[flang-commits] [flang] [AMDGPU] Add another SIFoldOperands instance after shrink (PR #67878)

[Stanislav Mekhanoshin via flang-commits]

rampitec wrote:

> I've taken another look at this. The patch does not show any benefit from running another `SIFoldOperands` pass _after_ `SIShrinkInstructions` per se; you get exactly the same results (modulo a couple of add instructions that have their operands commuted differently) if you put the second `SIFoldOperands` run _before_ `SIShrinkInstructions` instead.
> 
> In other words `SIFoldOperands` is not idempotent, and the reason for the that seems to be:
> 
> > And the reason it only happens for some SUBREV instructions is even more convoluted. It's because SIFoldOperands will sometimes shrink V_SUB_CO_U32_e64 to V_SUBREV_CO_U32_e32 even it does not manage to fold anything into it. This does seem wrong and is probably worth a closer look.
> 
> This goes back to https://reviews.llvm.org/D51345. Notice how the code that was added to `updateOperand` does the shrinking but does not actually do any folding; it returns before we get to `Old.ChangeToImmediate`/`Old.substVirtReg`. A second run of `SIFoldOperands` will see the shrunk instruction and fold into it.

Yes, this is mostly old targets without no-carry add/sub and the impact is on these 2 instructions which needs to be shrunk before folding. Although fold operands' shrinking capabilities are really limited compared to the shrink pass.

https://github.com/llvm/llvm-project/pull/67878