Mine.

This patch:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070319/046250.html

fixes the "simple" test case that Chris narrowed it down to:
define i1 @​test(i32 %tmp13) {
entry:
%tmp14 = shl i32 %tmp13, 12 ; [#uses=1]
%tmp15 = lshr i32 %tmp14, 12 ; [#uses=1]
%res = icmp ne i32 %tmp15, 0 ; :3 [#uses=1]
ret i1 %res
}

InstCombine now generates:
define i1 @​test(i32 %tmp13) {
entry:
%tmp14.mask = and i32 %tmp13, 1048575 ; [#uses=1]
%res = icmp ne i32 %tmp14.mask, 0 ; [#uses=1]
ret i1 %res
}

Unfortunately, this didn't fix SingleSource/Regression/C/2003-05-21-UnionBitFields.c

Test cases here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070319/046251.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070326/046257.html

Fixes here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070319/046250.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070319/046252.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070326/046256.html

Unfortunately, things are still broken: Qt crashes, llvm-gcc bootstrap too.
Haven't tried mozilla, but I think it's broken too. I'm working on testcase.

.ll before instcombine
Reduction from llvm-gcc bootstrap

.ll after instcombine

I've attached testcase from llvm-gcc bootstrap. Actually, it's
build/gengtype-yacc.o built on the stage1 (it was ok before! before fix :) ).

Fortunately, diff between .ll's is not so big. I've found only one suspicious
place (I can be surely wrong and miss something):

• %tmp52 = shl i32 %yystacksize.5, 1 ; [#uses=1]
• %tmp53 = add i32 %tmp52, 3 ; [#uses=1]
• %tmp55 = lshr i32 %tmp53, 2 ; [#uses=2]

• lshr i32 %yystacksize.5, 1 ; :4 [#uses=1]
• %tmp55 = and i32 %4, 3 ; [#uses=2]

It seems, that presence of "add" is not taken into account...

It looks like 2003-09-18-BitFieldTest.c is failing because
%tmp11 = shl i32 %tmp10, 30 ; [#uses=1]
%tmp13 = lshr i32 %tmp11, 31 ; [#uses=1]
%tmp1314 = trunc i32 %tmp13 to i8 ; [#uses=1]
%tmp131415 = zext i8 %tmp1314 to i32 ; [#uses=1]
is being turned into
lshr i32 %tmp10, 1 ; :0 [#uses=1]

Maybe it is related.

FWIW, the ppc llvm-gcc bootstrap is miscompiling itself, and the following new programs fail:

New Test Failures:
Applications/d/make_dparser [LLC compile, LLC-BETA compile, JIT codegen, CBE, LLC, LLC-BETA, JIT]
Applications/JM/ldecod/ldecod [LLC compile, LLC-BETA compile, JIT codegen, CBE, LLC, LLC-BETA, JIT]
Benchmarks/Fhourstones-3.1/fhourstones3.1 [LLC compile, LLC-BETA compile, JIT codegen, CBE, LLC,
LLC-BETA, JIT]
Benchmarks/MiBench/consumer-typeset/consumer-typeset [LLC compile, LLC-BETA compile, JIT
codegen, CBE, LLC, LLC-BETA, JIT]
SPEC/CFP2006/444.namd/444.namd [LLC compile, LLC-BETA compile, JIT codegen, CBE, LLC, LLC-
BETA, JIT]
SPEC/CINT2000/176.gcc/176.gcc [LLC compile, LLC-BETA compile, JIT codegen, CBE, LLC, LLC-BETA,
JIT]
SPEC/CINT2006/403.gcc/403.gcc [LLC compile, LLC-BETA compile, JIT codegen, , LLC, LLC-BETA]
UnitTests/2004-06-20-StaticBitfieldInit [LLC compile, LLC-BETA compile, JIT codegen, CBE, LLC, LLC-
BETA, JIT]
UnitTests/2006-01-23-InitializedBitField [LLC compile, LLC-BETA compile, JIT codegen, CBE, LLC, LLC-
BETA, JIT]

-Chris

With this patch:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070326/046268.html
all of the nightly test regressions are now passing, except for:
MultiSource/Applications/JM/ldecod

bugpoint has reduced an -instcombine test case with this result:
1998,2002c1998,2001
< %tmp133 = and i32 %b8, 1 ; [#uses=1]
< %tmp134 = shl i32 %tmp133, 3 ; [#uses=3]
< %tmp136 = ashr i32 %b8, 1 ; [#uses=1]
< %tmp137 = shl i32 %tmp136, 3 ; [#uses=3]
< %tmp139 = ashr i32 %tmp134, 2 ; [#uses=1]

  %tmp133 = shl i32 %b8, 3                ; <i32> [#uses=1]
  %tmp134 = and i32 %tmp133, 8            ; <i32> [#uses=2]
  shl i32 %b8, 2          ; <i32>:0 [#uses=1]
  %tmp137 = and i32 %0, -8                ; <i32> [#uses=3]

2004c2003
< store i32 %tmp139, i32* %tmp141

  store i32 0, i32* %tmp141

2008,2011c2007,2008
< icmp eq i32 %iftmp.430.0, 0 ; :0 [#uses=1]
< zext i1 %0 to i8 ; :1 [#uses=1]
< icmp ne i8 %1, 0 ; :2 [#uses=1]
< br i1 %2, label %cond_true147.exitStub, label %cond_false252.exitStub

  icmp eq i32 %iftmp.430.0, 0             ; <i1>:1 [#uses=1]
  br i1 %1, label %cond_true147.exitStub, label %cond_false252.exitStub

I'm looking at this now.

This was another bad shift mask. The bugpoint reduced test case is here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070326/046291.html

Fix patch is here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20070326/046290.html

Anton,

This has passed llvm/test, llvm-test, and a bootstrap build of llvm-gcc.

I'm calling it done. If you find more bugs in instcombine (e.g. Qt, SeaMonkey),
they are probably separate issues, please file a new bug.

Thanks,

Reid.

Seems ok to me. Thanks.