Change default opt-level for Config to speed.

[peterhuene]

This PR changes the default opt-level for a new Config to speed.

Fixes #981.

[peterhuene]

Looking into the various test failures.

[peterhuene]

Looks like some of the SIMD spec testsuite tests are failing when optimizing for speed:

Cranelift::simd::simd_address
Cranelift::simd::simd_align
Cranelift::simd::simd_bitwise
Cranelift::simd::simd_f32x4_cmp
Cranelift::simd::simd_f64x2_cmp
Cranelift::simd::simd_i16x8_cmp
Cranelift::simd::simd_i32x4_cmp
Cranelift::simd::simd_i8x16_cmp
Cranelift::simd::simd_store

[alexcrichton]

Want to update the wast testsuites with a FIXME where we set the optimization level to none, but only for the simd tests?

[peterhuene]

Can do. Pinging @abrown relating to the SIMD failures when we optimize for speed.

Looks like there are additional failures for trap-related tests on Windows that I'll investigate, possibly related to unwind.

[peterhuene]

For posterity, here's one such SIMD failure:

Error: failed directive on tests/spec_testsuite/proposals/simd/simd_f32x4_cmp.wast:7799:1

Caused by:
    0: WebAssembly failed to compile
    1: Compilation error: function u0:3(i64 vmctx [%rdi], i64 [%rsi]) system_v {
           gv0 = vmctx
           gv1 = load.i64 notrap aligned readonly gv0+4
           heap0 = static gv1, min 0, bound 0x0001_0000_0000, offset_guard 0x8000_0000, index_type i32

                                       block0(v0: i64, v1: i64):
       @0134 [RexOp1pu_id#b8]              v4 = iconst.i32 0
       @0136 [RexOp1umr#89]                v15 = uextend.i64 v4
       @0136 [RexOp1ldDisp8#808b]          v16 = load.i64 notrap aligned readonly v0+4
       @0136 [DynRexOp1rr#8001]            v5 = iadd v16, v15
       @0136 [DynRexOp2fld#410]            v6 = load_complex.i8x16 v16+v15
       ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
       ; error: inst2 (v6 = load_complex.i8x16 v16+v15): Instruction failed to re-encode DynRexOp2fld#410

       @013a [Op1jmpb#eb]                  jump block4(v6)

                                       block4(v3: i8x16):
       @013d [RexOp1pu_id#b8]              v8 = iconst.i32 1
       @013f [RexOp1umr#89]                v17 = uextend.i64 v8
       @013f [RexOp1ldDisp8#808b]          v18 = load.i64 notrap aligned readonly v0+4
       @013f [DynRexOp1rr#8001]            v9 = iadd v18, v17
       @013f [DynRexOp2fld#410]            v10 = load_complex.i8x16 v18+v17
       ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
       ; error: inst6 (v10 = load_complex.i8x16 v18+v17): Instruction failed to re-encode DynRexOp2fld#410

       @0143 [Op1jmpb#eb]                  jump block5(v10)

                                       block5(v7: i8x16):
       @0144 [null_fpr#00]                 v11 = raw_bitcast.f32x4 v3
       @0144 [null_fpr#00]                 v12 = raw_bitcast.f32x4 v7
       @0144 [RexOp2pfcmp#4c2]             v13 = fcmp le v11, v12
       @0146 [null_fpr#00]                 v14 = raw_bitcast.i8x16 v13
       @0146 [Op1jmpb#eb]                  jump block3(v14)

                                       block3(v2: i8x16):
       @0148 [Op1jmpb#eb]                  jump block2

                                       block2:
       @0149 [Op1jmpb#eb]                  jump block1

                                       block1:
       @0149 [Op1ret#c3]                   return
       }

       ; 2 verifier errors detected (see above). Compilation aborted.

[peterhuene]

The SIMD tests are the only ones that failing but I am seeing a bunch of additional panic output during the test executions:

thread '<unnamed>' panicked at 'assertion failed: ok', C:\Users\phuene\.cargo\registry\src\github.life-white.uk-1ecc6299db9ec823\cranelift-codegen-0.59.0\src\postopt.rs:353:5

[abrown]

@peterhuene, thanks for the heads up. I suspect this is related to bytecodealliance/cranelift#1388 (using REX prefixes to access more registers) but I'm not exactly sure I understand what the effects of speed are. I'll look at this more tomorrow.

[abrown]

The simd_f32x4_cmp.wast example above is actually failing because load_complex does not (yet) have an encoding for the vector types; that should be relatively easy to add. I looked at simd_address.wast and it is failing for the same reason; postop.rs is replacing load with load_complex and the assertion at postop.rs:353 is failing because that instruction has no encoding yet. I agree with @alexcrichton: turn off the optimization for now until I can imlement load_complex for SIMD.

[peterhuene]

@abrown thanks for looking into it! I'll disable optimization for the tests for now.

[peterhuene]

I've opened bytecodealliance/cranelift#1409 to track the SIMD issue.