================
@@ -3232,6 +3264,22 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl
GD, unsigned BuiltinID,
return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
*this, E, Intrinsic::sinh,
Intrinsic::experimental_constrained_sinh));
+ case Builtin::BIsincos:
+ case Builtin::BIsincosf:
+ case Builtin::BIsincosl:
+ case Builtin::BI__builtin_sincos:
+ case Builtin::BI__builtin_sincosf:
+ case Builtin::BI__builtin_sincosl:
+ // Only use the llvm.sincos.* builtin on AArch64 with optimizations.
+ // Currently, getting codegen that is no worse than the direct call
+ // requires using AA during codegen. This is not done at optlevel=none,
+ // and not all targets support this (AArch64 is one of the few known to).
+ if (!getTarget().getTriple().isAArch64() ||
+ CGM.getCodeGenOpts().OptimizationLevel == 0)
+ break;
----------------
MacDue wrote:
I can't control that most targets (outside of AArch64 and a few others) don't
enable AA at codegen, which means the current lowering for `llvm.sincos` will
generate worse code (an additional stack slot + load & store), that would not
be present if using the direct call. It's annoying, but right now I don't have
a fix for other targets, and I don't want to worsen codegen for them.
https://github.com/llvm/llvm-project/pull/114086
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
