JIT: Unblock Vector###<long> intrinsics on x86#112728
JIT: Unblock Vector###<long> intrinsics on x86#112728BruceForstall merged 20 commits intodotnet:mainfrom
Conversation
ghost
added
the
area-CodeGen-coreclr
dotnet-policy-service
bot
added
the
community-contribution
saucecontrol
force-pushed
the
createscalar64
branch
from
197fac5 to
628d4f8
Compare
saucecontrol
force-pushed
the
createscalar64
branch
from
628d4f8 to
3a130c8
Compare
| // Keep casts with operands usable from memory. | ||
| if (castOp->isContained() || castOp->IsRegOptional()) | ||
| { | ||
| return op; | ||
| } |
There was a problem hiding this comment.
This condition, added in #72719, made this method effectively useless. Removing it was a zero-diff change. I can look in future at containing the casts rather than removing them.
|
|
||
| GenTree* op2 = node->Op(2); | ||
|
|
||
| // TODO-XArch-AVX512 : Merge the NI_Vector512_Create and NI_Vector256_Create paths below. |
There was a problem hiding this comment.
The churn in this section is just taking care of this TODO
| tmp2 = InsertNewSimdCreateScalarUnsafeNode(TYP_SIMD16, op2, simdBaseJitType, 16); | ||
| LowerNode(tmp2); | ||
|
|
||
| node->ResetHWIntrinsicId(NI_SSE_MoveLowToHigh, tmp1, tmp2); |
There was a problem hiding this comment.
Changing this to UnpackLow shows up as a regression in a few places, because movlhps is one byte smaller, but it enables other optimizations since unpcklpd takes a memory operand plus mask and embedded broadcast.
Vector128.Create(double, 1.0):
- vmovups xmm0, xmmword ptr [reloc @RWD00]
- vmovlhps xmm0, xmm1, xmm0
+ vunpcklpd xmm0, xmm1, qword ptr [reloc @RWD00] {1to2}There was a problem hiding this comment.
This should probably be peepholed back to vmovlhps if both are from register.
There was a problem hiding this comment.
I was thinking the same but would rather save that for a followup. llvm has a replacement list of equivalent instructions that have different sizes, and unpcklpd is on it, as are things like vpermilps, which is replaced by pshufd.
It's worth having a discussion about whether we'd also want to do replacements that switch between float and integer domains. I'll open an issue.
There was a problem hiding this comment.
I looked at this again, and there's actually only a size difference for legacy SSE encoding, so it's probably not worth special casing.
| if (varDsc->lvIsParam) | ||
| { | ||
| // Promotion blocks combined read optimizations for SIMD loads of long params | ||
| return; | ||
| } |
There was a problem hiding this comment.
In isolation, this change produced a small number of diffs and was mostly an improvement. A few regressions show up in the SPMI reports, but the overall impact is good, especially considering the places we can load a long to vector with movq
|
It occurred to me the optimization to emit |
tannergooding
left a comment
tannergooding
left a comment
There was a problem hiding this comment.
CC. @dotnet/jit-contrib, @EgorBo, @BruceForstall for secondary review
4 participants
Resolves #11626
This resolves a large number of TODOs around HWIntrinsic expansion involving scalar longs on x86.
The most significant change here is in promoting
CreateScalarandToScalarto be code generating intrinsics instead of converting them to other intrinsics at lowering. This was necessary in order to handle emittingmovqfor scalar long loads/stores but also unlocks several other optimizations since we can now allowCreateScalarandToScalarto be contained and can specialize codegen depending on whether they end up loading/storing from/to memory or not. Some example improvements on x64:Vector128.CreateScalar(ref float):Vector128.CreateScalar(ref double):ref byte = Vector128<byte>.ToScalar():Vector<byte>.ToScalar()And the less realistic, but still interesting
Sse.AddScalar(Vector128.CreateScalar(ref float), Vector128.CreateScalar(ref float)).ToScalar():This also removes some redundant casts for
CreateScalarof small types. Previously, a zero-extending cast was inserted unconditionally and was sometimes removed by peephole opt on x64 but often wasn't.Vector128.CreateScalar(short):Vector128.CreateScalar(checked((byte)val)):cmp edx, 255 ja SHORT G_M000_IG04 mov eax, edx - movzx rax, al - vmovd xmm0, rax + vmovd xmm0, eaxVector128.CreateScalar(ref sbyte):x86 diffs are much more significant, because of the newly-enabled intrinsic expansion: