Skip to content

Set some missing GTF_GLOB_REF and improve its checking. - #61245

Closed
sandreenko wants to merge 4 commits into
dotnet:mainfrom
sandreenko:improveFlagsCheck
Closed

Set some missing GTF_GLOB_REF and improve its checking.#61245
sandreenko wants to merge 4 commits into
dotnet:mainfrom
sandreenko:improveFlagsCheck

Conversation

@sandreenko

@sandreenkosandreenko commented Nov 5, 2021•
edited
Loading

Copy link
Copy Markdown
Contributor

The code in fgDebugCheckFlags had some mistakes that were easy to fix, for harder errors I added a comment explaining them.

The diff is a regression but a small one.

The diffs look like correctness fixes, for example, for such comparison:

Core_Root\superpmi.exe "Core_Root_base\clrjit.dll" "Core_Root_diff\clrjit.dll" D:\Sergey\git\runtime\artifacts\spmi\mch\3df3e3ec-b1f6-4e6d-8439-2e7f3f7fa2ac.windows.x64\libraries_tests.pmi.windows.x64.checked.mch -a -c 78772

we forbid swapping call and IND(global static field), if I understand correctly, nothing prevents the call from making a change in this field, so it should not be hard to create a correctness repro for this bug.

***** BB01
++STMT00002 (IL ???... ???)++ [000013] -A--G------- * ASG byref ++ [000012] D------N---- +--* LCL_VAR byref V03 tmp2 ++ [000003] ----G------- \--* ADD byref ++ [000001] n---G------- +--* IND ref ++ [000000] I----------- | \--* CNS_INT(h) long 0xd1ffab1e static Fseq[hackishFieldName]++ [000002] ------------ \--* CNS_INT long 8 Fseq[#FirstElem]
***** BB01
STMT00001 (IL ???... ???)
[000011] I-C-G------- * CALL void System.TimeSpan..ctor (exactContextHnd=0x00000000D1FFAB1E)
[000010] ------------ this in rcx +--* ADDR byref [000009] -------N---- | \--* LCL_VAR struct<System.TimeSpan, 8> V02 tmp1 [000005] ------------ arg1 \--* CNS_INT long 1
***** BB01
STMT00002 (IL 0x00C... ???)
-- [000013] I-C-G------- * CALL struct System.TimeSpan.Add (exactContextHnd=0x00000000D1FFAB1E)-- [000003] ------------ this in rcx +--* ADD byref -- [000001] ------------ | +--* IND ref -- [000000] I----------- | | \--* CNS_INT(h) long 0xd1ffab1e static Fseq[hackishFieldName]-- [000002] ------------ | \--* CNS_INT long 8 Fseq[#FirstElem]-- [000015] n----------- arg1 \--* OBJ struct<System.TimeSpan, 8>-- [000014] ------------ \--* ADDR byref -- [000012] -------N---- \--* LCL_VAR struct<System.TimeSpan, 8> V02 tmp1 ++STMT00003 (IL 0x00C... ???)++ [000016] I-C-G------- * CALL struct System.TimeSpan.Add (exactContextHnd=0x00000000D1FFAB1E)++ [000014] ------------ this in rcx +--* LCL_VAR byref V03 tmp2 ++ [000018] n----------- arg1 \--* OBJ struct<System.TimeSpan, 8>++ [000017] ------------ \--* ADDR byref ++ [000015] -------N---- \--* LCL_VAR struct<System.TimeSpan, 8> V02 tmp1

so if inside [000011] I-C-G------- * CALL void System.TimeSpan..ctor we change what is stored in

[000001] n---G------- +--* IND ref [000000] I----------- | \--* CNS_INT(h) long 0xd1ffab1e static Fseq[hackishFieldName]

then before this change, we would read the changed value, not the correct one.

diffs with misleading total numbers Ignore the total diff, see #60124
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.x64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.x64.checked.4\diff
Total bytes of delta: -6956 (-0.10 % of base)
1 total files with Code Size differences (0 improved, 1 regressed), 1 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.x64.checked.5\diff
Total bytes of delta: -1515 (-0.00 % of base)
5 total files with Code Size differences (1 improved, 4 regressed), 4 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.x64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.x64.checked.4\diff
Total bytes of delta: -3695 (-0.01 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 2 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.x64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.x64.checked.4\diff
Total bytes of delta: -21142 (-0.02 % of base)
19 total files with Code Size differences (2 improved, 17 regressed), 2 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.arm64.checked.4\diff
Total bytes of delta: -1332 (-0.00 % of base)
1 total files with Code Size differences (0 improved, 1 regressed), 3 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.arm64.checked.4\diff
Total bytes of delta: -2920 (-0.01 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 1 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.arm64.checked.4\diff
Total bytes of delta: -20104 (-0.02 % of base)
17 total files with Code Size differences (0 improved, 17 regressed), 4 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.Linux.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.Linux.x64.checked.5\diff
Total bytes of delta: -7003 (-0.08 % of base)
1 total files with Code Size differences (0 improved, 1 regressed), 0 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.x64.checked.5\diff
Total bytes of delta: -1295 (-0.00 % of base)
5 total files with Code Size differences (1 improved, 4 regressed), 4 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.x64.checked.5\diff
Total bytes of delta: -2909 (-0.01 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 1 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.x64.checked.5\diff
Total bytes of delta: -23240 (-0.02 % of base)
19 total files with Code Size differences (3 improved, 16 regressed), 2 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.arm64.checked.4\diff
Total bytes of delta: -6992 (-0.09 % of base)
0 total files with Code Size differences (0 improved, 0 regressed), 2 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.arm64.checked.4\diff
Total bytes of delta: -1332 (-0.00 % of base)
1 total files with Code Size differences (0 improved, 1 regressed), 3 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.arm64.checked.4\diff
Total bytes of delta: -3848 (-0.01 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 2 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.arm64.checked.4\diff
Total bytes of delta: -20100 (-0.02 % of base)
17 total files with Code Size differences (0 improved, 17 regressed), 4 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.x86.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.x86.checked.3\diff
Total bytes of delta: -4922 (-0.09 % of base)
0 total files with Code Size differences (0 improved, 0 regressed), 3 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.x86.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.x86.checked.3\diff
Total bytes of delta: -2182 (-0.00 % of base)
7 total files with Code Size differences (0 improved, 7 regressed), 6 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.x86.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.x86.checked.3\diff
Total bytes of delta: -1532 (-0.00 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 3 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.x86.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.x86.checked.3\diff
Total bytes of delta: -18554 (-0.02 % of base)
8 total files with Code Size differences (0 improved, 8 regressed), 15 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.arm.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.arm.checked.3\diff
Total bytes of delta: -2176 (-0.00 % of base)
1 total files with Code Size differences (0 improved, 1 regressed), 6 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.arm.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.arm.checked.3\diff
Total bytes of delta: -1016 (-0.00 % of base)
3 total files with Code Size differences (2 improved, 1 regressed), 2 unchanged.
Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.arm.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.arm.checked.3\diff
Total bytes of delta: -27136 (-0.03 % of base)
8 total files with Code Size differences (0 improved, 8 regressed), 12 unchanged.

@ghostghost added area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI community-contribution Indicates that the PR has been added by a community member labels Nov 5, 2021
@ghost

ghost commented Nov 5, 2021

Copy link
Copy Markdown

Tagging subscribers to this area: @JulieLeeMSFT
See info in area-owners.md if you want to be subscribed.

Issue Details

The code in fgDebugCheckFlags had some mistakes that were easy to fix, for harder errors I added a comment explaining them.

The diff is a regression but a small one.

The diffs look like correctness fixes, for example, for such comparison:

Core_Root\superpmi.exe "Core_Root_base\clrjit.dll" "Core_Root_diff\clrjit.dll" D:\Sergey\git\runtime\artifacts\spmi\mch\3df3e3ec-b1f6-4e6d-8439-2e7f3f7fa2ac.windows.x64\libraries_tests.pmi.windows.x64.checked.mch -a -c 78772

we forbid swapping call and IND(global static field), if I understand correctly, nothing prevents the call from making a change in this field, so it should not be hard to create a correctness repro for this bug.

***** BB01
++STMT00002 (IL ???... ???)++ [000013] -A--G------- * ASG byref ++ [000012] D------N---- +--* LCL_VAR byref V03 tmp2 ++ [000003] ----G------- \--* ADD byref ++ [000001] n---G------- +--* IND ref ++ [000000] I----------- | \--* CNS_INT(h) long 0xd1ffab1e static Fseq[hackishFieldName]++ [000002] ------------ \--* CNS_INT long 8 Fseq[#FirstElem]
***** BB01
STMT00001 (IL ???... ???)
[000011] I-C-G------- * CALL void System.TimeSpan..ctor (exactContextHnd=0x00000000D1FFAB1E)
[000010] ------------ this in rcx +--* ADDR byref [000009] -------N---- | \--* LCL_VAR struct<System.TimeSpan, 8> V02 tmp1 [000005] ------------ arg1 \--* CNS_INT long 1
***** BB01
STMT00002 (IL 0x00C... ???)
-- [000013] I-C-G------- * CALL struct System.TimeSpan.Add (exactContextHnd=0x00000000D1FFAB1E)-- [000003] ------------ this in rcx +--* ADD byref -- [000001] ------------ | +--* IND ref -- [000000] I----------- | | \--* CNS_INT(h) long 0xd1ffab1e static Fseq[hackishFieldName]-- [000002] ------------ | \--* CNS_INT long 8 Fseq[#FirstElem]-- [000015] n----------- arg1 \--* OBJ struct<System.TimeSpan, 8>-- [000014] ------------ \--* ADDR byref -- [000012] -------N---- \--* LCL_VAR struct<System.TimeSpan, 8> V02 tmp1 ++STMT00003 (IL 0x00C... ???)++ [000016] I-C-G------- * CALL struct System.TimeSpan.Add (exactContextHnd=0x00000000D1FFAB1E)++ [000014] ------------ this in rcx +--* LCL_VAR byref V03 tmp2 ++ [000018] n----------- arg1 \--* OBJ struct<System.TimeSpan, 8>++ [000017] ------------ \--* ADDR byref ++ [000015] -------N---- \--* LCL_VAR struct<System.TimeSpan, 8> V02 tmp1

so if inside [000011] I-C-G------- * CALL void System.TimeSpan..ctor we change what is stored in

[000001] n---G------- +--* IND ref [000000] I----------- | \--* CNS_INT(h) long 0xd1ffab1e static Fseq[hackishFieldName]

then before this change, we would read the changed value, not the correct one.

diffs with misleading total numbers Ignore the total diff, see #60124 ``` Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.x64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.x64.checked.4\diff Total bytes of delta: -6956 (-0.10 % of base) 1 total files with Code Size differences (0 improved, 1 regressed), 1 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.x64.checked.5\diff
Total bytes of delta: -1515 (-0.00 % of base)
5 total files with Code Size differences (1 improved, 4 regressed), 4 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.x64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.x64.checked.4\diff
Total bytes of delta: -3695 (-0.01 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 2 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.x64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.x64.checked.4\diff
Total bytes of delta: -21142 (-0.02 % of base)
19 total files with Code Size differences (2 improved, 17 regressed), 2 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.arm64.checked.4\diff
Total bytes of delta: -1332 (-0.00 % of base)
1 total files with Code Size differences (0 improved, 1 regressed), 3 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.arm64.checked.4\diff
Total bytes of delta: -2920 (-0.01 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 1 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.arm64.checked.4\diff
Total bytes of delta: -20104 (-0.02 % of base)
17 total files with Code Size differences (0 improved, 17 regressed), 4 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.Linux.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.Linux.x64.checked.5\diff
Total bytes of delta: -7003 (-0.08 % of base)
1 total files with Code Size differences (0 improved, 1 regressed), 0 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.x64.checked.5\diff
Total bytes of delta: -1295 (-0.00 % of base)
5 total files with Code Size differences (1 improved, 4 regressed), 4 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.x64.checked.5\diff
Total bytes of delta: -2909 (-0.01 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 1 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.x64.checked.5\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.x64.checked.5\diff
Total bytes of delta: -23240 (-0.02 % of base)
19 total files with Code Size differences (3 improved, 16 regressed), 2 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.arm64.checked.4\diff
Total bytes of delta: -6992 (-0.09 % of base)
0 total files with Code Size differences (0 improved, 0 regressed), 2 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.arm64.checked.4\diff
Total bytes of delta: -1332 (-0.00 % of base)
1 total files with Code Size differences (0 improved, 1 regressed), 3 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.arm64.checked.4\diff
Total bytes of delta: -3848 (-0.01 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 2 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.arm64.checked.4\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.arm64.checked.4\diff
Total bytes of delta: -20100 (-0.02 % of base)
17 total files with Code Size differences (0 improved, 17 regressed), 4 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.x86.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.benchmarks.run.windows.x86.checked.3\diff
Total bytes of delta: -4922 (-0.09 % of base)
0 total files with Code Size differences (0 improved, 0 regressed), 3 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.x86.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.windows.x86.checked.3\diff
Total bytes of delta: -2182 (-0.00 % of base)
7 total files with Code Size differences (0 improved, 7 regressed), 6 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.x86.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.windows.x86.checked.3\diff
Total bytes of delta: -1532 (-0.00 % of base)
3 total files with Code Size differences (0 improved, 3 regressed), 3 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.x86.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.windows.x86.checked.3\diff
Total bytes of delta: -18554 (-0.02 % of base)
8 total files with Code Size differences (0 improved, 8 regressed), 15 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.arm.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.coreclr_tests.pmi.Linux.arm.checked.3\diff
Total bytes of delta: -2176 (-0.00 % of base)
1 total files with Code Size differences (0 improved, 1 regressed), 6 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.arm.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries.pmi.Linux.arm.checked.3\diff
Total bytes of delta: -1016 (-0.00 % of base)
3 total files with Code Size differences (2 improved, 1 regressed), 2 unchanged.

Creating dasm files: D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.arm.checked.3\base D:\Sergey\git\runtime\artifacts\spmi\asm.libraries_tests.pmi.Linux.arm.checked.3\diff
Total bytes of delta: -27136 (-0.03 % of base)
8 total files with Code Size differences (0 improved, 8 regressed), 12 unchanged.


</details>
<table>
<tr>
<th align="left">Author:</th>
<td>sandreenko</td>
</tr>
<tr>
<th align="left">Assignees:</th>
<td>-</td>
</tr>
<tr>
<th align="left">Labels:</th>
<td>
`area-CodeGen-coreclr`
</td>
</tr>
<tr>
<th align="left">Milestone:</th>
<td>-</td>
</tr>
</table>
</details>

@sandreenko

sandreenko commented Nov 5, 2021

Copy link
Copy Markdown
ContributorAuthor

Many regressions are caused by the need to spill such reads before inlinings or tail calls, so @jakobbotsch could you please review it?

@SingleAccretion

SingleAccretion commented Nov 5, 2021•
edited
Loading

Copy link
Copy Markdown
Contributor

so if inside [000011] I-C-G------- * CALL void System.TimeSpan..ctor we change what is stored in

[000001] n---G------- +--* IND ref
[000000] I----------- | \--* CNS_INT(h) long 0xd1ffab1e static Fseq[hackishFieldName]

Is that a legal thing to do? Since this is a box in which the struct static is stored, it seems like UB if managed code tried to access it in some way.

(Edit: it would be semantically equivalent to mutating the static's address)

jakobbotsch
jakobbotsch reviewed Nov 5, 2021
View reviewed changes
Comment threadsrc/coreclr/jit/morph.cpp
// must have it set as well. However, we clear it from the parent in cases
// ADDR(LCL_VAR or CLS_VAR).
if (oper == GT_ADDR && (op1->OperIs(GT_LCL_VAR, GT_CLS_VAR) ||
(op1->OperIs(GT_IND) && op1->AsOp()->gtOp1->OperIs(GT_CLS_VAR_ADDR))))

@jakobbotschjakobbotschNov 5, 2021

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is OperIs(GT_CLS_VAR_ADDR) ever true here? AFAIK we only use those opers very late.

@sandreenkosandreenkoNov 5, 2021

Copy link
Copy Markdown
ContributorAuthor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

you are right, I was adding it to sync with fgDebugCheckFlags but the only moment when we need it after rationalizing (created first GT_CLS_VAR_ADDR) and before lowering (don't call fgDebugCheckStmtsList->fgDebugCheckFlags in LIR). I will delete this change and add a not to why we need it in fgDebugCheckFlags, thanks for catching this.

jakobbotsch
jakobbotsch reviewed Nov 5, 2021
View reviewed changes

@jakobbotschjakobbotsch left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not super familiar with how these flags are meant to be propagated, maybe someone more familiar from @dotnet/jit-contrib could have another look?

@sandreenko

sandreenko commented Nov 5, 2021

Copy link
Copy Markdown
ContributorAuthor

so if inside [000011] I-C-G------- * CALL void System.TimeSpan..ctor we change what is stored in

[000001] n---G------- +--* IND ref
[000000] I----------- | \--* CNS_INT(h) long 0xd1ffab1e static Fseq[hackishFieldName]

Is that a legal thing to do? Since this is a box in which the struct static is stored, it seems like UB if managed code tried to access it in some way.

(Edit: it would be semantically equivalent to mutating the static's address)

I am not sure, probably @AndyAyersMS could help here.
The flag that we use here is CORINFO_FLG_FIELD_STATIC_IN_HEAP = 0x00000008, // See code:#StaticFields. This static field is in the GC heap as a boxed object and I was under impression that we are accessing a value of a static field with such IL, do you mean that

 [000001] n---G------- +--* IND ref [000000] I----------- | \--* CNS_INT(h) long 0xd1ffab1e static Fseq[hackishFieldName]

gives us the address of the field and so it can't change from managed code? Would it be right then to mark it as an invariant (it still can change due to GC, cant it?) and drop global?

@SingleAccretion

SingleAccretion commented Nov 5, 2021•
edited
Loading

Copy link
Copy Markdown
Contributor

gives us the address of the field and so it can't change from managed code?

Right. I guess technically it gives us "the base", we will also add an offset to it to get to the actual data, and that will be the value ldsflda gives us:

privatestaticVector128<int>_vtor;[MethodImpl(MethodImplOptions.NoInlining)]privatestaticrefVector128<int>AddressOfStructStatic(){returnref_vtor;}
N005 ( 7, 15) [000004] ------------*RETURN byref N004 ( 6, 14) [000003] ------------\--*ADD byref N002 ( 4, 12) [000001] n-----------+--*IND ref N001 ( 2, 10) [000000] I-----------|\--*CNS_INT(h) long 0xd1ffab1e static Fseq[_vtor]
N003 ( 1, 1) [000002] ------------\--*CNS_INT long 8Fseq[#FirstElem]

I don't think there is a combination of IL instructions that will allow one to get to the underlying box and mutate the reference itself.

Would it be right then to mark it as an invariant (it still can change due to GC, cant it?) and drop global?

I tried recently to mark it as invariant, but fgDebugCheckFlags was upset with me :), so I decided to look into it after some of my other work is finished. It is definitely "invariant" in the sense that it cannot be stored to, whether the VM can update it from under us I do not know. I guess I will say that we already model it as invariant when we don't expand it into its full form but instead use CLS_VAR, on 32 bit (and sometimes on 64 bit too):

staticreadonlyVector128<float>_vtor=Vector128.Create(1.0f,2.0f,3.0f,4.0f);staticint_justSomeField;privatestaticVector128<float>Vtors(){vara=_vtor;_justSomeField=1;for(inti=0;i<10;i++){a=Sse.Add(a,_vtor);}returna;}

- x86 codegen for the loop:

G_M54979_IG03: vaddps xmm1,xmm1,xmm0inceaxcmpeax,10jl SHORT G_M54979_IG03

IR in VN:

*****BB02, STMT00004(before)
N008 ( 19, 16) [000025] -A-XG---R---*ASG simd16 (copy)
N007 ( 3, 2) [000023] D------N----+--*LCL_VAR simd16<System.Runtime.Intrinsics.Vector128`1[Single]> V01 loc0 d:5N006 ( 15, 13) [000022] ---XG-------\--*HWINTRINSIC simd16 float AddN001 ( 3, 2) [000017] ------------+--*LCL_VAR simd16<System.Runtime.Intrinsics.Vector128`1[Single]> V01 loc0 u:4 (last use)
N005 ( 11, 10) [000021] ---XG-------\--*OBJ simd16<System.Runtime.Intrinsics.Vector128`1[Single]>
N004 ( 5, 6) [000020] ----G-------\--*ADD byref N002 ( 3, 4) [000018] ----G-------+--*CLS_VAR ref Hnd=0xd1ffab1eFseq[_vtor]
N003 ( 1, 1) [000019] ------------\--*CNS_INT int 4Fseq[#FirstElem]
N001 [000017] LCL_VARV01 loc0 u:4 (last use) => $2c0 {PhiDef($1, $4, $1c3)}
FieldSeq {_vtor} is $180N002 [000018] CLS_VARHnd=0xd1ffab1eFseq[_vtor] => $81 {PtrToStatic($180)}
N003 [000019] CNS_INT4Fseq[#FirstElem] => $42 {IntCns4}
FieldSeq {#FirstElem} is $181
fieldSeq $182 is {(_vtor, #FirstElem)}
N004 [000020] ADD=> $82 {PtrToStatic($182)}
KnowntypeVector128<float>
VNApplySelectors:VNForHandle(_vtor) is $140, fieldType is simd16, size =16AX2: $140!= $142==> select([$240]store($100, $142, $43), $140) ==> select($100, $140) remaining budget is 99.
VNForMapSelect($240, $140):simd16 returns $1c0 {$100[$140]}
N005 [000021] OBJ=> <l:$1c2 {norm=$1c0 {$100[$140]}, exc=$183 {NullPtrExc($82)}}, c:$1c4 {norm=$202 {MemOpaque:L00}, exc=$183 {NullPtrExc($82)}}>
N006 [000022] HWINTRINSIC=> <l:$1c8 {norm=$1c6 {HWI_SSE_Add($2c0, $1c0)}, exc=$183 {NullPtrExc($82)}}, c:$1c7 {norm=$1c5 {HWI_SSE_Add($2c0, $202)}, exc=$183 {NullPtrExc($82)}}>
N007 [000023] LCL_VARV01 loc0 d:5=> <l:$1c6 {HWI_SSE_Add($2c0, $1c0)}, c:$1c5 {HWI_SSE_Add($2c0, $202)}>
N008 [000025] ASG=> <l:$1c8 {norm=$1c6 {HWI_SSE_Add($2c0, $1c0)}, exc=$183 {NullPtrExc($82)}}, c:$1c7 {norm=$1c5 {HWI_SSE_Add($2c0, $202)}, exc=$183 {NullPtrExc($82)}}>

(We label it as PtrToStatic in its CLS_VAR form, we don't do it in the full expanded form because we think it is an actual struct field, but the two forms are equivalent semantically)

@sandreenko

sandreenko commented Nov 5, 2021

Copy link
Copy Markdown
ContributorAuthor

Very interesting,

whether the VM can update it from under us I do not know.

the only scenario that I can imagine is GC.collect but then its VM responsibility to update the result of the IND.

I tried recently to mark it as invariant, but fgDebugCheckFlags was upset with me :)

we can try to combine these fixes for fgDebugCheckFlags with your work, could you share the diffs so I can try to play with them?

@AndyAyersMS

AndyAyersMS commented Nov 5, 2021

Copy link
Copy Markdown
Member

It would be nice to separate out refactorings/cleanups (like wider use of GenTreeFlags) from behavior changes. So maybe just do a refactor PR first?

As others have noted, access to some statics might require more than one indir; I believe all save the outermost should be invariant and non-faulting.

@sandreenko

sandreenko commented Nov 5, 2021

Copy link
Copy Markdown
ContributorAuthor

It would be nice to separate out refactorings/cleanups (like wider use of GenTreeFlags) from behavior changes. So maybe just do a refactor PR first?

sure, the refactorings came last moment when I was analyzing regressions and was surprised that the flags are not shown under the debugger, I will extract this change.

@SingleAccretion

SingleAccretion commented Nov 5, 2021

Copy link
Copy Markdown
Contributor

could you share the diffs so I can try to play with them?

I did not do much - just marked the address indirection as invariant. The problem was that the handle under it implies a mutable static, so that's why the checker was complaining. I then (after the replays came back red) solved the problem that prompted to me to look into this in the first place in a different way.

@sandreenkosandreenko mentioned this pull request Nov 6, 2021
Merged
@AndyAyersMS

AndyAyersMS commented Nov 11, 2021

Copy link
Copy Markdown
Member

@sandreenko is this ready for re-review?

@sandreenko
sandreenko marked this pull request as draft November 12, 2021 01:46
@sandreenko

sandreenko commented Nov 12, 2021

Copy link
Copy Markdown
ContributorAuthor

@sandreenko is this ready for re-review?

no, looks like we need to understand when to set GTF_IND_INVARIANT , when not to set GTF_GLOB_REF for these static fields in this PR otherwise it leads to regressions. And, looks like we don't have a good understanding of it now. I will play with it when I have time.

@ghostghost closed this Dec 12, 2021
@ghost

ghost commented Dec 12, 2021

Copy link
Copy Markdown

Draft Pull Request was automatically closed for inactivity. Please let us know if you'd like to reopen it.

@ghostghost locked as resolved and limited conversation to collaborators Jan 11, 2022
This pull request was closed.
Sign up for freeto subscribe to this conversation on GitHub. Already have an account? Sign in.

Reviewers

@jakobbotschjakobbotschjakobbotsch left review comments
@AndyAyersMSAndyAyersMSAwaiting requested review from AndyAyersMS

Assignees

@AndyAyersMSAndyAyersMS

@sandreenkosandreenko

Labels

area-CodeGen-coreclrCLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMIcommunity-contributionIndicates that the PR has been added by a community member

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@sandreenko@SingleAccretion@AndyAyersMS@jakobbotsch