JIT: contained memory safety analysis fixes and improvements by AndyAyersMS · Pull Request #64843 · dotnet/runtime

AndyAyersMS · 2022-02-05T03:20:14Z

Fixes a couple of issues exposed by forward sub, where containment analysis
was allowing unsafe reordering of operands. More checks of this sort may be
needed.

See issues in #64828.

Generalize the safety check so that a store to a local not live into a handler
can be reordered with respect to node causing exceptions. Happily this leads
to almost uniformly better code despite the more stringent checking added above.

Add special handling for "transparent" nodes like CreateScalarUnsafe that allow
their children to be contained by their parent.

Add a workaround for the late callbacks into the containment checker made on
unlinked nodes. Assume these are always safe.

Fixes a couple of issues exposed by forward sub, where containment analysis was allowing unsafe reordering of operands. More checks of this sort may be needed. See issues in dotnet#64828. Generalize the safety check so that a store to a local not live into a handler can be reordered with respect to node causing exceptions. Happily this leads to almost uniformly better code despite the more stringent checking added above. Add a workaround for the late callbacks into the containment checker made on unlinked nodes. Assume these are always safe.

ghost · 2022-02-05T03:20:22Z

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

Issue Details

Fixes a couple of issues exposed by forward sub, where containment analysis
was allowing unsafe reordering of operands. More checks of this sort may be
needed.

See issues in #64828.

Generalize the safety check so that a store to a local not live into a handler
can be reordered with respect to node causing exceptions. Happily this leads
to almost uniformly better code despite the more stringent checking added above.

Add a workaround for the late callbacks into the containment checker made on
unlinked nodes. Assume these are always safe.

Author:	AndyAyersMS
Assignees:	AndyAyersMS
Labels:	`area-CodeGen-coreclr`
Milestone:	-

AndyAyersMS · 2022-02-05T03:20:25Z

cc @dotnet/jit-contrib

tannergooding · 2022-02-05T04:27:34Z

 if (op1Type == op2Type)
 {
- if (IsContainableMemoryOp(op1))
+ if (IsContainableMemoryOp(op1) && IsSafeToContainMem(cmp, op1))


There looked to be a couple outer places in lowerxarch that do IsContainableMemoryOp followed by an MakeSrcContained without checking IsSafeToContainMem
https://github.com/dotnet/runtime/blob/main/src/coreclr/jit/lowerxarch.cpp#L4726-L4729
https://github.com/dotnet/runtime/blob/main/src/coreclr/jit/lowerxarch.cpp#L4856-L4859
https://github.com/dotnet/runtime/blob/main/src/coreclr/jit/lowerxarch.cpp#L5253-L5256
https://github.com/dotnet/runtime/blob/main/src/coreclr/jit/lowerxarch.cpp#L5282-L5286
https://github.com/dotnet/runtime/blob/main/src/coreclr/jit/lowerxarch.cpp#L6290-L6298
We also have https://github.com/dotnet/runtime/blob/main/src/coreclr/jit/lowerxarch.cpp#L5072-L5075, where its not clear if it doing IsContainableMemoryOp is right and perhaps indirOpSource->isContained() would be better?

Thanks, added checks.
I'll probably leave that last bit as is, IsContainableMemoryOp is not that expensive.

tannergooding · 2022-02-05T04:28:44Z

+ //
+ if (transparentParentNode != nullptr)
+ {
+ canBeContained = IsSafeToContainMem(containingNode, transparentParentNode, node);


We don't need this for the Load* intrinsic cases below because its known to always be safe, is that right?

I think so, yes. But not 100% sure.
By my understanding, for unary containing nodes, the safety check should immediately return safe, as the child's gtNext is the node, so there is nothing "in between" that can interfere. For higher arity operations the children can interfere and also there can be other interference in between them from COMMA expansion.

I'm going to add a quick early out to IsSafeToContainMem for the unary node case, so we are less tempted to optimize out the safety check call and possibly miss something.

By my understanding, for unary containing nodes, the safety check should immediately return safe, as the child's gtNext is the node, so there is nothing "in between" that can interfere
Do you mean that for unary operators operand->gtNext == parent? But that's not true for the linear order.
t1 =IND(addr) STOREIND(addr, ...) // Updates [addr] to t2UNARY_USER(t1) // Should observe t1, not t2.
Is valid LIR.

Do you know of any way we'd create such a pattern? It would be nice to have more examples that fail if we mess this up.
At any rate, best not to be too clever here. With the added safety calls and follow-up check in MakeSrcContained we should have things covered, I hope.

Do you know of any way we'd create such a pattern?
Yep:
privatestaticuintProblem(uinta,uint*b){uintzero=0;return*b+Bmi2.MultiplyNoFlags(a,a,b)*zero;}
N003 (???,???) [000015] ------------IL_OFFSET void INLRT@0x002[E-] REGNAN005 ( 1, 1) [000003] -----------Z t3 =LCL_VAR int V01 arg1 u:1 edx REG edx $81/--* t3 int N007 ( 3, 2) [000004] *--XG------- t4 =*IND int REG eax <l:$1c1, c:$1c0> N009 ( 1, 1) [000005] ------------ t5 =LCL_VAR int V00 arg0 u:1 ecx REG ecx $80N011 ( 1, 1) [000006] ------------ t6 =LCL_VAR int V00 arg0 u:1 ecx (last use) REG ecx $80N013 ( 1, 1) [000007] -----------z t7 =LCL_VAR int V01 arg1 u:1 esi (last use) REG esi $81/--* t5 int +--* t6 int +--* t7 int N015 ( 4, 4) [000008] ---XG------- t8 =*HWINTRINSIC int MultiplyNoFlagsREG edx $83/--* t4 int N017 ( 10, 9) [000012] ---XG-------*RETURN int REGNA $181

AndyAyersMS · 2022-02-05T16:51:08Z

Given that Lowering::MakeSrcContained takes the parent node as an arg it seems like we can add memory safety verification there and catch cases where the caller missed checks.

Not sure if this will be super-costly for checked builds, but it seems worth trying.

AndyAyersMS · 2022-02-05T17:30:34Z

Small number of diffs from the extra checks. Here's one: we think we should be null checking this before bounds checking and this leads to some messy codegen.

;; BEFORE; Assembly listing for method System.Runtime.Intrinsics.Vector128`1[Byte][System.Byte]:get_Item(int):ubyte:this; Emitting BLENDED_CODE for X64 CPU with AVX - Windows; optimized code; rsp based frame; partially interruptible; No matching PGO data; Final local variable assignments;; V00 this [V00,T01] ( 3, 3 ) byref -> rcx this single-def; V01 arg1 [V01,T00] ( 4, 4 ) int -> rdx single-def; V02 OutArgs [V02 ] ( 1, 1 ) lclBlk (32) [rsp+00H] "OutgoingArgSpace";; Lcl frame size = 40G_M59860_IG01: ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref, nogc <-- Prolog IGsubrsp,40vzeroupper ;; bbWeight=1 PerfScore 1.25G_M59860_IG02: ; gcrefRegs=00000000 {}, byrefRegs=00000002 {rcx}, byref, isz ; byrRegs +[rcx]cmpedx,16jae SHORT G_M59860_IG04movzxrax, byte ptr [rcx+rdx] ;; bbWeight=1 PerfScore 3.25G_M59860_IG03: ; , epilog, nogc, extendaddrsp,40ret ;; bbWeight=1 PerfScore 1.25G_M59860_IG04: ; gcVars=0000000000000000 {}, gcrefRegs=00000000 {}, byrefRegs=00000000 {}, gcvars, byref ; byrRegs -[rcx]call CORINFO_HELP_THROW_ARGUMENTOUTOFRANGEEXCEPTION ; gcr arg pop 0int3;; AFTER; Assembly listing for method System.Runtime.Intrinsics.Vector128`1[Byte][System.Byte]:get_Item(int):ubyte:this; Emitting BLENDED_CODE for X64 CPU with AVX - Windows; optimized code; rsp based frame; partially interruptible; No matching PGO data; Final local variable assignments;; V00 this [V00,T01] ( 3, 3 ) byref -> rcx this single-def; V01 arg1 [V01,T00] ( 4, 4 ) int -> rdx single-def; V02 OutArgs [V02 ] ( 1, 1 ) lclBlk (32) [rsp+00H] "OutgoingArgSpace"; V03 rat0 [V03 ] ( 1, 1 ) simd16 -> [rsp+28H] "SIMDInitTempVar";; Lcl frame size = 56G_M59860_IG01: ; gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref, nogc <-- Prolog IGsubrsp,56vzeroupper ;; bbWeight=1 PerfScore 1.25G_M59860_IG02: ; gcrefRegs=00000000 {}, byrefRegs=00000002 {rcx}, byref, isz ; byrRegs +[rcx]vmovupdxmm0, xmmword ptr [rcx]cmpedx,16jae SHORT G_M59860_IG04vmovupd xmmword ptr [rsp+28H],xmm0movzxrax, byte ptr [rsp+rdx+28H] ;; bbWeight=1 PerfScore 8.25G_M59860_IG03: ; , epilog, nogc, extendaddrsp,56ret ;; bbWeight=1 PerfScore 1.25G_M59860_IG04: ; gcVars=0000000000000000 {}, gcrefRegs=00000000 {}, byrefRegs=00000000 {}, gcvars, byref ; byrRegs -[rcx]call CORINFO_HELP_THROW_ARGUMENTOUTOFRANGEEXCEPTION ; gcr arg pop 0int3

Full set of regressions from the extra checks is

aspnet.run.windows.x64.checked.mch:


Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 19733334 (overridden on cmd)
Total bytes of diff: 19733350 (overridden on cmd)
Total bytes of delta: 16 (0.00 % of base)
diff is a regression.
relative diff is a regression.

Detail diffs


Top file regressions (bytes):
2 : 46659.dasm (1.49% of base)
2 : 50687.dasm (0.32% of base)
2 : 51092.dasm (1.49% of base)
2 : 51969.dasm (0.35% of base)
2 : 45076.dasm (1.49% of base)
2 : 55877.dasm (1.49% of base)
2 : 48677.dasm (0.37% of base)
2 : 44556.dasm (0.32% of base)
8 total files with Code Size differences (0 improved, 8 regressed), 0 unchanged.
Top method regressions (bytes):
2 ( 1.49% of base) : 46659.dasm - PathString:StartsWithSegments(PathString,int):bool:this
2 ( 1.49% of base) : 51092.dasm - PathString:StartsWithSegments(PathString,int):bool:this
2 ( 1.49% of base) : 45076.dasm - PathString:StartsWithSegments(PathString,int):bool:this
2 ( 1.49% of base) : 55877.dasm - PathString:StartsWithSegments(PathString,int):bool:this
2 ( 0.32% of base) : 50687.dasm - RuntimeType:FilterApplyMethodBase(MethodBase,int,int,int,ref):bool
2 ( 0.35% of base) : 51969.dasm - RuntimeType:FilterApplyMethodBase(MethodBase,int,int,int,ref):bool
2 ( 0.37% of base) : 48677.dasm - RuntimeType:FilterApplyMethodBase(MethodBase,int,int,int,ref):bool
2 ( 0.32% of base) : 44556.dasm - RuntimeType:FilterApplyMethodBase(MethodBase,int,int,int,ref):bool
Top method regressions (percentages):
2 ( 1.49% of base) : 46659.dasm - PathString:StartsWithSegments(PathString,int):bool:this
2 ( 1.49% of base) : 51092.dasm - PathString:StartsWithSegments(PathString,int):bool:this
2 ( 1.49% of base) : 45076.dasm - PathString:StartsWithSegments(PathString,int):bool:this
2 ( 1.49% of base) : 55877.dasm - PathString:StartsWithSegments(PathString,int):bool:this
2 ( 0.37% of base) : 48677.dasm - RuntimeType:FilterApplyMethodBase(MethodBase,int,int,int,ref):bool
2 ( 0.35% of base) : 51969.dasm - RuntimeType:FilterApplyMethodBase(MethodBase,int,int,int,ref):bool
2 ( 0.32% of base) : 50687.dasm - RuntimeType:FilterApplyMethodBase(MethodBase,int,int,int,ref):bool
2 ( 0.32% of base) : 44556.dasm - RuntimeType:FilterApplyMethodBase(MethodBase,int,int,int,ref):bool
8 total methods with Code Size differences (0 improved, 8 regressed), 0 unchanged.

coreclr_tests.pmi.windows.x64.checked.mch:


Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 128321453 (overridden on cmd)
Total bytes of diff: 128321565 (overridden on cmd)
Total bytes of delta: 112 (0.00 % of base)
diff is a regression.
relative diff is a regression.

Detail diffs


Top file regressions (bytes):
18 : 239319.dasm (3.01% of base)
18 : 239317.dasm (2.95% of base)
18 : 239318.dasm (2.85% of base)
11 : 84869.dasm (42.31% of base)
11 : 84868.dasm (39.29% of base)
11 : 84867.dasm (40.74% of base)
11 : 84896.dasm (40.74% of base)
11 : 84870.dasm (39.29% of base)
2 : 239321.dasm (0.32% of base)
2 : 89419.dasm (0.09% of base)
Top file improvements (bytes):
-1 : 239320.dasm (-0.16% of base)
11 total files with Code Size differences (1 improved, 10 regressed), 0 unchanged.
Top method regressions (bytes):
18 ( 2.95% of base) : 239317.dasm - VectorArrayTest`1[Byte][System.Byte]:VectorArray(ubyte):int
18 ( 2.85% of base) : 239318.dasm - VectorArrayTest`1[Int16][System.Int16]:VectorArray(short):int
18 ( 3.01% of base) : 239319.dasm - VectorArrayTest`1[Int32][System.Int32]:VectorArray(int):int
11 (40.74% of base) : 84867.dasm - System.Runtime.Intrinsics.Vector128`1[Byte][System.Byte]:get_Item(int):ubyte:this
11 (39.29% of base) : 84870.dasm - System.Runtime.Intrinsics.Vector128`1[Double][System.Double]:get_Item(int):double:this
11 (39.29% of base) : 84868.dasm - System.Runtime.Intrinsics.Vector128`1[Int16][System.Int16]:get_Item(int):short:this
11 (42.31% of base) : 84869.dasm - System.Runtime.Intrinsics.Vector128`1[Int32][System.Int32]:get_Item(int):int:this
11 (40.74% of base) : 84896.dasm - System.Runtime.Intrinsics.Vector128`1[Int64][System.Int64]:get_Item(int):long:this
2 ( 0.09% of base) : 89419.dasm - ILGEN_0x537f7b0:Method_0x323f83b5(double,byte,int,short,ubyte,double,short,ushort,double,long,int,long,ushort,long,int):float
2 ( 0.32% of base) : 239321.dasm - VectorArrayTest`1[Int64][System.Int64]:VectorArray(long):int
Top method improvements (bytes):
-1 (-0.16% of base) : 239320.dasm - VectorArrayTest`1[Double][System.Double]:VectorArray(double):int
Top method regressions (percentages):
11 (42.31% of base) : 84869.dasm - System.Runtime.Intrinsics.Vector128`1[Int32][System.Int32]:get_Item(int):int:this
11 (40.74% of base) : 84867.dasm - System.Runtime.Intrinsics.Vector128`1[Byte][System.Byte]:get_Item(int):ubyte:this
11 (40.74% of base) : 84896.dasm - System.Runtime.Intrinsics.Vector128`1[Int64][System.Int64]:get_Item(int):long:this
11 (39.29% of base) : 84870.dasm - System.Runtime.Intrinsics.Vector128`1[Double][System.Double]:get_Item(int):double:this
11 (39.29% of base) : 84868.dasm - System.Runtime.Intrinsics.Vector128`1[Int16][System.Int16]:get_Item(int):short:this
18 ( 3.01% of base) : 239319.dasm - VectorArrayTest`1[Int32][System.Int32]:VectorArray(int):int
18 ( 2.95% of base) : 239317.dasm - VectorArrayTest`1[Byte][System.Byte]:VectorArray(ubyte):int
18 ( 2.85% of base) : 239318.dasm - VectorArrayTest`1[Int16][System.Int16]:VectorArray(short):int
2 ( 0.32% of base) : 239321.dasm - VectorArrayTest`1[Int64][System.Int64]:VectorArray(long):int
2 ( 0.09% of base) : 89419.dasm - ILGEN_0x537f7b0:Method_0x323f83b5(double,byte,int,short,ubyte,double,short,ushort,double,long,int,long,ushort,long,int):float
Top method improvements (percentages):
-1 (-0.16% of base) : 239320.dasm - VectorArrayTest`1[Double][System.Double]:VectorArray(double):int
11 total methods with Code Size differences (1 improved, 10 regressed), 0 unchanged.

libraries.pmi.windows.x64.checked.mch:


Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 46583443 (overridden on cmd)
Total bytes of diff: 46583502 (overridden on cmd)
Total bytes of delta: 59 (0.00 % of base)
diff is a regression.
relative diff is a regression.

Detail diffs


Top file regressions (bytes):
11 : 20673.dasm (40.74% of base)
11 : 20675.dasm (42.31% of base)
11 : 20674.dasm (39.29% of base)
11 : 20676.dasm (39.29% of base)
11 : 20701.dasm (40.74% of base)
2 : 58129.dasm (0.41% of base)
2 : 58130.dasm (0.71% of base)
7 total files with Code Size differences (0 improved, 7 regressed), 0 unchanged.
Top method regressions (bytes):
11 (40.74% of base) : 20673.dasm - System.Runtime.Intrinsics.Vector128`1[Byte][System.Byte]:get_Item(int):ubyte:this
11 (39.29% of base) : 20676.dasm - System.Runtime.Intrinsics.Vector128`1[Double][System.Double]:get_Item(int):double:this
11 (39.29% of base) : 20674.dasm - System.Runtime.Intrinsics.Vector128`1[Int16][System.Int16]:get_Item(int):short:this
11 (42.31% of base) : 20675.dasm - System.Runtime.Intrinsics.Vector128`1[Int32][System.Int32]:get_Item(int):int:this
11 (40.74% of base) : 20701.dasm - System.Runtime.Intrinsics.Vector128`1[Int64][System.Int64]:get_Item(int):long:this
2 ( 0.71% of base) : 58130.dasm - Microsoft.CodeAnalysis.VisualBasic.LocalRewriter:AppendToBlock(Microsoft.CodeAnalysis.VisualBasic.BoundBlock,Microsoft.CodeAnalysis.VisualBasic.BoundStatement):Microsoft.CodeAnalysis.VisualBasic.BoundBlock:this
2 ( 0.41% of base) : 58129.dasm - Microsoft.CodeAnalysis.VisualBasic.LocalRewriter:Concat(Microsoft.CodeAnalysis.VisualBasic.BoundStatement,Microsoft.CodeAnalysis.VisualBasic.BoundStatement):Microsoft.CodeAnalysis.VisualBasic.BoundStatement:this
Top method regressions (percentages):
11 (42.31% of base) : 20675.dasm - System.Runtime.Intrinsics.Vector128`1[Int32][System.Int32]:get_Item(int):int:this
11 (40.74% of base) : 20673.dasm - System.Runtime.Intrinsics.Vector128`1[Byte][System.Byte]:get_Item(int):ubyte:this
11 (40.74% of base) : 20701.dasm - System.Runtime.Intrinsics.Vector128`1[Int64][System.Int64]:get_Item(int):long:this
11 (39.29% of base) : 20676.dasm - System.Runtime.Intrinsics.Vector128`1[Double][System.Double]:get_Item(int):double:this
11 (39.29% of base) : 20674.dasm - System.Runtime.Intrinsics.Vector128`1[Int16][System.Int16]:get_Item(int):short:this
2 ( 0.71% of base) : 58130.dasm - Microsoft.CodeAnalysis.VisualBasic.LocalRewriter:AppendToBlock(Microsoft.CodeAnalysis.VisualBasic.BoundBlock,Microsoft.CodeAnalysis.VisualBasic.BoundStatement):Microsoft.CodeAnalysis.VisualBasic.BoundBlock:this
2 ( 0.41% of base) : 58129.dasm - Microsoft.CodeAnalysis.VisualBasic.LocalRewriter:Concat(Microsoft.CodeAnalysis.VisualBasic.BoundStatement,Microsoft.CodeAnalysis.VisualBasic.BoundStatement):Microsoft.CodeAnalysis.VisualBasic.BoundStatement:this
7 total methods with Code Size differences (0 improved, 7 regressed), 0 unchanged.

libraries_tests.pmi.windows.x64.checked.mch:


Summary of Code Size diffs:
(Lower is better)
Total bytes of base: 119636782 (overridden on cmd)
Total bytes of diff: 119637034 (overridden on cmd)
Total bytes of delta: 252 (0.00 % of base)
diff is a regression.
relative diff is a regression.

Detail diffs


Top file regressions (bytes):
33 : 305492.dasm (6.53% of base)
27 : 305489.dasm (5.76% of base)
27 : 305491.dasm (5.81% of base)
27 : 305494.dasm (5.72% of base)
27 : 305504.dasm (5.57% of base)
27 : 305490.dasm (5.73% of base)
21 : 305503.dasm (4.67% of base)
21 : 305502.dasm (4.61% of base)
21 : 305501.dasm (4.64% of base)
21 : 305506.dasm (4.59% of base)
10 total files with Code Size differences (0 improved, 10 regressed), 0 unchanged.
Top method regressions (bytes):
33 ( 6.53% of base) : 305492.dasm - <>c__DisplayClass309_0`1[Double][System.Double]:<TestAdditionOverflow>b__0(int,double):this
27 ( 5.76% of base) : 305489.dasm - <>c__DisplayClass309_0`1[Byte][System.Byte]:<TestAdditionOverflow>b__0(int,ubyte):this
27 ( 5.73% of base) : 305490.dasm - <>c__DisplayClass309_0`1[Int16][System.Int16]:<TestAdditionOverflow>b__0(int,short):this
27 ( 5.81% of base) : 305491.dasm - <>c__DisplayClass309_0`1[Int32][System.Int32]:<TestAdditionOverflow>b__0(int,int):this
27 ( 5.72% of base) : 305494.dasm - <>c__DisplayClass309_0`1[Int64][System.Int64]:<TestAdditionOverflow>b__0(int,long):this
27 ( 5.57% of base) : 305504.dasm - <>c__DisplayClass329_0`1[Double][System.Double]:<TestSubtractionOverflow>b__0(int,double):this
21 ( 4.64% of base) : 305501.dasm - <>c__DisplayClass329_0`1[Byte][System.Byte]:<TestSubtractionOverflow>b__0(int,ubyte):this
21 ( 4.61% of base) : 305502.dasm - <>c__DisplayClass329_0`1[Int16][System.Int16]:<TestSubtractionOverflow>b__0(int,short):this
21 ( 4.67% of base) : 305503.dasm - <>c__DisplayClass329_0`1[Int32][System.Int32]:<TestSubtractionOverflow>b__0(int,int):this
21 ( 4.59% of base) : 305506.dasm - <>c__DisplayClass329_0`1[Int64][System.Int64]:<TestSubtractionOverflow>b__0(int,long):this
Top method regressions (percentages):
33 ( 6.53% of base) : 305492.dasm - <>c__DisplayClass309_0`1[Double][System.Double]:<TestAdditionOverflow>b__0(int,double):this
27 ( 5.81% of base) : 305491.dasm - <>c__DisplayClass309_0`1[Int32][System.Int32]:<TestAdditionOverflow>b__0(int,int):this
27 ( 5.76% of base) : 305489.dasm - <>c__DisplayClass309_0`1[Byte][System.Byte]:<TestAdditionOverflow>b__0(int,ubyte):this
27 ( 5.73% of base) : 305490.dasm - <>c__DisplayClass309_0`1[Int16][System.Int16]:<TestAdditionOverflow>b__0(int,short):this
27 ( 5.72% of base) : 305494.dasm - <>c__DisplayClass309_0`1[Int64][System.Int64]:<TestAdditionOverflow>b__0(int,long):this
27 ( 5.57% of base) : 305504.dasm - <>c__DisplayClass329_0`1[Double][System.Double]:<TestSubtractionOverflow>b__0(int,double):this
21 ( 4.67% of base) : 305503.dasm - <>c__DisplayClass329_0`1[Int32][System.Int32]:<TestSubtractionOverflow>b__0(int,int):this
21 ( 4.64% of base) : 305501.dasm - <>c__DisplayClass329_0`1[Byte][System.Byte]:<TestSubtractionOverflow>b__0(int,ubyte):this
21 ( 4.61% of base) : 305502.dasm - <>c__DisplayClass329_0`1[Int16][System.Int16]:<TestSubtractionOverflow>b__0(int,short):this
21 ( 4.59% of base) : 305506.dasm - <>c__DisplayClass329_0`1[Int64][System.Int64]:<TestSubtractionOverflow>b__0(int,long):this
10 total methods with Code Size differences (0 improved, 10 regressed), 0 unchanged.

Going to push this next batch of changes and will come back to addressing regressions when I have time.

AndyAyersMS · 2022-02-05T17:50:09Z

Seems plausible that for value type instance methods we could allow exceptions from an indir based on this to reorder with other exceptions in containment safety checks (at least for small offsets?).

AndyAyersMS · 2022-02-05T22:23:52Z

So at least per innerloop we're not missing any safety checks.

The checks still seem a bit heavy handed in places, like the regression noted above. Not sure it's worth trying to relax constraints for these isolated calls though; typically such methods will inline and the problem will likely go away.

The cases I've seen so far where we might want to relax:

two indir trees that can only cause NREs -- possibly allow them to reorder. This would fix the small regressions in the Test_Sse_CompareScalarOrderedLessThan_* tests, which are passed two byref arguments.
a byref this indir tree vs a bounds check for SIMD -- get_Item as noted above

Am going to look at some of the other regressions in libraries_tests next.

Turns out <>c__DisplayClass309_0`1[Double][System.Double]:<TestAdditionOverflow>b__0(int,double):this and friends are methods with multiple get_Item calls, so same case as above.

Another thought on how to deal with the SIMD case -- split the INDIR into an upfront nullcheck and a containable non-faulting INDIR. That should avoid the need to "spill" the SIMD value like we're seeing. I have been tempted to do similar things for inlining args that could be forward subbed but may cause exceptions (but no other side effects) -- just evaluate them twice, the first time just for effect. I may give this a try since it seems to be specific to just this one intrinsic.

Or (even more far out there perhaps) for this particular case we could do the bounds check first, then when we branch on failure, we could do the null check before calling the bounds check helper.

AndyAyersMS · 2022-02-05T22:36:54Z

Also fixing an outerloop test I broke in #64828. Will next try running outerloop.

AndyAyersMS · 2022-02-05T22:57:23Z

/azp run runtime-coreclr outerloop

AndyAyersMS · 2022-02-05T23:10:34Z

Hmm, not clear where to try and "fix" the get_Item issues.

Perhaps better to live with the small regressions for now and close the correctness loophole.

@dotnet/jit-contrib PTAL

azure-pipelines · 2022-02-05T23:20:55Z

Azure Pipelines successfully started running 1 pipeline(s).

jakobbotsch

LGTM. It would be nice to get this in before the fuzzers run tomorrow.

jakobbotsch · 2022-02-05T23:29:58Z

+//------------------------------------------------------------------------
+// IsSafeToContainMem: Checks for conflicts between childNode and grandParentNode
+// and returns 'true' iff memory operand childNode can be contained in ancestorNode
+//
+// Arguments:
+// grandParentNode - any non-leaf node
+// parentNode - parent of `childNode` and an input to `grandParentNode`
+// childNode - some node that is an input to `parentNode`
+//
+// Return value:
+// true if it is safe to make childNode a contained memory operand.
+//


Seems like there's a few argument names here that need to be updated, ancestorNode and grandParentNode => grandparentNode?

Thanks -- will fix this subsequently so I don't retrigger all the tests.

AndyAyersMS · 2022-02-06T02:00:27Z

Libraries test failure (Linux x64 Debug):

Process terminated. Assertion failed.
at System.Number.FormatHalf(ValueStringBuilder& sb, Half value, ReadOnlySpan`1 format, NumberFormatInfo info)

I have seen this before (#61359).

AndyAyersMS · 2022-02-06T02:28:42Z

x86 crossgen2 failure on windows, may be related?

 Fatal error. 0xC0000005
at Internal.JitInterface.CorInfoImpl.JitCompileMethod(IntPtr ByRef, IntPtr, IntPtr, IntPtr, Internal.JitInterface.CORINFO_METHOD_INFO ByRef, UInt32, IntPtr ByRef, UInt32 ByRef)
at Internal.JitInterface.CorInfoImpl.CompileMethodInternal(ILCompiler.DependencyAnalysis.IMethodNode, Internal.IL.MethodIL)
at Internal.JitInterface.CorInfoImpl.CompileMethod(ILCompiler.DependencyAnalysis.ReadyToRun.MethodWithGCInfo, ILCompiler.Logger)
at ILCompiler.ReadyToRunCodegenCompilation.<ComputeDependencyNodeDependencies>b__35_0(ILCompiler.DependencyAnalysisFramework.DependencyNodeCore`1<ILCompiler.DependencyAnalysis.NodeFactory>)
at System.Threading.Tasks.Parallel+<>c__DisplayClass33_0`2[[System.__Canon, System.Private.CoreLib, Version=6.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e],[System.__Canon, System.Private.CoreLib, Version=6.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e]].<ForEachWorker>b__0(Int32)
at System.Threading.Tasks.Parallel+<>c__DisplayClass19_0`1[[System.__Canon, System.Private.CoreLib, Version=6.0.0.0, Culture=neutral, PublicKeyToken=7cec85d7bea7798e]].<ForWorker>b__1(System.Threading.Tasks.RangeWorker ByRef, Int32, Boolean ByRef)
at System.Threading.Tasks.TaskReplicator+Replica.Execute()
at System.Threading.Tasks.TaskReplicator+Replica+<>c.<.ctor>b__4_0(System.Object)
at System.Threading.Tasks.Task.InnerInvoke()
at System.Threading.Tasks.Task+<>c.<.cctor>b__271_0(System.Object)
at System.Threading.ExecutionContext.RunFromThreadPoolDispatchLoop(System.Threading.Thread, System.Threading.ExecutionContext, System.Threading.ContextCallback, System.Object)
at System.Threading.Tasks.Task.ExecuteWithThreadLocal(System.Threading.Tasks.Task ByRef, System.Threading.Thread)
at System.Threading.Tasks.Task.ExecuteEntryUnsafe(System.Threading.Thread)
at System.Threading.ThreadPoolWorkQueue.Dispatch()
at System.Threading.PortableThreadPool+WorkerThread.WorkerThreadStart()
at System.Threading.Thread.StartCallback()

Can't repro it locally. Looking at the 1GB "mini dump" from CI, we are OOMing while in the jit:

00 3007ddb0 728baa33 coreclr!EEPolicy::HandleFatalError+0x72
01 3007df30 727d1f0e coreclr!CPFH_RealFirstPassHandler+0xe8b0c
02 (Inline) -------- coreclr!CPFH_FirstPassHandler+0xc6
03 3007df74 774d3272 coreclr!COMPlusFrameHandler+0x13e
04 3007df98 774d3244 ntdll!ExecuteHandler2+0x26
05 3007e060 774c0dbf ntdll!ExecuteHandler+0x24
06 3007e538 6f7c149b ntdll!KiUserExceptionDispatcher+0xf
07 3007e538 6f7c7168 clrjit_win_x86_x86!ArenaAllocator::allocateNewPage+0x55
08 3007e554 6f7c71ac clrjit_win_x86_x86!ArenaAllocator::allocateMemory+0x65
09 3007e564 6f87594f clrjit_win_x86_x86!ArenaAllocator::MemStatsAllocator::allocateMemory+0x20
0a 3007e574 6f877ebc clrjit_win_x86_x86!CompAllocator::allocate<HashTableBase<int,jitstd::list<GenTree *,jitstd::allocator<GenTree *> > *,HashTableInfo<int>,CompAllocator>::Bucket>+0x19
0b 3007e598 6f875d99 clrjit_win_x86_x86!HashTableBase<int,jitstd::list<GenTree *,jitstd::allocator<GenTree *> > *,HashTableInfo<int>,CompAllocator>::Resize+0x25
0c 3007e5b4 6f87600d clrjit_win_x86_x86!HashTableBase<int,jitstd::list<GenTree *,jitstd::allocator<GenTree *> > *,HashTableInfo<int>,CompAllocator>::AddOrUpdate+0x4d
0d 3007e61c 6f8766bd clrjit_win_x86_x86!CheckLclVarSemanticsHelper::Check+0xed
0e 3007eb50 6f811304 clrjit_win_x86_x86!LIR::Range::CheckLIR+0x50a
0f 3007eb88 6f8825fd clrjit_win_x86_x86!Compiler::fgUpdateFlowGraph+0x7da
10 3007ebac 6f8c3b38 clrjit_win_x86_x86!Lowering::DoPhase+0xed
11 3007ebe0 6f7e642d clrjit_win_x86_x86!Phase::Run+0x38

so this is unrelated.

We might want to handle OOMing this way a bit more gracefully. Not sure if a minopts fallback is sensible but it may be an option to recover.

AndyAyersMS · 2022-02-06T06:40:18Z

I think all failures are unrelated.

Several timeouts and the two issues above.

ghost assigned AndyAyersMS Feb 5, 2022

ghost added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Feb 5, 2022

tannergooding reviewed Feb 5, 2022
View reviewed changes

extra checks; fast path early out; assertion in MakeSrcContained
bd67329

fix broken outerloop test
0026f9c

jakobbotsch approved these changes Feb 5, 2022
View reviewed changes

AndyAyersMS merged commit d936a66 into dotnet:mainFeb 6, 2022

AndyAyersMS mentioned this pull request Feb 7, 2022
System.Numerics.Vectors.Tests: Assertion failed 'intrinsicId == NI_Vector128_GetElement' #64918
Closed

ghost locked as resolved and limited conversation to collaborators Mar 8, 2022

Uh oh!

Conversation

AndyAyersMS commented Feb 5, 2022• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ghost commented Feb 5, 2022

Uh oh!

AndyAyersMS commented Feb 5, 2022

Uh oh!

tannergoodingFeb 5, 2022• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AndyAyersMSFeb 5, 2022• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AndyAyersMS commented Feb 5, 2022

Uh oh!

AndyAyersMS commented Feb 5, 2022

aspnet.run.windows.x64.checked.mch:

coreclr_tests.pmi.windows.x64.checked.mch:

libraries.pmi.windows.x64.checked.mch:

libraries_tests.pmi.windows.x64.checked.mch:

Uh oh!

AndyAyersMS commented Feb 5, 2022

Uh oh!

AndyAyersMS commented Feb 5, 2022• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

AndyAyersMS commented Feb 5, 2022

Uh oh!

AndyAyersMS commented Feb 5, 2022

Uh oh!

AndyAyersMS commented Feb 5, 2022

Uh oh!

azure-pipelinesBot commented Feb 5, 2022

Uh oh!

jakobbotsch left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AndyAyersMS commented Feb 6, 2022

Uh oh!

AndyAyersMS commented Feb 6, 2022• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

AndyAyersMS commented Feb 6, 2022• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

AndyAyersMS commented Feb 5, 2022•
edited
Loading

tannergoodingFeb 5, 2022•
edited
Loading

AndyAyersMSFeb 5, 2022•
edited
Loading

AndyAyersMS commented Feb 5, 2022•
edited
Loading

AndyAyersMS commented Feb 6, 2022•
edited
Loading

AndyAyersMS commented Feb 6, 2022•
edited
Loading