Instruction selection failures on SPARC/Solaris

[ggreif]

Bugzilla Link	2105
Resolution	FIXED
Resolved on	Nov 07, 2018 00:17
Version	unspecified
OS	Solaris
CC	@asl

Extended Description

Some tests fail on SPARCv8:

1. ---

FAIL: .../llvm/test/CodeGen/Generic/2008-01-25-dag-combine-mul.ll
Failed with signal(SIGABRT) at line 1
while running: llvm-as < .../llvm/test/CodeGen/Generic/2008-01-25-dag-combine-mul.ll | llc
Cannot yet select: 0xd42d28: i32,i32 = umul_lohi 0xd42c38, 0xd42a58

2. ---

FAIL: .../llvm/test/CodeGen/Generic/storetrunc-fp.ll
Failed with signal(SIGABRT) at line 1
while running: llvm-as < .../llvm/test/CodeGen/Generic/storetrunc-fp.ll | llc
Cannot yet select: 0xd3fe68: ch = store 0xd3fcf8, 0xd403d8, 0xd39320, 0xd404f8 <0xd36278:0> alignment=4

3. ---

FAIL: .../llvm/test/DebugInfo/funccall.ll
Failed with signal(SIGABRT) at line 1
while running: llvm-upgrade < .../llvm/test/DebugInfo/funccall.ll | llvm-as | llc
Cannot yet select llvm.dbg.declare: 0xd3adc8: ch = declare 0xd496b8, 0xd3d6c8, 0xd4b228

Anybody who is at home in this area, can you fix this?

Otherwise some guidance welcome and I try to do it myself.

[asl]

Gabor, as for the third - it looks like debug intrinsics should be marked as 'expand' on Sparc target, so they will be effectively lowered to zero there. This will make test fail, but not abort.

[ggreif]

As a side info, LLVM on Sparc/Solaris10 is in pretty good shape:

            ===  Summary ===

of expected passes 2256

of unexpected failures 5

of unexpected successes 1

of expected failures 9

2 of the 5 fails is related to inline asm,
3 are mentioned above.

The single XPASS comes from the new libm that finally implements powf, sinf etc. for Solaris10.

[lattner]

Fixed:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080225/058920.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080225/058921.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080225/058922.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080225/058923.html

Note that the sparc backend isn't generating the 32 x 32 -> 64 multiplies. In the meantime we generate correct but non crashing code.

Feel free to un-xfail the powf test if you're interested.

[lattner]

that should have been "correct but non-optimal" code :)

[ggreif]

Wonderful!

We are down to

of expected passes 2263

of unexpected failures 2

of unexpected successes 1

of expected failures 8