[PATCH] D86364: [ValueTracking] Interpret GEPs as a series of adds multiplied by the related scaling factor

[Quentin Colombet via Phabricator via llvm-commits]

qcolombet added inline comments.


================
Comment at: llvm/test/Transforms/InstCombine/constant-fold-gep.ll:47
+; CHECK: store i32 1, i32* getelementptr inbounds ([3 x %struct.X], [3 x %struct.X]* @Y, i64 0, i64 2, i32 1, i64 2), align 4
+  store i32 1, i32* getelementptr ([3 x %struct.X], [3 x %struct.X]* @Y, i64 0, i64 0, i32 0, i64 17), align 4
 ; CHECK: store i32 1, i32* getelementptr inbounds ([3 x %struct.X], [3 x %struct.X]* @Y, i64 1, i64 0, i32 0, i64 0), align 8
----------------
Note: The change in the input of the test is because the alignment was not consistent with the base pointer.
On this store the alignment cannot be 8, since the base pointer is aligned on 16 (17 * 4 + 16 == 17 * 4 + 4 * 4 == 21 * 4 == 10 * 8 + 4 i.e., remainder with 8 cannot be zero). More easily seen looking at the previous load: @Y + 16 is aligned to 8 then we add 4 to make @Y + 17, and we claimed it was aligned to 8.


================
Comment at: llvm/test/Transforms/InstCombine/constant-fold-gep.ll:107
+  %res = load i32, i32* %NS7, align 1
+  ret i32 %res
+}
----------------
Test case for the structure access bug found during pretesting.
With the offset calculation bug of the previous version of the patch we would have got `align 4`.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D86364/new/

https://reviews.llvm.org/D86364