JIT: "Mismatch in align instruction" in superpmi-replay

[jakobbotsch]

superpmi-replay seems to be failing with "Mismatch in align instruction" when run with JitStressRegs=3:

[00:26:08] Mismatch in align instruction.

[00:26:08] Containing IG: IG117

[00:26:08] loopHeadPredIG: IG117

[00:26:08] loopHeadIG: IG118

[00:26:08] igInLoop: IG121

[00:26:08] igInLoop->backEdge: IG119

[00:26:08] igInLoop has align instruction for : IG122

[00:26:08] Loop:

[00:26:08] 	IG118

[00:26:08] 	IG119

[00:26:08] 	IG120

[00:26:08] 	IG121

[00:26:08] 	IG122

[00:26:08] 	IG123

[00:26:08] 	IG124

[00:26:08] 	IG125

[00:26:08] 	IG126

[00:26:08] 	IG127

[00:26:08] 	IG128

[00:26:08] 	IG129

[00:26:08] 	IG130

[00:26:08] 	IG131

[00:26:08] 	IG132

[00:26:08] 	IG133

[00:26:08] 	IG134

[00:26:08] 	IG135

[00:26:08] 	IG136

[00:26:08] 	IG137

[00:26:08] 	IG138

[00:26:08] 	IG139

[00:26:08] 	IG140

[00:26:08] 	IG141

[00:26:08] 	IG142

[00:26:08] 	IG143

[00:26:08] 	IG144

[00:26:08] 	IG145

[00:26:08] 	IG146

[00:26:08] 	IG147

[00:26:08] 	IG148

[00:26:08] 	IG149

[00:26:08] 	IG150

[00:26:08] 	IG151

[00:26:08] 	IG152

[00:26:08] 	IG153

[00:26:08] 	IG154

[00:26:08] 	IG155

[00:26:08] 	IG156

[00:26:08] 	IG157

[00:26:08] 	IG158

[00:26:08] 	IG159

[00:26:08] 	IG160

[00:26:08] 	IG161

[00:26:08] 	IG162

[00:26:08] 	IG163

[00:26:08] 	IG164

[00:26:08] 	IG165

[00:26:08] 	IG166

[00:26:08] 	IG167

[00:26:08] 	IG168

[00:26:08] 	IG169

[00:26:08] 	IG170

[00:26:08] 	IG171

[00:26:08] 	IG172

[00:26:08] 	IG173

[00:26:08] 	IG174

[00:26:08] 	IG175

[00:26:08] 	IG176

[00:26:08] 	IG177

[00:26:08] 	IG178

[00:26:08] 	IG179

[00:26:08] 	IG180

[00:26:08] 	IG181

[00:26:08] 	IG182

[00:26:08] 	IG183

[00:26:08] 	IG184

[00:26:08] 	IG185

[00:26:08] 	IG186

[00:26:08] 	IG187

[00:26:08] 	IG188

[00:26:08] 	IG189

[00:26:08] 	IG190

[00:26:08] 	IG191

[00:26:08] 	IG192

[00:26:08] 	IG193

[00:26:08] 	IG194

[00:26:08] 	IG195

[00:26:08] 	IG196

[00:26:08] 	IG197

[00:26:08] 	IG198

[00:26:08] 	IG199

[00:26:08] 	IG200

[00:26:08] 	IG201

[00:26:08] 	IG202

[00:26:08] 	IG203

[00:26:08] 	IG204

[00:26:08] 	IG205

[00:26:08] 	IG206

[00:26:08] 	IG207

[00:26:08] 	IG208

[00:26:08] 	IG209

[00:26:08] 	IG210

[00:26:08] 	IG211

[00:26:08] 	IG212

[00:26:08] 	IG213

[00:26:08] 	IG214

[00:26:08] 	IG215

[00:26:08] 	IG216

[00:26:08] ISSUE: <ASSERT> #530064 D:\a\_work\1\s\src\coreclr\jit\emit.cpp (5883) - Assertion failed 'false' in 'System.Numerics.BigIntegerCalculator:PowCore(System.Span`1[uint],int,System.ReadOnlySpan`1[uint],System.ReadOnlySpan`1[uint],System.Span`1[uint],System.Span`1[uint])' during 'Generate code' (IL size 511; hash 0xd0657e4f; Tier1)

[00:26:08] 

Console log: https://helixre107v0xdcypoyl9e7f.blob.core.windows.net/dotnet-runtime-refs-pull-91827-merge-6e8d1e4418334afa89/windows-x64-2/1/console.c2bbff16.log?helixlogtype=result

cc @kunalspathak

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

Issue Details

---

superpmi-replay seems to be failing with "Mismatch in align instruction" when run with JitStressRegs=3:

[00:26:08] Mismatch in align instruction.

[00:26:08] Containing IG: IG117

[00:26:08] loopHeadPredIG: IG117

[00:26:08] loopHeadIG: IG118

[00:26:08] igInLoop: IG121

[00:26:08] igInLoop->backEdge: IG119

[00:26:08] igInLoop has align instruction for : IG122

[00:26:08] Loop:

[00:26:08] 	IG118

[00:26:08] 	IG119

[00:26:08] 	IG120

[00:26:08] 	IG121

[00:26:08] 	IG122

[00:26:08] 	IG123

[00:26:08] 	IG124

[00:26:08] 	IG125

[00:26:08] 	IG126

[00:26:08] 	IG127

[00:26:08] 	IG128

[00:26:08] 	IG129

[00:26:08] 	IG130

[00:26:08] 	IG131

[00:26:08] 	IG132

[00:26:08] 	IG133

[00:26:08] 	IG134

[00:26:08] 	IG135

[00:26:08] 	IG136

[00:26:08] 	IG137

[00:26:08] 	IG138

[00:26:08] 	IG139

[00:26:08] 	IG140

[00:26:08] 	IG141

[00:26:08] 	IG142

[00:26:08] 	IG143

[00:26:08] 	IG144

[00:26:08] 	IG145

[00:26:08] 	IG146

[00:26:08] 	IG147

[00:26:08] 	IG148

[00:26:08] 	IG149

[00:26:08] 	IG150

[00:26:08] 	IG151

[00:26:08] 	IG152

[00:26:08] 	IG153

[00:26:08] 	IG154

[00:26:08] 	IG155

[00:26:08] 	IG156

[00:26:08] 	IG157

[00:26:08] 	IG158

[00:26:08] 	IG159

[00:26:08] 	IG160

[00:26:08] 	IG161

[00:26:08] 	IG162

[00:26:08] 	IG163

[00:26:08] 	IG164

[00:26:08] 	IG165

[00:26:08] 	IG166

[00:26:08] 	IG167

[00:26:08] 	IG168

[00:26:08] 	IG169

[00:26:08] 	IG170

[00:26:08] 	IG171

[00:26:08] 	IG172

[00:26:08] 	IG173

[00:26:08] 	IG174

[00:26:08] 	IG175

[00:26:08] 	IG176

[00:26:08] 	IG177

[00:26:08] 	IG178

[00:26:08] 	IG179

[00:26:08] 	IG180

[00:26:08] 	IG181

[00:26:08] 	IG182

[00:26:08] 	IG183

[00:26:08] 	IG184

[00:26:08] 	IG185

[00:26:08] 	IG186

[00:26:08] 	IG187

[00:26:08] 	IG188

[00:26:08] 	IG189

[00:26:08] 	IG190

[00:26:08] 	IG191

[00:26:08] 	IG192

[00:26:08] 	IG193

[00:26:08] 	IG194

[00:26:08] 	IG195

[00:26:08] 	IG196

[00:26:08] 	IG197

[00:26:08] 	IG198

[00:26:08] 	IG199

[00:26:08] 	IG200

[00:26:08] 	IG201

[00:26:08] 	IG202

[00:26:08] 	IG203

[00:26:08] 	IG204

[00:26:08] 	IG205

[00:26:08] 	IG206

[00:26:08] 	IG207

[00:26:08] 	IG208

[00:26:08] 	IG209

[00:26:08] 	IG210

[00:26:08] 	IG211

[00:26:08] 	IG212

[00:26:08] 	IG213

[00:26:08] 	IG214

[00:26:08] 	IG215

[00:26:08] 	IG216

[00:26:08] ISSUE: <ASSERT> #530064 D:\a\_work\1\s\src\coreclr\jit\emit.cpp (5883) - Assertion failed 'false' in 'System.Numerics.BigIntegerCalculator:PowCore(System.Span`1[uint],int,System.ReadOnlySpan`1[uint],System.ReadOnlySpan`1[uint],System.Span`1[uint],System.Span`1[uint])' during 'Generate code' (IL size 511; hash 0xd0657e4f; Tier1)

[00:26:08] 

Console log: https://helixre107v0xdcypoyl9e7f.blob.core.windows.net/dotnet-runtime-refs-pull-91827-merge-6e8d1e4418334afa89/windows-x64-2/1/console.c2bbff16.log?helixlogtype=result

cc @kunalspathak

Author:	jakobbotsch
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged
Milestone:	-

[kunalspathak]

Seems to be same as #90693.

[kunalspathak]

In placeLoopAlignInstructions(), we try to make loops unalign if we find out that the body of the loop starts before the loopTop. However, we only check if the body block is just before the loopTop block. In the failure case, there are entire set of blocks inserted by various optimizations. Below, block BB43 is a loop top whose other blocks appear before it i.e. B44, B45. As such we should unmark BB43 as not alignable.

BB44  [0174]  2       BB41,BB43            57.72  1039  4 [015..016)-> BB46  ( cond )                     i bwd IBC LIR 
BB45  [0175]  1       BB44                 57.72  1039  4 [015..016)-> BB43  ( cond )                     i bwd bwd-src IBC LIR 
BB46  [0176]  2       BB44,BB45            42.67   768  1 [015..016)-> BB275 ( cond )                     i bwd IBC LIR 
BB47  [0180]  1       BB46                 42.67   768  1 [015..016)-> BB54  ( cond )                     i hascall gcsafe bwd IBC LIR 
BB48  [0184]  1       BB47                 42.67   768  1 [015..016)-> BB275 ( cond )                     i hascall gcsafe bwd IBC LIR 
BB49  [0189]  1       BB48                 42.67   768  1 [015..016)                                      i hascall gcsafe bwd LoopPH IBC LIR 
BB51  [0196]  2       BB49,BB50            57.72  1039  5 [015..016)-> BB53  ( cond )                     i bwd IBC LIR 
BB52  [0197]  1       BB51                 57.72  1039  5 [015..016)-> BB50  ( cond )                     i bwd bwd-src IBC LIR 
BB53  [0198]  2       BB51,BB52            42.67   768  1 [015..016)                                      i bwd IBC LIR 
BB55  [0186]  2       BB53,BB54            42.67   768  1 [015..016)-> BB07  ( cond )                     i bwd IBC LIR 
BB56  [0053]  1       BB55                  1.33    24  0 [015..016)-> BB04  ( cond )                     i bwd IBC LIR 
BB57  [0055]  2       BB03,BB56             0.33     6    [015..016)-> BB110 ( cond )                     i IBC LIR 
BB58  [0806]  2       BB57,BB109           10      180    [015..016)-> BB87  ( cond )                     i bwd IBC LIR 
BB60  [0202]  1       BB58                  5.33    96  6 [015..016)-> BB275 ( cond )                     i bwd IBC LIR 
BB61  [0211]  1       BB60                  5.33    96  6 [015..016)-> BB65  ( cond )                     i hascall gcsafe bwd IBC LIR 
BB62  [0217]  1       BB61                  5.27    95  6 [015..016)-> BB275 ( cond )                     i gcsafe bwd IBC LIR 
BB63  [0223]  1       BB62                  5.27    95  6 [015..016)-> BB275 ( cond )                     i hascall gcsafe bwd IBC LIR 
BB64  [0230]  1       BB63                  5.27    95  6 [015..016)-> BB68  (always)                     i hascall gcsafe bwd IBC LIR 
BB43  [0173]  1       BB45                 15.05   271  4 [015..016)-> BB44  (always)                     i Loop Loop0 bwd bwd-target IBC LIR align 

[kunalspathak]

BEFORE:

BB40  [0162]  1       BB38                  1.48    27  1 [015..016)                                      i hascall gcsafe bwd IBC 
BB41  [0163]  3       BB37,BB39,BB40       42.67   768  1 [015..016)-> BB44  (always)                     i gcsafe bwd LoopPH IBC 
BB43  [0173]  1       BB45                 15.05   271  4 [015..016)                                      i Loop Loop0 bwd bwd-target IBC align 
BB44  [0174]  2       BB41,BB43            57.72  1039  4 [015..016)-> BB46  ( cond )                     i bwd IBC 
BB45  [0175]  1       BB44                 57.72  1039  4 [015..016)-> BB43  ( cond )                     i bwd bwd-src IBC 
BB46  [0176]  2       BB44,BB45            42.67   768  1 [015..016)-> BB275 ( cond )                     i bwd IBC 
BB47  [0180]  1       BB46                 42.67   768  1 [015..016)-> BB54  ( cond )                     i hascall gcsafe bwd IBC 
BB48  [0184]  1       BB47                 42.67   768  1 [015..016)-> BB275 ( cond )                     i hascall gcsafe bwd IBC 
BB49  [0189]  1       BB48                 42.67   768  1 [015..016)-> BB51  (always)                     i hascall gcsafe bwd LoopPH IBC 

Decided to straighten unconditional branch at block BB41 branch to BB44 because of IBC profile data
fgFindInsertPoint(regionIndex=0, putInTryRegion=true, startBlk=BB01, endBlk=BB00, nearBlk=BB44, jumpBlk=BB00, runRarely=false)
Relocated uncommon block BB43
Relocated block [BB43..BB43] inserted after BB49
Changed an unconditional jump from BB41 to the next block BB44 into a BBJ_NONE block
Block BB43 ended with a BBJ_NONE, Changed to an unconditional jump to BB44

AFTER:

BB40  [0162]  1       BB38                  1.48    27  1 [015..016)                                      i hascall gcsafe bwd IBC 
BB41  [0163]  3       BB37,BB39,BB40       42.67   768  1 [015..016)                                      i gcsafe bwd LoopPH IBC 
BB44  [0174]  2       BB41,BB43            57.72  1039  4 [015..016)-> BB46  ( cond )                     i bwd IBC 
BB45  [0175]  1       BB44                 57.72  1039  4 [015..016)-> BB43  ( cond )                     i bwd bwd-src IBC 
BB46  [0176]  2       BB44,BB45            42.67   768  1 [015..016)-> BB275 ( cond )                     i bwd IBC 
BB47  [0180]  1       BB46                 42.67   768  1 [015..016)-> BB54  ( cond )                     i hascall gcsafe bwd IBC 
BB48  [0184]  1       BB47                 42.67   768  1 [015..016)-> BB275 ( cond )                     i hascall gcsafe bwd IBC 
BB49  [0189]  1       BB48                 42.67   768  1 [015..016)-> BB51  (always)                     i hascall gcsafe bwd LoopPH IBC 
BB43  [0173]  1       BB45                 15.05   271  4 [015..016)-> BB44  (always)                     i Loop Loop0 bwd bwd-target IBC align 

This is the precise transformation that block reordering performs, where it moves the loop heap below the loop body. I looked in the code and it is tricky to find all the cases where this can happen so that we can unmark the alignment bits.

Perhaps, such scenarios should be handled in placeLoopAlignment() phase. In that phase, when we iterate through all the blocks, if we see a new bbNatLoopNum that is not already seen, we add it in a hashmap. Ideally, that should be the loop head block, and if that loop needs alignment, should be the one marked as such. However, if we encounter a block that is marked for alignment, but whose bbNatLoopNum is already in the hashmap, we will know that the loop body was encountered earlier than loop top and hence that loop should be unmarked for alignment.