JIT: Accelerate long -> floating casts on x86#113930
JIT: Accelerate long -> floating casts on x86#113930BruceForstall merged 7 commits intodotnet:mainfrom
Conversation
ghost
added
the
area-CodeGen-coreclr
dotnet-policy-service
Bot
added
the
community-contribution
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
saucecontrol
changed the title
|
This is ready for review. To summarize, it simply replaces helper calls with AVX-512 conversion instructions, as follows:
Latest diffs show no regressions other than those expected from inlining. |
|
|
||
| #if defined(FEATURE_HW_INTRINSICS) && defined(TARGET_X86) | ||
| if (!tree->TypeIs(TYP_LONG) && | ||
| !(tree->OperIs(GT_CAST) && varTypeIsLong(tree->AsCast()->CastOp()) && varTypeIsFloating(tree))) |
There was a problem hiding this comment.
nit: negated conditions like this can be hard to read. A small comment covering that we want to handle nodes that produce long or GT_CAST float->long would be beneficial IMO.
There was a problem hiding this comment.
Agreed. I actually plan on extending this to handle casts in the opposite direction as well, and that will make this check even more hairy. I'll do something to simplify it then.
| #endif // FEATURE_HW_INTRINSICS && TARGET_X86 | ||
| { | ||
| return tree->gtNext; | ||
| } |
There was a problem hiding this comment.
The fact that from this point onwards it can now also be GT_CAST float rather than only some NODE long seems like a tricky thing that might trip people up in the future.
tannergooding
left a comment
tannergooding
left a comment
There was a problem hiding this comment.
Changes LGTM. Just a couple minor bits of feedback
CC. @dotnet/jit-contrib, @EgorBo for secondary review
|
Ping @dotnet/jit-contrib, @EgorBo for secondary review |
|
/azp run runtime-coreclr outerloop, Fuzzlyn |
|
Azure Pipelines successfully started running 2 pipeline(s). |
5 participants
This adds support for using EVEX SIMD conversion instructions to handle long/ulong to float/double casts on x86 rather than going through helper calls. With this, all integral -> floating casts are accelerated on x86 with AVX-512.
Typical diff:
G_M57389_IG01: ; bbWeight=1, gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref, nogc <-- Prolog IG - sub esp, 8 - vzeroupper - ;; size=6 bbWeight=1 PerfScore 1.25 + ;; size=0 bbWeight=1 PerfScore 0.00 G_M57389_IG02: ; bbWeight=1, gcrefRegs=00000000 {}, byrefRegs=00000000 {}, byref - push dword ptr [esp+0x10+0x04] - ; npt arg push 0 - push dword ptr [esp+0x0C+0x08] - ; npt arg push 1 - call CORINFO_HELP_LNG2DBL - ; gcr arg pop 2 - fstp qword ptr [esp] - vmovsd xmm0, qword ptr [esp] - vcvtsd2ss xmm0, xmm0, xmm0 + vmovq xmm0, qword ptr [esp+0x04] + vcvtqq2ps xmm0, xmm0 vmovd eax, xmm0 - ;; size=29 bbWeight=1 PerfScore 12.50 + ;; size=16 bbWeight=1 PerfScore 9.00 G_M57389_IG03: ; bbWeight=1, epilog, nogc, extend - add esp, 8 ret 8 - ;; size=6 bbWeight=1 PerfScore 2.25 + ;; size=3 bbWeight=1 PerfScore 2.00Full diffs