[llvm] [Matrix] Propagate shape information through cast insts (PR #141869)

[Florian Hahn via llvm-commits]

================
@@ -0,0 +1,202 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -passes='lower-matrix-intrinsics' -S < %s | FileCheck %s
+
+define void @fneg_2x2(ptr %in, ptr %out) {
+; CHECK-LABEL: @fneg_2x2(
+; CHECK-NEXT:    [[COL_LOAD:%.*]] = load <2 x float>, ptr [[IN:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr float, ptr [[IN]], i64 2
+; CHECK-NEXT:    [[COL_LOAD1:%.*]] = load <2 x float>, ptr [[VEC_GEP]], align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg <2 x float> [[COL_LOAD]]
+; CHECK-NEXT:    [[TMP2:%.*]] = fneg <2 x float> [[COL_LOAD1]]
+; CHECK-NEXT:    store <2 x float> [[TMP1]], ptr [[OUT:%.*]], align 16
+; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr float, ptr [[OUT]], i64 2
+; CHECK-NEXT:    store <2 x float> [[TMP2]], ptr [[VEC_GEP2]], align 8
+; CHECK-NEXT:    ret void
+;
+  %inv = load <4 x float>, ptr %in
+  %op = fneg <4 x float> %inv
+  %opt  = call <4 x float> @llvm.matrix.transpose(<4 x float> %op, i32 2, i32 2)
----------------
fhahn wrote:

Could we use a single `matrix.column.major.load` to get some matrix type info instead of 2 transposes?

Using transpose(transpose()) is a neat trick, but if you don't know the trick it might be a bit surprising to see here ;)

https://github.com/llvm/llvm-project/pull/141869