[llvm] [VectorCombine] Align shrinkLoadForShuffles to register boundaries (PR #209775)

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Tue Jul 21 07:34:19 PDT 2026


================
@@ -252,6 +252,25 @@ define <2 x i4> @vec_with_2elts_128bits_i4(ptr align 16 dereferenceable(16) %p)
   ret <2 x i4> %r
 }
 
+; Shrinking this to <9 x double> would change the legalized type.
+
+define <8 x double> @load_v16f64(ptr %p) {
+; SSE-LABEL: @load_v16f64(
+; SSE-NEXT:    [[TMP1:%.*]] = load <10 x double>, ptr [[P:%.*]], align 512
+; SSE-NEXT:    [[S:%.*]] = shufflevector <10 x double> [[TMP1]], <10 x double> poison, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>
+; SSE-NEXT:    ret <8 x double> [[S]]
+;
+; AVX-LABEL: @load_v16f64(
+; AVX-NEXT:    [[TMP1:%.*]] = load <12 x double>, ptr [[P:%.*]], align 512
+; AVX-NEXT:    [[S:%.*]] = shufflevector <12 x double> [[TMP1]], <12 x double> poison, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>
+; AVX-NEXT:    ret <8 x double> [[S]]
+;
+  %l = load <16 x double>, ptr %p, align 512
+  %s = shufflevector <16 x double> %l, <16 x double> poison, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>
+  ret <8 x double> %s
+}
+
----------------
arsenm wrote:

Test a pointer case

https://github.com/llvm/llvm-project/pull/209775


More information about the llvm-commits mailing list