LLVM 3.6 giving incorrectly signed zeros

[simonbyrne]

Using LLVM 3.6, branches involving comparison to zero can sometimes return the incorrect sign.

This has come up in #9336 (comment) and #9880 (comment).

Here is a minimal example:

julia> versioninfo()
Julia Version 0.4.0-dev+3639
Commit 7f7e9ae (2015-03-01 22:49 UTC)
Platform Info:
  System: Linux (x86_64-linux-gnu)
  CPU: Intel(R) Xeon(R) CPU E7- 8850  @ 2.00GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Nehalem)
  LAPACK: libopenblas
  LIBM: libopenlibm
  LLVM: libLLVM-3.6.0

julia> function foo(x,y)
           if x==y==0
               return y
           end
           y+1
       end
foo (generic function with 1 method)

julia> foo(0.0,-0.0)
0.0

And the @code_llvm foo(0.0,-0.0):

define double @julia_foo_42667(double, double) {
top:
  call void @llvm.dbg.value(metadata double %0, i64 0, metadata !10, metadata !16)
  call void @llvm.dbg.value(metadata double %1, i64 0, metadata !17, metadata !16)
  %2 = fcmp une double %0, %1, !dbg !18
  br i1 %2, label %L3, label %L1, !dbg !18

L1:                                               ; preds = %top
  call void @llvm.dbg.value(metadata double 0.000000e+00, i64 0, metadata !20, metadata !16)
  %phitmp = fcmp une double %0, 0.000000e+00, !dbg !18
  br i1 %phitmp, label %L3, label %if2, !dbg !18

if2:                                              ; preds = %L1
  ret double %0, !dbg !21

L3:                                               ; preds = %top, %L1
  %3 = fadd double %1, 1.000000e+00, !dbg !22
  ret double %3, !dbg !22
}

The problem here is that the if2 label it is returning %0 instead of %1.

[ViralBShah]

Should we have a label for llvm bugs?

[tkelman]

We can probably rename the "fixed in LLVM 3.5" label to be a general "LLVM issue" instead... or just keep using "upstream" where it's usually pretty clear which upstream is being referred to.

[simonbyrne]

BTW, I haven't filed an upstream issue, as I wasn't really sure where the problem lay. Could someone who understands more about the internals help me out?

[JeffBezanson]

"fixed in LLVM 3.5" is actually a pretty useful label since we'll know exactly which issues to confirm and close when we upgrade LLVM.

[tkelman]

"fixed in LLVM 3.5" is actually a pretty useful label since we'll know exactly which issues to confirm and close when we upgrade LLVM

True, but we can re-appropriate it after that's done. Either "fixed in LLVM 3.7" or something else.

[kmsquire]

Or just add a new label.

[tkelman]

BTW, I haven't filed an upstream issue, as I wasn't really sure where the problem lay. Could someone who understands more about the internals help me out?

I don't really understand the internals very well, but I think the way to communicate these things upstream would be to try to create a standalone .ll file with IR that returns incorrect results on 3.6.0 but correct on 3.5.1?

[simonbyrne]

The problem is that the generated LLVM IR is incorrect (see above), so I'm not sure where the actual problem lies.

[tkelman]

Hm, then the upstream label is probably wrong.

[mbauman]

Isn't it just likely to be an LLVM optimization pass bug? Perhaps try disabling all the optimizations and see if it persists.

[simonbyrne]

Looking at the debug info, here's the two optimisation passes between which the wrong value gets returned:

*** IR Dump After Combine redundant instructions ***
; Function Attrs: sspreq
define double @julia_foo_64477(double, double) #0 {
top:
  call void @llvm.dbg.value(metadata double %0, i64 0, metadata !10, metadata !16)
  call void @llvm.dbg.value(metadata double %1, i64 0, metadata !17, metadata !16)
  %2 = fcmp une double %0, %1, !dbg !18
  br i1 %2, label %L3, label %L1, !dbg !18

L1:                                               ; preds = %top
  call void @llvm.dbg.value(metadata double 0.000000e+00, i64 0, metadata !20, metadata !16)
  %phitmp = fcmp une double %1, 0.000000e+00, !dbg !18
  br i1 %phitmp, label %L3, label %if2, !dbg !18

if2:                                              ; preds = %L1
  ret double %1, !dbg !21

L3:                                               ; preds = %top, %L1
  %3 = fadd double %1, 1.000000e+00, !dbg !22
  ret double %3, !dbg !22
}
*** IR Dump After Global Value Numbering ***
; Function Attrs: sspreq
define double @julia_foo_64477(double, double) #0 {
top:
  call void @llvm.dbg.value(metadata double %0, i64 0, metadata !10, metadata !16)
  call void @llvm.dbg.value(metadata double %1, i64 0, metadata !17, metadata !16)
  %2 = fcmp une double %0, %1, !dbg !18
  br i1 %2, label %L3, label %L1, !dbg !18

L1:                                               ; preds = %top
  call void @llvm.dbg.value(metadata double 0.000000e+00, i64 0, metadata !20, metadata !16)
  %phitmp = fcmp une double %0, 0.000000e+00, !dbg !18
  br i1 %phitmp, label %L3, label %if2, !dbg !18

if2:                                              ; preds = %L1
  ret double %0, !dbg !21

L3:                                               ; preds = %top, %L1
  %3 = fadd double %1, 1.000000e+00, !dbg !22
  ret double %3, !dbg !22
}