[llvm] [AMDGPU] Remove setcc by using add/sub carryout (PR #155255)

[via llvm-commits]

================
@@ -16036,6 +16036,61 @@ SDValue SITargetLowering::performSetCCCombine(SDNode *N,
     }
   }
 
+  // Eliminate setcc by using carryout from add/sub instruction
+
+  // LHS = ADD i64 RHS, Z          LHSlo = UADDO       i32 RHSlo, Zlo
+  // setcc LHS ult RHS     ->      LHSHi = UADDO_CARRY i32 RHShi, Zhi
+  // similarly for subtraction
+
+  // LHS = ADD i64 Y, 1            LHSlo = UADDO       i32 Ylo, 1
+  // setcc LHS eq 0        ->      LHSHi = UADDO_CARRY i32 Yhi, 0
+
+  // Don't split a 64-bit add/sub into two 32-bit add/sub instructions for
+  // non-divergent operations.  This can result in lo/hi 32-bit operations
+  // being done in SGPR and VGPR with additional operations being needed
+  // to move operands and/or generate the intermediate carry.
+  if (VT == MVT::i64 && N->isDivergent() &&
----------------
LU-JOHN wrote:

If isDivergent is removed, SALU isel produces:

```
S_UADDO_PSEUDO
S_ADD_CO_PSEUDO

```
finalize-isel regenerates the carry between the adds:

```
S_ADD_I32
S_CSELECT_B64
S_CMP_LG_U64
S_ADDC_U32
```

For comparison, with VALU, isel produces:

```
S_UADDO_PSEUDO
V_ADDC_U32_e64

```

Then si-fix-sgpr-copies produces:

```
V_ADD_CO_U32_e64
V_ADDC_U32_e64

```

https://github.com/llvm/llvm-project/pull/155255