pre-commit: PR115945

[dtcxzyw]

Link: llvm/llvm-project#115945
Requested by: @dtcxzyw

[github-actions]

runner: buildkitsandbox
baseline runner: cseadmin-PowerEdge-R830
baseline: llvm/llvm-project@59da1af
patch: llvm/llvm-project#115945
sha256: d328b8f83e48993aca460c5c10ebebc4c96c4db2c11e3ecbc2146c28bcf7fa8b
commit: 03a463d
84 files changed, 23240 insertions(+), 23133 deletions(-)

Compilation time result (by files):
Top 5 improvements:
  faiss/sorting.cpp.ll 2737312975 -> 2682824211 -1.99%
  libquic/p224-64.c.ll 10121950763 -> 10094333211 -0.27%
  openmpi/ad_nfs_read.ll 314846615 -> 314296061 -0.17%
  spike/f32_rem.ll 78047962 -> 77943078 -0.13%
  recastnavigation/RecastRasterization.cpp.ll 238372457 -> 238074990 -0.12%
Top 5 regressions:
  darktable/SamsungV0Decompressor.cpp.ll 805391542 -> 900985346 +11.87%
  proj/mlfn.cpp.ll 74339013 -> 76726187 +3.21%
  clamav/consio.cpp.ll 63020601 -> 63999468 +1.55%
  typst-rs/4w2wosufr3zrj3ba.ll 522528365 -> 529413302 +1.32%
  abc/ltl_parser.c.ll 287902968 -> 291225915 +1.15%

Overall: -0.00773396%

Compilation time result (by projects):
Top 5 improvements:
Top 5 regressions:
  nanosvg 3622521963 -> 3627561774 +0.14%
  darktable 134232439007 -> 134375971761 +0.11%

Overall: -0.00229482%

Improvements:
  indvars.NumElimIdentity 1769 -> 1798 +1.64%
  licm.NumAddSubHoisted 896 -> 898 +0.22%
  indvars.NumFoldedUser 1843 -> 1846 +0.16%
  indvars.NumReplaced 99165 -> 99307 +0.14%
  scalar-evolution.NumExitCountsComputed 3663461 -> 3664325 +0.02%
  indvars.NumElimCmp 46860 -> 46869 +0.02%
Regressions:
  div-rem-pairs.NumRecomposed 2830 -> 2822 -0.28%
  licm.NumSunk 249155 -> 248644 -0.21%
  correlated-value-propagation.NumUDivURemsNarrowedExpanded 887 -> 886 -0.11%
  correlated-value-propagation.NumAShrsConverted 3806 -> 3802 -0.11%
  correlated-value-propagation.NumSRems 1199 -> 1198 -0.08%
  simple-loop-unswitch.NumTrivial 2565 -> 2563 -0.08%
  gvn.NumGVNEqProp 338497 -> 338328 -0.05%
  instcombine.NumAggregateReconstructionsSimplified 9229 -> 9225 -0.04%
  scalar-evolution.NumExitCountsNotComputed 11524283 -> 11520504 -0.03%
  instcombine.NegatorNumTreesNegated 35982 -> 35971 -0.03%

velox/Expr.cpp.ll crash
03a463d pre-commit: Update
18 18 bench/abseil-cpp/optimized/hash_test.cc.ll
19 20 bench/cmake/optimized/archive_write_set_format_pax.c.ll
48 44 bench/delta-rs/optimized/2braxl0lj34anf5z.ll
2 2 bench/faiss/optimized/IndexFlat.cpp.ll
207 207 bench/graphviz/optimized/stress.c.ll
43 43 bench/gromacs/optimized/lmmin.cpp.ll
1 1 bench/harfbuzz/optimized/harfbuzz.cc.ll
11 11 bench/hyperscan/optimized/rose_build_add_mask.cpp.ll
101 85 bench/icu/optimized/ustrcase.ll
73 78 bench/imgui/optimized/imgui_tables.cpp.ll
181 182 bench/meshlab/optimized/filter_create.cpp.ll
71 72 bench/openusd/optimized/tessellation.cpp.ll
83 82 bench/php/optimized/phpdbg_cmd.ll
43 45 bench/proj/optimized/mlfn.cpp.ll
6 24 bench/qemu/optimized/hw_usb_hcd-xhci.c.ll
186 187 bench/recastnavigation/optimized/DetourDebugDraw.cpp.ll
14 14 bench/redis/optimized/sentinel.ll
4 4 bench/rust-analyzer-rs/optimized/egul20e4uygvok2.ll
6 6 bench/tree-sitter-rs/optimized/55e7aotywrgrb7st.ll
108 108 bench/typst-rs/optimized/4m3ebbqd1xx21e5m.ll
3 3 bench/wireshark/optimized/packet-gvcp.c.ll

[dtcxzyw]

Regression.

[akshayrdeodhar]

I'm unable to reproduce this- with and without the change, the output of opt -O3 ustrcase.ll is the same. Should I be using a different pass pipeline?

[akshayrdeodhar]

Also synced my fork to friday (to a commit hash after llvm-opt-benchmark was run), there's no source diff between a version with the phi simplification change and without.

[dtcxzyw]

Did you use the optimized IR? You should use the original IR files under bench/<project>/original/.

With llvm/llvm-project@52361d0:

bin/opt -O3 ../../llvm-opt-benchmark/bench/icu/original/ustrcase.ll -S -o out1.ll
<apply patch and rebuild opt>
bin/opt -O3 ../../llvm-opt-benchmark/bench/icu/original/ustrcase.ll -S -o out2.ll
diff out1.ll out2.ll

992c992
<   br i1 %cmp6149, label %while.body, label %for.end
---
>   br i1 %cmp6149, label %while.body.preheader, label %for.end
994,1000c994,1004
< while.body:                                       ; preds = %while.cond.outer, %while.cond.backedge
<   %srcIndex.1150 = phi i32 [ %srcIndex.1.be, %while.cond.backedge ], [ %srcIndex.1.ph, %while.cond.outer ]
<   %idxprom = sext i32 %srcIndex.1150 to i64
<   %arrayidx = getelementptr inbounds i16, ptr %src, i64 %idxprom
<   %1 = load i16, ptr %arrayidx, align 2
<   %conv = zext i16 %1 to i32
<   %cmp7 = icmp ult i16 %1, 383
---
> while.body.preheader:                             ; preds = %while.cond.outer
>   %1 = sext i32 %srcIndex.1.ph to i64
>   br label %while.body
> 
> while.body:                                       ; preds = %while.body.preheader, %while.cond.backedge
>   %indvars.iv = phi i64 [ %1, %while.body.preheader ], [ %indvars.iv.next.pre-phi, %while.cond.backedge ]
>   %indvars = trunc i64 %indvars.iv to i32
>   %arrayidx = getelementptr inbounds i16, ptr %src, i64 %indvars.iv
>   %2 = load i16, ptr %arrayidx, align 2
>   %conv = zext i16 %2 to i32
>   %cmp7 = icmp ult i16 %2, 383
1004c1008
<   %idxprom9 = zext nneg i16 %1 to i64
---
>   %idxprom9 = zext nneg i16 %2 to i64
1006,1007c1010,1011
<   %2 = load i8, ptr %arrayidx10, align 1
<   %cmp12 = icmp eq i8 %2, -128
---
>   %3 = load i8, ptr %arrayidx10, align 1
>   %cmp12 = icmp eq i8 %3, -128
1011,1012c1015,1017
<   %inc = add nsw i32 %srcIndex.1150, 1
<   %cmp16 = icmp eq i8 %2, 0
---
>   %4 = add nsw i64 %indvars.iv, 1
>   %cmp16 = icmp eq i8 %3, 0
>   %5 = trunc nsw i64 %4 to i32
1016c1021,1022
<   %srcIndex.1.be = phi i32 [ %inc, %if.end14 ], [ %inc39, %if.end38 ]
---
>   %indvars.iv.next.pre-phi = phi i64 [ %4, %if.end14 ], [ %10, %if.end38 ]
>   %srcIndex.1.be = phi i32 [ %5, %if.end14 ], [ %inc39, %if.end38 ]
1021c1027
<   %3 = sext i8 %2 to i16
---
>   %6 = sext i8 %3 to i16
1025c1031
<   %cmp22 = icmp ugt i16 %1, -10241
---
>   %cmp22 = icmp ugt i16 %2, -10241
1029c1035
<   %4 = load ptr, ptr %call, align 8
---
>   %7 = load ptr, ptr %call, align 8
1032,1034c1038,1040
<   %arrayidx28 = getelementptr inbounds i16, ptr %4, i64 %idxprom27
<   %5 = load i16, ptr %arrayidx28, align 2
<   %conv29 = zext i16 %5 to i32
---
>   %arrayidx28 = getelementptr inbounds i16, ptr %7, i64 %idxprom27
>   %8 = load i16, ptr %arrayidx28, align 2
>   %conv29 = zext i16 %8 to i32
1039,1041c1045,1047
<   %arrayidx34 = getelementptr inbounds i16, ptr %4, i64 %idxprom33
<   %6 = load i16, ptr %arrayidx34, align 2
<   %conv35 = zext i16 %6 to i32
---
>   %arrayidx34 = getelementptr inbounds i16, ptr %7, i64 %idxprom33
>   %9 = load i16, ptr %arrayidx34, align 2
>   %conv35 = zext i16 %9 to i32
1047c1053,1054
<   %inc39 = add nsw i32 %srcIndex.1150, 1
---
>   %10 = add nsw i64 %indvars.iv, 1
>   %inc39 = add nsw i32 %indvars, 1
1050c1057
<   %cmp46 = icmp ult i16 %6, 128
---
>   %cmp46 = icmp ult i16 %9, 128
1055c1062,1063
<   %7 = ashr i16 %6, 7
---
>   %11 = trunc nsw i64 %10 to i32
>   %12 = ashr i16 %9, 7
1059,1061c1067,1069
<   %srcIndex.2 = phi i32 [ %inc, %if.end18 ], [ %inc39, %if.end50.loopexit ]
<   %delta.0 = phi i16 [ %3, %if.end18 ], [ %7, %if.end50.loopexit ]
<   %add54 = add i16 %delta.0, %1
---
>   %srcIndex.2 = phi i32 [ %5, %if.end18 ], [ %11, %if.end50.loopexit ]
>   %delta.0 = phi i16 [ %6, %if.end18 ], [ %12, %if.end50.loopexit ]
>   %add54 = add i16 %delta.0, %2
1064,1065c1072,1073
<   %8 = xor i32 %prev.1.ph.ph, -1
<   %sub56 = add i32 %srcIndex.2, %8
---
>   %13 = xor i32 %prev.1.ph.ph, -1
>   %sub56 = add i32 %srcIndex.2, %13
1136c1144
<   %inc71 = add nsw i32 %srcIndex.1150, 1
---
>   %inc71 = add nsw i32 %indvars, 1
1146,1147c1154,1155
<   %9 = load i16, ptr %arrayidx78, align 2
<   %conv79 = zext i16 %9 to i32
---
>   %14 = load i16, ptr %arrayidx78, align 2
>   %conv79 = zext i16 %14 to i32
1156c1164
<   %inc88 = add nsw i32 %srcIndex.1150, 2
---
>   %inc88 = add nsw i32 %indvars, 2
1165c1173
<   store i32 %srcIndex.1150, ptr %cpStart94, align 4
---
>   store i32 %indvars, ptr %cpStart94, align 4
1182c1190
<   %sub103 = sub nsw i32 %srcIndex.1150, %prev.1.ph.ph
---
>   %sub103 = sub nsw i32 %indvars, %prev.1.ph.ph
1218,1220c1226,1228
<   %10 = load ptr, ptr %s, align 8
<   %sub107 = sub nsw i32 %srcIndex.3, %srcIndex.1150
<   %call108 = call fastcc noundef i32 @_ZN6icu_7512_GLOBAL__N_112appendResultEPDsiiiPKDsijPNS_5EditsE(ptr noundef %dest, i32 noundef %retval.0.i94, i32 noundef %destCapacity, i32 noundef %c.1, ptr noundef %10, i32 noundef %sub107, i32 noundef %options, ptr noundef %edits)
---
>   %15 = load ptr, ptr %s, align 8
>   %sub107 = sub nsw i32 %srcIndex.3, %indvars
>   %call108 = call fastcc noundef i32 @_ZN6icu_7512_GLOBAL__N_112appendResultEPDsiiiPKDsijPNS_5EditsE(ptr noundef %dest, i32 noundef %retval.0.i94, i32 noundef %destCapacity, i32 noundef %c.1, ptr noundef %15, i32 noundef %sub107, i32 noundef %options, ptr noundef %edits)
1225c1233
<   %srcIndex.1.lcssa = phi i32 [ %srcIndex.1.be, %while.cond.backedge ], [ %srcIndex.1.ph, %while.cond.outer ]
---
>   %srcIndex.1.lcssa = phi i32 [ %srcLimit, %while.cond.backedge ], [ %srcIndex.1.ph, %while.cond.outer ]
1962c1970
<   br i1 %cmp1139.i, label %while.body.i, label %for.end.i
---
>   br i1 %cmp1139.i, label %while.body.preheader.i, label %for.end.i
1964,1970c1972,1982
< while.body.i:                                     ; preds = %while.cond.outer.i, %while.cond.backedge.i
<   %srcIndex.1140.i = phi i32 [ %srcIndex.1.be.i, %while.cond.backedge.i ], [ %srcIndex.1.ph.i, %while.cond.outer.i ]
<   %idxprom.i = sext i32 %srcIndex.1140.i to i64
<   %arrayidx.i = getelementptr inbounds i16, ptr %src, i64 %idxprom.i
<   %2 = load i16, ptr %arrayidx.i, align 2
<   %conv.i = zext i16 %2 to i32
<   %cmp2.i = icmp ult i16 %2, 383
---
> while.body.preheader.i:                           ; preds = %while.cond.outer.i
>   %2 = sext i32 %srcIndex.1.ph.i to i64
>   br label %while.body.i
> 
> while.body.i:                                     ; preds = %while.cond.backedge.i, %while.body.preheader.i
>   %indvars.iv.i = phi i64 [ %2, %while.body.preheader.i ], [ %indvars.iv.next.pre-phi.i, %while.cond.backedge.i ]
>   %indvars.i = trunc i64 %indvars.iv.i to i32
>   %arrayidx.i = getelementptr inbounds i16, ptr %src, i64 %indvars.iv.i
>   %3 = load i16, ptr %arrayidx.i, align 2
>   %conv.i = zext i16 %3 to i32
>   %cmp2.i = icmp ult i16 %3, 383
1974c1986
<   %idxprom4.i = zext nneg i16 %2 to i64
---
>   %idxprom4.i = zext nneg i16 %3 to i64
1976,1977c1988,1989
<   %3 = load i8, ptr %arrayidx5.i, align 1
<   %cmp7.i = icmp eq i8 %3, -128
---
>   %4 = load i8, ptr %arrayidx5.i, align 1
>   %cmp7.i = icmp eq i8 %4, -128
1981,1982c1993,1995
<   %inc.i = add nsw i32 %srcIndex.1140.i, 1
<   %cmp11.i = icmp eq i8 %3, 0
---
>   %5 = add nsw i64 %indvars.iv.i, 1
>   %cmp11.i = icmp eq i8 %4, 0
>   %6 = trunc nsw i64 %5 to i32
1986c1999,2000
<   %srcIndex.1.be.i = phi i32 [ %inc.i, %if.end9.i ], [ %inc33.i, %if.end32.i ]
---
>   %indvars.iv.next.pre-phi.i = phi i64 [ %5, %if.end9.i ], [ %11, %if.end32.i ]
>   %srcIndex.1.be.i = phi i32 [ %6, %if.end9.i ], [ %inc33.i, %if.end32.i ]
1991c2005
<   %4 = sext i8 %3 to i16
---
>   %7 = sext i8 %4 to i16
1995c2009
<   %cmp17.i = icmp ugt i16 %2, -10241
---
>   %cmp17.i = icmp ugt i16 %3, -10241
1999c2013
<   %5 = load ptr, ptr %call.i, align 8
---
>   %8 = load ptr, ptr %call.i, align 8
2002,2004c2016,2018
<   %arrayidx23.i = getelementptr inbounds i16, ptr %5, i64 %idxprom22.i
<   %6 = load i16, ptr %arrayidx23.i, align 2
<   %conv24.i = zext i16 %6 to i32
---
>   %arrayidx23.i = getelementptr inbounds i16, ptr %8, i64 %idxprom22.i
>   %9 = load i16, ptr %arrayidx23.i, align 2
>   %conv24.i = zext i16 %9 to i32
2009,2011c2023,2025
<   %arrayidx28.i = getelementptr inbounds i16, ptr %5, i64 %idxprom27.i
<   %7 = load i16, ptr %arrayidx28.i, align 2
<   %conv29.i = zext i16 %7 to i32
---
>   %arrayidx28.i = getelementptr inbounds i16, ptr %8, i64 %idxprom27.i
>   %10 = load i16, ptr %arrayidx28.i, align 2
>   %conv29.i = zext i16 %10 to i32
2017c2031,2032
<   %inc33.i = add nsw i32 %srcIndex.1140.i, 1
---
>   %11 = add nsw i64 %indvars.iv.i, 1
>   %inc33.i = add nsw i32 %indvars.i, 1
2020c2035
<   %cmp39.i = icmp ult i16 %7, 128
---
>   %cmp39.i = icmp ult i16 %10, 128
2025c2040,2041
<   %8 = ashr i16 %7, 7
---
>   %12 = trunc nsw i64 %11 to i32
>   %13 = ashr i16 %10, 7
2029,2031c2045,2047
<   %srcIndex.2.i = phi i32 [ %inc.i, %if.end13.i ], [ %inc33.i, %if.end43.loopexit.i ]
<   %delta.0.i = phi i16 [ %4, %if.end13.i ], [ %8, %if.end43.loopexit.i ]
<   %add47.i = add i16 %delta.0.i, %2
---
>   %srcIndex.2.i = phi i32 [ %6, %if.end13.i ], [ %12, %if.end43.loopexit.i ]
>   %delta.0.i = phi i16 [ %7, %if.end13.i ], [ %13, %if.end43.loopexit.i ]
>   %add47.i = add i16 %delta.0.i, %3
2034,2035c2050,2051
<   %9 = xor i32 %prev.1.ph.i.ph, -1
<   %sub49.i = add i32 %srcIndex.2.i, %9
---
>   %14 = xor i32 %prev.1.ph.i.ph, -1
>   %sub49.i = add i32 %srcIndex.2.i, %14
2106,2107c2122,2123
<   %inc64.i = add nsw i32 %srcIndex.1140.i, 1
<   store i32 %srcIndex.1140.i, ptr %cpStart65.i, align 4
---
>   %inc64.i = add nsw i32 %indvars.i, 1
>   store i32 %indvars.i, ptr %cpStart65.i, align 4
2117,2118c2133,2134
<   %10 = load i16, ptr %arrayidx72.i, align 2
<   %conv73.i = zext i16 %10 to i32
---
>   %15 = load i16, ptr %arrayidx72.i, align 2
>   %conv73.i = zext i16 %15 to i32
2127c2143
<   %inc82.i = add nsw i32 %srcIndex.1140.i, 2
---
>   %inc82.i = add nsw i32 %indvars.i, 2
2141c2157
<   %sub91.i = sub nsw i32 %srcIndex.1140.i, %prev.1.ph.i.ph
---
>   %sub91.i = sub nsw i32 %indvars.i, %prev.1.ph.i.ph
2177,2179c2193,2195
<   %11 = load ptr, ptr %s.i, align 8
<   %sub95.i = sub nsw i32 %srcIndex.3.i, %srcIndex.1140.i
<   %call96.i = call fastcc noundef i32 @_ZN6icu_7512_GLOBAL__N_112appendResultEPDsiiiPKDsijPNS_5EditsE(ptr noundef %dest, i32 noundef %retval.0.i84.i, i32 noundef %destCapacity, i32 noundef %call86.i, ptr noundef %11, i32 noundef %sub95.i, i32 noundef %options, ptr noundef %edits)
---
>   %16 = load ptr, ptr %s.i, align 8
>   %sub95.i = sub nsw i32 %srcIndex.3.i, %indvars.i
>   %call96.i = call fastcc noundef i32 @_ZN6icu_7512_GLOBAL__N_112appendResultEPDsiiiPKDsijPNS_5EditsE(ptr noundef %dest, i32 noundef %retval.0.i84.i, i32 noundef %destCapacity, i32 noundef %call86.i, ptr noundef %16, i32 noundef %sub95.i, i32 noundef %options, ptr noundef %edits)
2184c2200
<   %srcIndex.1.lcssa.i = phi i32 [ %srcIndex.1.be.i, %while.cond.backedge.i ], [ %srcIndex.1.ph.i, %while.cond.outer.i ]
---
>   %srcIndex.1.lcssa.i = phi i32 [ %srcLength, %while.cond.backedge.i ], [ %srcIndex.1.ph.i, %while.cond.outer.i ]
2228,2229c2244,2245
<   %12 = load i32, ptr %errorCode, align 4
<   %cmp.i.i14 = icmp sgt i32 %12, 0
---
>   %17 = load i32, ptr %errorCode, align 4
>   %cmp.i.i14 = icmp sgt i32 %17, 0

[akshayrdeodhar]

Sorry! Missed this! Am taking a look at the comment in the main MR now!

[dtcxzyw]

/add-label reviewed,regression