3 1. Create a new DAG -> DAG instruction selector, by adding patterns to the
9 SparcV8 backend skeleton
10 ------------------------
12 This directory houses a 32-bit SPARC V8 backend employing an expander-based
13 instruction selector. It is not yet functionally complete. Watch
14 this space for more news coming soon!
16 Current expected test failures
17 ------------------------------
19 Here are the currently-expected SingleSource failures for V8
20 (Some C++ programs are crashing in libstdc++ at the moment;
23 (llc) SingleSource/Regression/C++/EH/exception_spec_test
24 (llc) SingleSource/Regression/C++/EH/throw_rethrow_test
26 Here are the currently-expected MultiSource failures for V8:
28 (llc,cbe) MultiSource/Applications/d/make_dparser
29 (llc,cbe) MultiSource/Applications/hexxagon
30 (llc) MultiSource/Benchmarks/Fhourstones
31 (llc,cbe) MultiSource/Benchmarks/McCat/03-testtrie
32 (llc) MultiSource/Benchmarks/McCat/18-imp
33 (llc,cbe) MultiSource/Benchmarks/Prolangs-C/bison/mybison
34 (llc,cbe) MultiSource/Benchmarks/Prolangs-C/fixoutput
35 (llc,cbe) MultiSource/Benchmarks/Prolangs-C/gnugo
36 (llc,cbe) MultiSource/Benchmarks/Prolangs-C/plot2fig
37 (llc,cbe) MultiSource/Benchmarks/Ptrdist/anagram
38 (llc,cbe) MultiSource/Benchmarks/FreeBench/analyzer
39 * DANGER * analyzer will run the machine out of VM
40 (I don't know whether the following fail in cbe:)
41 (llc) MultiSource/Benchmarks/FreeBench/distray
42 (llc) MultiSource/Benchmarks/FreeBench/fourinarow
43 (llc) MultiSource/Benchmarks/FreeBench/pifft
44 (llc) MultiSource/Benchmarks/MallocBench/gs
45 (llc) MultiSource/Benchmarks/Prolangs-C++/deriv1
46 (llc) MultiSource/Benchmarks/Prolangs-C++/deriv2
48 Known SPEC failures for V8 (probably not an exhaustive list):
52 (llc) 188.ammp -- FPMover bug?
55 (native,llc,cbe) 126.gcc
56 (native,llc,cbe) 255.vortex
61 * support shl on longs (fourinarow needs this)
62 * support casting 64-bit integers to FP types (fhourstones needs this)
63 * support FP rem (call fmod)
65 * Keep the address of the constant pool in a register instead of forming its
66 address all of the time.
68 * Change code like this:
69 or %o0, %lo(.CPI_main_0), %o0
72 ld [%o0+%lo(.CPI_main_0)], %o0
73 for constant pool access.
75 * We can fold small constant offsets into the %hi/%lo references to constant
76 pool addresses as well.
78 * Directly support select instructions, and fold setcc instructions into them
79 where possible. I think this is what afflicts the inner loop of Olden/tsp
80 (hot block = tsp():no_exit.1.i, overall GCC/LLC = 0.03).