[PATCH] D105360: [PowerPC] Fix popcntb XL Compat Builtin for 32bit

[Eli Friedman via Phabricator via cfe-commits]

efriedma added inline comments.


================
Comment at: llvm/lib/Target/PowerPC/PPCInstr64Bit.td:831
+// 64bit version of popcntb for 64bit sized unsigned long. 
+let isCodeGenOnly = 1 in
+def POPCNTB8 : XForm_11<31, 122, (outs g8rc:$rA), (ins g8rc:$rS),
----------------
nemanjai wrote:
> efriedma wrote:
> > I'm sort of confused by the instruction variations here... let me see if I'm understanding correctly:
> > 
> > 1. popcntb always produces a 64-bit result.
> > 2. The 32-bit variation of the intrinsic just throws away the high bits.
> > 3. We can't use the 64-bit instruction in "32-bit" mode because we marked the register class illegal in isel, and Feature64BitRegs has been marked "beta" for the last 15 years.
> > 4. Therefore, there are two versions of the instruction: the real instruction that produces a 64-bit result, and a fake version of the instruction we can use in 32-bit mode.
> This is correct. We have two GPR register classes that model the same register file. The `gprc` registers pretend the high bits don't exist. The two versions of the instructions that have GPR operands are quite common in PPC.
> 
> There have historically been some operating systems on which context switches spill the full width GPRs even in 32-bit mode. Others spill only the low 32 bits of the GPRs. The addition of two different register classes and `Feature64BitRegs` likely had to do with that.
> There have historically been some operating systems on which context switches 

Oh, that makes sense.  You can't really use the high bits of a 64-bit register if they can get randomly zeroed out from under you.

Are any commonly used operating systems today in the "spill only the low 32 bits" category?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D105360/new/

https://reviews.llvm.org/D105360