JIT: Improve x86 unsigned to floating cast codegen #111595
Conversation
|
cc @dotnet/jit-contrib this is ready for review |
* main: System.Net.Http.WinHttpHandler.StartRequestAsync assertion failed (dotnet#109799) Keep test PDB in helix payload for native AOT (dotnet#111949) Build the RID-specific System.IO.Ports packages in the VMR (dotnet#112054) Always inline number conversions (dotnet#112061) Use Contains{Any} in Regex source generator (dotnet#112065) Update dependencies from https://github.com/dotnet/arcade build 20250130.5 (dotnet#112013) JIT: Transform single-reg args to FIELD_LIST in physical promotion (dotnet#111590) Ensure that math calls into the CRT are tracked as needing vzeroupper (dotnet#112011) Use double.ConvertToIntegerNative where safe to do in System.Random (dotnet#112046) JIT: Compute `fgCalledCount` after synthesis (dotnet#112041) Simplify boolean logic in `TimeZoneInfo` (dotnet#112062) JIT: Update type when return temp is freshly created (dotnet#111948) Remove unused build controls and simplify DotNetBuild.props (dotnet#111986) Fix case-insensitive JSON deserialization of enum member names (dotnet#112028) WasmAppBuilder: Remove double computation of a value (dotnet#112047) Disable LTCG for brotli and zlibng. (dotnet#111805) JIT: Improve x86 unsigned to floating cast codegen (dotnet#111595) simplify x86 special intrinsic imports (dotnet#111836) JIT: Try to retain entry weight during profile synthesis (dotnet#111971) Fix explicit offset of ByRefLike fields. (dotnet#111584)
|
@saucecontrol we're seeing some test failures (#112324, #112325, #112329) in our stress pipelines on x64 related to floating-point casts. Do those look related to this PR? |
|
Yeah, if it's only showing up under JitStressRegs, it's likely bad codegen from an existing bug this PR exposed by changing float->double->float casts to float->float. In which case, #112217 should resolve it. |
|
Actually, not so sure about that last one. This PR changed integral->floating casts, but those failures are on floating->integral, which I don't think has changed recently |
Thanks for confirming; I'll wait to see if #112217 resolves them... |
Fixes #77658
This improves codegen mostly for unsigned to floating types but catches a few other redundant conversions.
Adds support for using AVX-512
vcvtusi2s[sd]for uint -> float/double (only ulong was handled previously) on both x64 and x86.Improves codegen for uint -> float conversions on x64 without AVX-512, removing the intermediate conversion to double.
mov eax, edi xorps xmm0, xmm0 - cvtsi2sd xmm0, rax - cvtsd2ss xmm0, xmm0 + cvtsi2ss xmm0, raxAdds support for direct ulong -> float cast to the x64 SSE2 fallback, resolving a difference in behavior between hardware with AVX-512 vs without, and saving an extra
cvtsd2ssinstruction.xorps xmm0, xmm0 mov rax, rdi shr rax, 1 mov rsi, edi and rsi, 1 or rsi, rax test rdi, rdi cmovns rsi, rdi - cvtsi2sd xmm0, rsi + cvtsi2ss xmm0, rsi jns SHORT G_M37561_IG56 - addsd xmm0, xmm0 + addss xmm0, xmm0 G_M37561_IG56: - cvtsd2ss xmm0, xmm0Removes some redundant float -> double -> float casts.
SPMI Diffs
The only code size regressions are the insertion of
xorpsto clear the upper elements of the target reg for the AVX-512 unsigned conversion instructions. These were previously omitted but should have been there since the unsigned conversions have the same behavior as the signed (i.e. preserving/copying upper elements) and are subject to the same false dependency penalties.GCC emits the
xorpsfor all conversions; Clang skips it for all conversions in simple examples but may emit it in more complex scenarios.https://godbolt.org/z/6aY7fdE3d