[llvm] [AMDGPU][Attributor] Rework update of `AAAMDWavesPerEU` (PR #123995)

[Johannes Doerfert via llvm-commits]

================
@@ -1323,6 +1298,60 @@ struct AAAMDGPUNoAGPR
 
 const char AAAMDGPUNoAGPR::ID = 0;
 
+/// The final check and update of the attribute 'amdgpu-waves-per-eu' based on
+/// the determined 'amdgpu-flat-work-group-size' attribute. We can't do this
+/// during attributor run because the two attributes grow in opposite direction,
+/// we should not use any intermediate value to calculate waves per eu until we
+/// have a determined flat workgroup size.
+static void updateWavesPerEU(Module &M, TargetMachine &TM) {
+  for (Function &F : M) {
+    if (F.isDeclaration())
+      continue;
+
+    const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
+
+    std::optional<std::pair<unsigned, std::optional<unsigned>>>
+        FlatWgrpSizeAttr =
+            AMDGPU::getIntegerPairAttribute(F, "amdgpu-flat-work-group-size");
+
+    unsigned MinWavesPerEU = ST.getMinWavesPerEU();
+    unsigned MaxWavesPerEU = ST.getMaxWavesPerEU();
+
+    unsigned MinFlatWgrpSize = ST.getMinFlatWorkGroupSize();
+    unsigned MaxFlatWgrpSize = ST.getMaxFlatWorkGroupSize();
+    if (FlatWgrpSizeAttr.has_value()) {
+      MinFlatWgrpSize = FlatWgrpSizeAttr->first;
+      MaxFlatWgrpSize = *(FlatWgrpSizeAttr->second);
+    }
+
+    // Start with the max range.
----------------
jdoerfert wrote:

"best range", not "max"

https://github.com/llvm/llvm-project/pull/123995