Summary:
Previously, if the threshold was 2, we were willing to speculatively
execute 2 cheap instructions in both basic blocks (thus we were willing
to speculatively execute cost = 4), but weren't willing to speculate
when one BB had 3 instructions and other one had no instructions,
even thought that would have total cost of 3.
This looks inconsistent to me.
I don't think `cmov`-like instructions will start executing
until both of it's inputs are available: https://godbolt.org/z/zgHePf
So i don't see why the existing behavior is the correct one.
Also, let's add it's own `cl::opt` for this threshold,
with default=4, so it is not stricter than the previous threshold:
will allow to fold when there are 2 BB's each with cost=2.
And since the logic has changed, it will also allow to fold when
one BB has cost=3 and other cost=1, or there is only one BB with cost=4.
This is an alternative solution to D65148:
This fix is mainly motivated by `signbit-like-value-extension.ll` test.
That pattern comes up in JPEG decoding, see e.g.
`Figure F.12 – Extending the sign bit of a decoded value in V`
of `ITU T.81` (JPEG specification).
That branch is not predictable, and it is within the innermost loop,
so the fact that that pattern ends up being stuck with a branch
instead of `select` (i.e. `CMOV` for x86) is unlikely to be beneficial.
This has great results on the final assembly (vanilla test-suite + RawSpeed): (metric pass - D67240)
| metric | old | new | delta | % |
| x86-mi-counting.NumMachineFunctions | 37720 | 37721 | 1 | 0.00% |
| x86-mi-counting.NumMachineBasicBlocks | 773545 | 771181 | -2364 | -0.31% |
| x86-mi-counting.NumMachineInstructions | 7488843 | 7486442 | -2401 | -0.03% |
| x86-mi-counting.NumUncondBR | 135770 | 135543 | -227 | -0.17% |
| x86-mi-counting.NumCondBR | 423753 | 422187 | -1566 | -0.37% |
| x86-mi-counting.NumCMOV | 24815 | 25731 | 916 | 3.69% |
| x86-mi-counting.NumVecBlend | 17 | 17 | 0 | 0.00% |
We significantly decrease basic block count, notably decrease instruction count,
significantly decrease branch count and very significantly increase `cmov` count.
Performance-wise, unsurprisingly, this has great effect on
target RawSpeed benchmark. I'm seeing 5 **major** improvements:
```
Benchmark Time CPU Time Old Time New CPU Old CPU New
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Samsung/NX3000/_3184416.SRW/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Samsung/NX3000/_3184416.SRW/threads:8/process_time/real_time_mean -0.3064 -0.3064 226.9913 157.4452 226.9800 157.4384
Samsung/NX3000/_3184416.SRW/threads:8/process_time/real_time_median -0.3057 -0.3057 226.8407 157.4926 226.8282 157.4828
Samsung/NX3000/_3184416.SRW/threads:8/process_time/real_time_stddev -0.4985 -0.4954 0.3051 0.1530 0.3040 0.1534
Kodak/DCS760C/86L57188.DCR/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Kodak/DCS760C/86L57188.DCR/threads:8/process_time/real_time_mean -0.1747 -0.1747 80.4787 66.4227 80.4771 66.4146
Kodak/DCS760C/86L57188.DCR/threads:8/process_time/real_time_median -0.1742 -0.1743 80.4686 66.4542 80.4690 66.4436
Kodak/DCS760C/86L57188.DCR/threads:8/process_time/real_time_stddev +0.6089 +0.5797 0.0670 0.1078 0.0673 0.1062
Sony/DSLR-A230/DSC08026.ARW/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Sony/DSLR-A230/DSC08026.ARW/threads:8/process_time/real_time_mean -0.1598 -0.1598 171.6996 144.2575 171.6915 144.2538
Sony/DSLR-A230/DSC08026.ARW/threads:8/process_time/real_time_median -0.1598 -0.1597 171.7109 144.2755 171.7018 144.2766
Sony/DSLR-A230/DSC08026.ARW/threads:8/process_time/real_time_stddev +0.4024 +0.3850 0.0847 0.1187 0.0848 0.1175
Canon/EOS 77D/IMG_4049.CR2/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Canon/EOS 77D/IMG_4049.CR2/threads:8/process_time/real_time_mean -0.0550 -0.0551 280.3046 264.8800 280.3017 264.8559
Canon/EOS 77D/IMG_4049.CR2/threads:8/process_time/real_time_median -0.0554 -0.0554 280.2628 264.7360 280.2574 264.7297
Canon/EOS 77D/IMG_4049.CR2/threads:8/process_time/real_time_stddev +0.7005 +0.7041 0.2779 0.4725 0.2775 0.4729
Canon/EOS 5DS/2K4A9929.CR2/threads:8/process_time/real_time_pvalue 0.0000 0.0000 U Test, Repetitions: 49 vs 49
Canon/EOS 5DS/2K4A9929.CR2/threads:8/process_time/real_time_mean -0.0354 -0.0355 316.7396 305.5208 316.7342 305.4890
Canon/EOS 5DS/2K4A9929.CR2/threads:8/process_time/real_time_median -0.0354 -0.0356 316.6969 305.4798 316.6917 305.4324
Canon/EOS 5DS/2K4A9929.CR2/threads:8/process_time/real_time_stddev +0.0493 +0.0330 0.3562 0.3737 0.3563 0.3681
```
That being said, it's always best-effort, so there will likely
be cases where this worsens things.
Reviewers: efriedma, craig.topper, dmgreen, jmolloy, fhahn, Carrot, hfinkel, chandlerc
Reviewed By: jmolloy
Subscribers: xbolva00, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67318
llvm-svn: 372009
As it's causing some bot failures (and per request from kbarton).
This reverts commit r358543/ab70da07286e618016e78247e4a24fcb84077fda.
llvm-svn: 358546
This is the last general step to allow more IR-level speculation with a safety harness in place in CodeGenPrepare.
The intent is to restore the behavior enabled by:
http://reviews.llvm.org/rL228826
but prevent bad performance such as:
https://llvm.org/bugs/show_bug.cgi?id=24818
Earlier patches in this sequence:
D12882 (disable SimplifyCFG speculation for expensive instructions)
D13297 (have CGP despeculate expensive ops)
D14630 (have CGP despeculate special versions of cttz/ctlz)
As shown in the test cases, we only have two instructions currently affected: ctz for some x86 and fdiv generally.
Allowing exactly one expensive instruction is a bit of a hack, but it lines up with what is currently implemented
in CGP. If we make the despeculation more general in CGP, we can make the speculation here more liberal.
A follow-up patch will adjust the cost for sqrt and possibly other typically expensive math intrinsics (currently
everything is cheap by default). GPU targets would likely want to override those expensive default costs (just as
they probably should already override the cost of div/rem) because just about any math is cheaper than control-flow
on those targets.
Differential Revision: http://reviews.llvm.org/D15213
llvm-svn: 255660
...because that's what the cost model was intended to do.
As discussed in D12882, this fix has a temporary unintended consequence for
SimplifyCFG: it causes us to not speculate an fdiv. However, two wrongs make
PR24818 right, and two wrongs make PR24343 act right even though it's really
still wrong.
I intend to correct SimplifyCFG and add to CodeGenPrepare to account for this
cost model change and preserve the righteousness for the bug report cases.
https://llvm.org/bugs/show_bug.cgi?id=24818https://llvm.org/bugs/show_bug.cgi?id=24343
Differential Revision: http://reviews.llvm.org/D12882
llvm-svn: 248439
This reverts isSafeToSpeculativelyExecute's use of ReadNone until we
split ReadNone into two pieces: one attribute which reasons about how
the function reasons about memory and another attribute which determines
how it may be speculated, CSE'd, trap, etc.
llvm-svn: 246331
Any call which is side effect free is trivially OK to speculate. We
already had similar logic in EarlyCSE and GVN but we were missing it
from isSafeToSpeculativelyExecute.
This fixes PR24601.
llvm-svn: 246232
These are named following the IEEE-754 names for these
functions, rather than the libm fmin / fmax to avoid
possible ambiguities. Some languages may implement something
resembling fmin / fmax which return NaN if either operand is
to propagate errors. These implement the IEEE-754 semantics
of returning the other operand if either is a NaN representing
missing data.
llvm-svn: 220341
Roman Lebedev