fix regression on VI with SDWA gfx9 support

[airlied]

Summary

The SDWA gfx9 support broke VI, this fixes the assert at least.

[arsenm]

Needs test

[sarnex]

This patch fixes the bug I sent to llvm-commits, https://bugs.freedesktop.org/show_bug.cgi?id=101561

Tested-by: Nick Sarnie <[e-mail]>

Thanks,
Sarnex

[airlied]

Consider this a bug report for Sam, I'm not even sure where to start writing a test.

[SamWot]

Hi,
I have no access to bug report: https://bugs.freedesktop.org/show_bug.cgi?id=101561
Can you send me failing .ll?

Thanks

[SamWot]

This change should in fact fix problem, but it is only workaround.
Ultimately problem is that instruction V_CVT_U32_F32 allow omod operand (see SIInstrInfo.td:1435). In fact this operand shouldn't be allowed here.
This should work as quickfix for assert if you need this. But I think it is better to provide full fix.

[sarnex]

Hi Sam,

I just emailed you the good and bad shader dumps.

Thanks,
Sarnex

[arsenm]

lib/Target/AMDGPU/SIPeepholeSDWA.cpp

718

This is also no reason to use hasModifiersSet. That is a convenience wrapper when you don't have the operand, you can check the value of it directly here

[airlied]

At the moment this is more than an assert fix, we have broken VI in favour of GFX9 which nobody has. Please get a workaround, revert or fix into llvm master asap.

lib/Target/AMDGPU/SIPeepholeSDWA.cpp

 
   // Copy omod if present, initialize otherwise if needed
   MachineOperand *OMod = TII->getNamedOperand(MI, AMDGPU::OpName::omod);
-  if (OMod) {
+  if (OMod && TII->hasModifiersSet(MI, AMDGPU::OpName::omod)) {
     assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::omod) != -1);
     SDWAInst.add(*OMod);
   } else if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::omod) != -1) {