[llvm] [ARM] Swap sides of cmp/cmn based on folding ability (PR #191915)

[John Brawn via llvm-commits]

john-brawn-arm wrote:

Looking at these simple test cases:
```
void cmp_shiftx(int x, int y) {
  if ((x<<1) == y)
    otherfn();
}

void cmp_shifty(int x, int y) {
  if (x == (y<<1))
    otherfn();
}

void cmn_shiftx(int x, int y) {
  if ((x<<1) == -y)
    otherfn();
}

void cmn_shifty(int x, int y) {
  if (x == -(y<<1))
    otherfn();
}
```
and compiling with ``--target=arm-none-eabi -mcpu=cortex-m33 -O2`` it looks like currently we fail to use cmn_shiftx, and it looks like just adding an extra pattern fixes that:
```
diff --git a/llvm/lib/Target/ARM/ARMInstrThumb2.td b/llvm/lib/Target/ARM/ARMInstrThumb2.td
index 4f0cea8b00aa..ad4d2b8c2187 100644
--- a/llvm/lib/Target/ARM/ARMInstrThumb2.td
+++ b/llvm/lib/Target/ARM/ARMInstrThumb2.td
@@ -3556,6 +3556,8 @@ def : T2Pat<(ARMcmpZ GPRnopc:$Rn, (ineg rGPR:$Rm)),
             (t2CMNrr GPRnopc:$Rn, rGPR:$Rm)>;
 def : T2Pat<(ARMcmpZ GPRnopc:$Rn, (ineg t2_so_reg:$ShiftedRm)),
             (t2CMNrs GPRnopc:$Rn, t2_so_reg:$ShiftedRm)>;
+def : T2Pat<(ARMcmpZ (ineg GPRnopc:$Rn), t2_so_reg:$ShiftedRm),
+            (t2CMNrs GPRnopc:$Rn, t2_so_reg:$ShiftedRm)>;
 
 defm t2TST  : T2I_cmp_irs<0b0000, "tst", rGPR,
                           IIC_iTSTi, IIC_iTSTr, IIC_iTSTsi,
```
Do you have any examples which your approach would optimize where the above would not? If so please add tests for them.

https://github.com/llvm/llvm-project/pull/191915