[PATCH] D147330: [Matrix] Refine cost estimate for dot-product.

[Florian Hahn via Phabricator via llvm-commits]

fhahn marked 2 inline comments as done.
fhahn added inline comments.


================
Comment at: llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp:1345
+    // the returned cost is < 0, the argument is cheaper to use in the
+    // dot-product lowering. An invalid cost is return if the operand is not
+    // supported by the lowering code.
----------------
thegameg wrote:
> 
I adjusted the code slightly to include an estimate of the cost of embedding the columns into a vector, so we can compute valid costs for all cases. I removed that part.


================
Comment at: llvm/test/Transforms/LowerMatrixIntrinsics/dot-product-int.ll:326
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[TMP0:%.*]] = mul <8 x i64> [[A:%.*]], [[B:%.*]]
-; CHECK-NEXT:    [[TMP1:%.*]] = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> [[TMP0]])
-; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <1 x i64> poison, i64 [[TMP1]], i64 0
-; CHECK-NEXT:    ret <1 x i64> [[TMP2]]
+; CHECK-NEXT:    [[SPLIT:%.*]] = shufflevector <8 x i64> [[A:%.*]], <8 x i64> poison, <1 x i32> zeroinitializer
+; CHECK-NEXT:    [[SPLIT1:%.*]] = shufflevector <8 x i64> [[A]], <8 x i64> poison, <1 x i32> <i32 1>
----------------
thegameg wrote:
> Wonder which part made this go back, the cost of the mul maybe?
Yep, I think vector integer multiply doesn't support i64 elements on ARM64.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D147330/new/

https://reviews.llvm.org/D147330