[PATCH] D151358: [LegalizeTypes] Improve expansion of wide SMIN/SMAX/UMIN/UMAX

[Eli Friedman via Phabricator via llvm-commits]

efriedma added inline comments.


================
Comment at: llvm/test/CodeGen/AMDGPU/min.ll:552
 
-; EG: MIN_UINT
-; EG: MIN_UINT
+; EG: SETE_INT
+; EG: SETGT_UINT
----------------
Looked more carefully at the output; I don't think this is actually a regression, the total instructions appears the same.

Before:

```
          LSHR * T0.X, KC0[2].Y, literal.x,
        2(2.802597e-45), 0(0.000000e+00)
          SETGE_UINT * T0.W, KC0[3].X, KC0[3].Z,
          CNDE_INT   T0.Z, PV.W, KC0[2].W, KC0[3].Y,
          MIN_UINT * T0.W, KC0[2].W, KC0[3].Y,
          SETE_INT * T1.W, KC0[3].X, KC0[3].Z,
          CNDE_INT   T1.X, PV.W, T0.Z, T0.W,
          MIN_UINT * T1.Y, KC0[3].X, KC0[3].Z,
```

After:

```
          SETE_INT   T0.Z, KC0[3].X, KC0[3].Z,
          SETGT_UINT * T0.W, KC0[3].Z, KC0[3].X,
          SETGT_UINT * T1.W, KC0[3].Y, KC0[2].W,
          CNDE_INT * T0.W, T0.Z, T0.W, PV.W,
          CNDE_INT * T0.Y, PV.W, KC0[3].Z, KC0[3].X,
          CNDE_INT * T0.X, T0.W, KC0[3].Y, KC0[2].W,
          LSHR * T1.X, KC0[2].Y, literal.x,
        2(2.802597e-45), 0(0.000000e+00)
```


================
Comment at: llvm/test/CodeGen/AMDGPU/r600-legalize-umax-bug.ll:23
+; CHECK-NEXT:    4(5.605194e-45), 0(0.000000e+00)
+; CHECK-NEXT:     CNDE_INT T0.X, PV.W, literal.x, T0.X,
 ; CHECK-NEXT:     MOV T0.Y, 0.0,
----------------
Regression.


================
Comment at: llvm/test/CodeGen/ARM/fpclamptosat.ll:2205
 ; SOFT-NEXT:  @ %bb.1: @ %entry
-; SOFT-NEXT:    movs r0, #0
-; SOFT-NEXT:    mvns r0, r0
+; SOFT-NEXT:    mvns r0, r2
 ; SOFT-NEXT:  .LBB28_2: @ %entry
----------------
Regression.  Might need to special-case the constant 4294967295.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D151358/new/

https://reviews.llvm.org/D151358