[llvm] [VectorCombine] Try to scalarize vector loads feeding bitcast instructions. (PR #164682)

[Simon Pilgrim via llvm-commits]

================
@@ -0,0 +1,136 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -passes=vector-combine -mtriple=arm64-apple-darwinos -S %s | FileCheck %s
+
+define i32 @load_v4i8_bitcast_to_i32(ptr %x) {
+; CHECK-LABEL: define i32 @load_v4i8_bitcast_to_i32(
+; CHECK-SAME: ptr [[X:%.*]]) {
+; CHECK-NEXT:    [[R_SCALAR:%.*]] = load i32, ptr [[X]], align 4
+; CHECK-NEXT:    ret i32 [[R_SCALAR]]
+;
+  %lv = load <4 x i8>, ptr %x
+  %r = bitcast <4 x i8> %lv to i32
+  ret i32 %r
+}
+
+define i64 @load_v2i32_bitcast_to_i64(ptr %x) {
+; CHECK-LABEL: define i64 @load_v2i32_bitcast_to_i64(
+; CHECK-SAME: ptr [[X:%.*]]) {
+; CHECK-NEXT:    [[R_SCALAR:%.*]] = load i64, ptr [[X]], align 8
+; CHECK-NEXT:    ret i64 [[R_SCALAR]]
+;
+  %lv = load <2 x i32>, ptr %x
+  %r = bitcast <2 x i32> %lv to i64
+  ret i64 %r
+}
+
+define float @load_v4i8_bitcast_to_float(ptr %x) {
+; CHECK-LABEL: define float @load_v4i8_bitcast_to_float(
+; CHECK-SAME: ptr [[X:%.*]]) {
+; CHECK-NEXT:    [[R_SCALAR:%.*]] = load float, ptr [[X]], align 4
+; CHECK-NEXT:    ret float [[R_SCALAR]]
+;
+  %lv = load <4 x i8>, ptr %x
+  %r = bitcast <4 x i8> %lv to float
+  ret float %r
+}
+
+define float @load_v2i16_bitcast_to_float(ptr %x) {
+; CHECK-LABEL: define float @load_v2i16_bitcast_to_float(
+; CHECK-SAME: ptr [[X:%.*]]) {
+; CHECK-NEXT:    [[R_SCALAR:%.*]] = load float, ptr [[X]], align 4
+; CHECK-NEXT:    ret float [[R_SCALAR]]
+;
+  %lv = load <2 x i16>, ptr %x
+  %r = bitcast <2 x i16> %lv to float
+  ret float %r
+}
+
+define double @load_v4i16_bitcast_to_double(ptr %x) {
+; CHECK-LABEL: define double @load_v4i16_bitcast_to_double(
+; CHECK-SAME: ptr [[X:%.*]]) {
+; CHECK-NEXT:    [[LV:%.*]] = load <4 x i16>, ptr [[X]], align 8
+; CHECK-NEXT:    [[R_SCALAR:%.*]] = bitcast <4 x i16> [[LV]] to double
+; CHECK-NEXT:    ret double [[R_SCALAR]]
+;
+  %lv = load <4 x i16>, ptr %x
+  %r = bitcast <4 x i16> %lv to double
+  ret double %r
+}
+
+define double @load_v2i32_bitcast_to_double(ptr %x) {
+; CHECK-LABEL: define double @load_v2i32_bitcast_to_double(
+; CHECK-SAME: ptr [[X:%.*]]) {
+; CHECK-NEXT:    [[LV:%.*]] = load <2 x i32>, ptr [[X]], align 8
+; CHECK-NEXT:    [[R_SCALAR:%.*]] = bitcast <2 x i32> [[LV]] to double
+; CHECK-NEXT:    ret double [[R_SCALAR]]
+;
+  %lv = load <2 x i32>, ptr %x
+  %r = bitcast <2 x i32> %lv to double
+  ret double %r
+}
+
+; Multiple users with the same bitcast type should be scalarized.
+define i32 @load_v4i8_bitcast_multiple_users_same_type(ptr %x) {
+; CHECK-LABEL: define i32 @load_v4i8_bitcast_multiple_users_same_type(
+; CHECK-SAME: ptr [[X:%.*]]) {
+; CHECK-NEXT:    [[LV_SCALAR:%.*]] = load i32, ptr [[X]], align 4
+; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[LV_SCALAR]], [[LV_SCALAR]]
+; CHECK-NEXT:    ret i32 [[ADD]]
+;
+  %lv = load <4 x i8>, ptr %x
+  %r1 = bitcast <4 x i8> %lv to i32
+  %r2 = bitcast <4 x i8> %lv to i32
+  %add = add i32 %r1, %r2
+  ret i32 %add
+}
----------------
RKSimon wrote:

Aren't other passes likely to have already folded these duplication or are you seeing this kind of thing in real world code?

https://github.com/llvm/llvm-project/pull/164682