N043 ( 2, 2) [000013] ------------ il_offset void IL offset: 25 REG NA
N045 ( 1, 1) [000011] ------------ t11 = const int 100 REG NA $42
/--* t11 int
N047 ( 2, 2) [000012] ------------ * return int REG NA $1c0
There was a problem hiding this comment.
This is hugely picky of me, but is there any chance we could avoid printing the message before the operands? No big deal if there isn't.
|
@pgavlin @dotnet/jit-contrib any additional comments? |
|
@pgavlin Done. Sample codegen output now: |
|
LGTM. Thanks for making the change! |
1. Use the LIR node dumper to display nodes to be generated by codegen, since we're in LIR form at that point. Add a new "prefix message" argument to allow "Generating: " to prefix all such lines. 2. Fix off-by-one error in LIR dump due to `#ifdef` versus `#if`. 3. Remove extra trailing line for each LIR node. This interfered with dotnet#1. But I always thought it was unnecessarily verbose; I don't believe there is any ambiguity without that extra space. 4. Add dTreeLIR()/cTreeLIR() functions for use in the debugger.
codegen, since we're in LIR form at that point. Add a new
"prefix message" argument to allow "Generating: " to prefix all
such lines.
#ifdefversus#if.with (1). But I always thought it was unnecessarily verbose; I don't
believe there is any ambiguity without that extra space.