[LLVM] Update to llvm/llvm-project@1f20eee6dc36 (#5308)

This pulls in the AMDGPU backend support for the
gfx950 target.

We need to fix the rewrites in `Combine.td` given that
llvm/llvm-project#112700 adds
a new attribute for denorm mode for `arith.addf`.

---------

Co-authored-by: Lei Zhang <[email protected]>

Author: jungpark-mlir

cmake/llvm-hash.txt

@@ -1 +1 @@
-86b69c31642e98f8357df62c09d118ad1da4e16a
+1f20eee6dc367bd202895e3eedb03974a628ef16

lib/Dialect/Triton/Transforms/Combine.td

@@ -17,7 +17,7 @@ def CombineDotAddIPattern : Pat<
         [(Constraint<CPred<"isZero($0)">> $c),
          (Constraint<CPred<"res->hasOneUse()">, "dot result has a single use">)]>;
 def CombineDotAddFPattern : Pat<
-        (Arith_AddFOp $d, (TT_DotOp:$res $a, $b, $c, $inputPrecision, $maxNumImpreciseAcc), $fastmath),
+        (Arith_AddFOp $d, (TT_DotOp:$res $a, $b, $c, $inputPrecision, $maxNumImpreciseAcc), $fastmath, $denorm),
         (TT_DotOp $a, $b, $d, $inputPrecision, $maxNumImpreciseAcc, (location $res)),
         [(Constraint<CPred<"isZero($0)">> $c),
          (Constraint<CPred<"::llvm::cast<::mlir::IntegerAttr>($0).getInt() == 0">> $maxNumImpreciseAcc),
@@ -29,7 +29,7 @@ def CombineDotAddIRevPattern : Pat<
         [(Constraint<CPred<"isZero($0)">> $c),
          (Constraint<CPred<"res->hasOneUse()">, "dot result has a single use">)]>;
 def CombineDotAddFRevPattern : Pat<
-        (Arith_AddFOp (TT_DotOp:$res $a, $b, $c, $inputPrecision, $maxNumImpreciseAcc), $d, $fastmath),
+        (Arith_AddFOp (TT_DotOp:$res $a, $b, $c, $inputPrecision, $maxNumImpreciseAcc), $d, $fastmath, $denorm),
         (TT_DotOp $a, $b, $d, $inputPrecision, $maxNumImpreciseAcc, (location $res)),
         [(Constraint<CPred<"isZero($0)">> $c),
          (Constraint<CPred<"::llvm::cast<::mlir::IntegerAttr>($0).getInt() == 0">> $maxNumImpreciseAcc),