JIT: Ensure last-use copy omission candidates are marked with GTF_GLOB_REF #89088
Conversation
Last-use copy omission allows the JIT to avoid creating copies of struct values that are passed implicitly by-ref when the value is a last-use of a local. When we do this we effectively address expose the local for the lifetime of the call, so we mark the local as address exposed as part of doing so. However, there is a problem where morph may reorder two uses of a such a local and break the last use information. For example, consider the following case: ``` [000015] --CXG------ ▌ CALL void Program:Foo(int,int) [000010] ----------- arg0 ├──▌ LCL_FLD int V00 loc0 [+0] [000012] --CXG------ arg1 └──▌ CALL int Program:Bar(S):int [000011] ----------- arg0 └──▌ LCL_VAR struct<S, 32> V00 loc0 (last use) ``` V00 is used both at [000010] and at [000011], the latter as a last use. Since we only address expose V00 when we get to [000011] we do not mark [000010] with GTF_GLOB_REF; the net effect is the following call args morphing, where we have reordered the two occurrences illegally: ``` [000015] --CXG+----- ▌ CALL void Program:Foo(int,int) [000037] DACXG------ arg1 setup ├──▌ STORE_LCL_VAR int V04 tmp3 [000012] --CXG+----- │ └──▌ CALL int Program:Bar(S):int [000011] -----+----- arg0 in rcx │ └──▌ LCL_ADDR long V00 loc0 [+0] [000038] ----------- arg1 in rdx ├──▌ LCL_VAR int V04 tmp3 [000010] -----+----- arg0 in rcx └──▌ LCL_FLD int (AX) V00 loc0 [+0] ``` This change fixes the problem by running a separate pass over the IR before morph to identify and address expose the candidates for last-use copy omission. The net result is then the following correct IR: ``` [000015] --CXG+----- ▌ CALL void Runtime_85611:Foo(int,int) [000039] DA--G------ arg0 setup ├──▌ STORE_LCL_VAR int V05 tmp4 [000010] ----G+----- │ └──▌ LCL_FLD int (AX) V00 loc0 [+0] [000037] DACXG------ arg1 setup ├──▌ STORE_LCL_VAR int V04 tmp3 [000012] --CXG+----- │ └──▌ CALL int Runtime_85611:Bar(Runtime_85611+S):int [000011] ----G+----- arg0 in rcx │ └──▌ LCL_ADDR long V00 loc0 [+0] [000038] ----------- arg1 in rdx ├──▌ LCL_VAR int V04 tmp3 [000040] ----------- arg0 in rcx └──▌ LCL_VAR int V05 tmp4 ``` The pass has some TP impact but it is mitigated since we only need to visit statements with GTF_CALL. Fix dotnet#85611
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsLast-use copy omission allows the JIT to avoid creating copies of struct values that are passed implicitly by-ref when the value is a last-use of a local. When we do this we effectively address expose the local for the lifetime of the call, so we mark the local as address exposed as part of doing so. However, there is a problem where morph may reorder two uses of a such a local and break the last use information. For example, consider the following case: [000015] --CXG------ ▌ CALL void Program:Foo(int,int)
[000010] ----------- arg0 ├──▌ LCL_FLD int V00 loc0 [+0]
[000012] --CXG------ arg1 └──▌ CALL int Program:Bar(S):int
[000011] ----------- arg0 └──▌ LCL_VAR struct<S, 32> V00 loc0 (last use)V00 is used both at [000015] --CXG+----- ▌ CALL void Program:Foo(int,int)
[000037] DACXG------ arg1 setup ├──▌ STORE_LCL_VAR int V04 tmp3
[000012] --CXG+----- │ └──▌ CALL int Program:Bar(S):int
[000011] -----+----- arg0 in rcx │ └──▌ LCL_ADDR long V00 loc0 [+0]
[000038] ----------- arg1 in rdx ├──▌ LCL_VAR int V04 tmp3
[000010] -----+----- arg0 in rcx └──▌ LCL_FLD int (AX) V00 loc0 [+0]This change fixes the problem by running a separate pass over the IR before morph to identify the candidates for last-use copy omission. These candidates then have [000015] --CXG+----- ▌ CALL void Runtime_85611:Foo(int,int)
[000039] DA--G------ arg0 setup ├──▌ STORE_LCL_VAR int V05 tmp4
[000010] ----G+----- │ └──▌ LCL_FLD int (AX) V00 loc0 [+0]
[000037] DACXG------ arg1 setup ├──▌ STORE_LCL_VAR int V04 tmp3
[000012] --CXG+----- │ └──▌ CALL int Runtime_85611:Bar(Runtime_85611+S):int
[000011] ----G+----- arg0 in rcx │ └──▌ LCL_ADDR long V00 loc0 [+0]
[000038] ----------- arg1 in rdx ├──▌ LCL_VAR int V04 tmp3
[000040] ----------- arg0 in rcx └──▌ LCL_VAR int V05 tmp4I first started out with a fix that simply address exposed the candidates in the pass, but this has some semi-large regressions (diffs) since it disables lots of optimizations that morph will do. Fix #85611 Tiny number of diffs are expected with positive TP diffs in minopts are expected due to the removal of
|
|
/azp run runtime |
|
No commit pushedDate could be found for PR 89088 in repo dotnet/runtime |
|
/azp run runtime |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
cc @dotnet/jit-contrib PTAL @AndyAyersMS The failure is #88507. Small number of diffs, with a small TP impact (sometimes positive, sometimes negative). |
ghost
locked as resolved and limited conversation to collaborators
Last-use copy omission allows the JIT to avoid creating copies of struct values that are passed implicitly by-ref when the value is a last-use of a local. When we do this we effectively address expose the local for the lifetime of the call, so we mark the local as address exposed as part of doing so.
However, there is a problem where morph may reorder two uses of a such a local and break the last use information. For example, consider the following case:
V00 is used both at
[000010]and at[000011], the latter as a last use. Since we only address expose V00 when we get to[000011]we do not mark[000010]withGTF_GLOB_REF; the net effect is the following call args morphing, where we have reordered the two occurrences illegally:This change fixes the problem by running a separate pass over the IR before morph to identify the candidates for last-use copy omission. These candidates then have
GTF_GLOB_REFadded as part of morph. The net result is then the following correct IR:I first started out with a fix that simply address exposed the candidates in the pass, but this has some semi-large regressions (diffs) since it disables lots of optimizations that morph will do.
It's still not ideal since address exposing the local late just has the effect of making these optimizations depend on the block visit order; for .NET 9 I want to consider a fix where we instead do this optimization in the backend.
Fix #85611
Tiny number of diffs are expected with positive TP diffs in minopts due to the removal of
fgRemarkGlobalUses. The TP impact of the pass is minor due to the use ofGTF_CALLand the locals linked list to exit very early (about 0.03%, less than what is gained by the removal offgRemarkGlobalUses).