[clang] [llvm] [mlir] [AMDGPU] Add the support for 45-bit buffer resource (PR #159702)

[Krzysztof Drewniak via llvm-commits]

================
@@ -5905,33 +5905,46 @@ bool AMDGPULegalizerInfo::legalizePointerAsRsrcIntrin(
   Register Flags = MI.getOperand(5).getReg();
 
   LLT S32 = LLT::scalar(32);
+  LLT S64 = LLT::scalar(64);
 
   B.setInsertPt(B.getMBB(), ++B.getInsertPt());
-  auto Unmerge = B.buildUnmerge(S32, Pointer);
-  Register LowHalf = Unmerge.getReg(0);
-  Register HighHalf = Unmerge.getReg(1);
-
-  auto AndMask = B.buildConstant(S32, 0x0000ffff);
-  auto Masked = B.buildAnd(S32, HighHalf, AndMask);
-
-  MachineInstrBuilder NewHighHalf = Masked;
-  std::optional<ValueAndVReg> StrideConst =
-      getIConstantVRegValWithLookThrough(Stride, MRI);
-  if (!StrideConst || !StrideConst->Value.isZero()) {
-    MachineInstrBuilder ShiftedStride;
-    if (StrideConst) {
-      uint32_t StrideVal = StrideConst->Value.getZExtValue();
-      uint32_t ShiftedStrideVal = StrideVal << 16;
-      ShiftedStride = B.buildConstant(S32, ShiftedStrideVal);
-    } else {
-      auto ExtStride = B.buildAnyExt(S32, Stride);
-      auto ShiftConst = B.buildConstant(S32, 16);
-      ShiftedStride = B.buildShl(S32, ExtStride, ShiftConst);
-    }
-    NewHighHalf = B.buildOr(S32, Masked, ShiftedStride);
+
+  if (ST.has45BitNumRecordsBufferResource()) {
+    // Build the lower 64-bit value, which has a 57-bit base and the lower 7-bit
+    // num_records.
+    LLT PtrIntTy = LLT::scalar(MRI.getType(Pointer).getSizeInBits());
+    auto PointerInt = B.buildPtrToInt(PtrIntTy, Pointer);
+    auto ExtPointer = B.buildAnyExtOrTrunc(S64, PointerInt);
+    auto NumRecordsLHS = B.buildShl(S64, NumRecords, B.buildConstant(S32, 57));
+    Register LowHalf = B.buildOr(S64, ExtPointer, NumRecordsLHS).getReg(0);
+
+    // Build the higher 64-bit value, which has the higher 38-bit num_records,
+    // 6-bit zero (omit), 16-bit stride and scale and 4-bit flag.
+    auto NumRecordsRHS = B.buildLShr(S64, NumRecords, B.buildConstant(S32, 7));
+    auto ExtStride = B.buildAnyExt(S64, Stride);
+    auto ShiftedStride = B.buildShl(S64, ExtStride, B.buildConstant(S32, 44));
+    auto ExtFlags = B.buildAnyExt(S64, Flags);
+    auto NewFlags = B.buildAnd(S64, ExtFlags, B.buildConstant(S64, 0x3));
----------------
krzysz00 wrote:

In the interests of not needing to touch this once someone uses the type field for something, can we drop this AND and go with "we trust the programmer to the right thing here"?

https://github.com/llvm/llvm-project/pull/159702