[llvm] [VPlan] Add simple driver option to run some individual transforms. (PR #178522)

[Andrei Elovikov via llvm-commits]

================
@@ -0,0 +1,74 @@
+; REQUIRES: asserts
+
+; Test the widen-from-metadata VPlan transform that converts VPInstructions to
+; recipes based on !vplan.widen metadata.
+
+; RUN: opt -passes=loop-vectorize -vplan-test-transform='widen-from-metadata,print' -disable-output %s 2>&1 | FileCheck %s --check-prefix=WIDEN
+; RUN: opt -passes=loop-vectorize -vplan-test-transform='widen-from-metadata,print' -disable-output %s 2>&1 | FileCheck %s --check-prefix=REPLICATE
+
+; Test widen conversion: VPInstruction with !vplan.widen !{!"widen"}
+; should be converted to VPWidenRecipe (WIDEN instead of EMIT).
+; WIDEN-LABEL: VPlan ' for UF>=1' {
+; WIDEN: <x1> vector loop: {
+; WIDEN:   vector.body:
+; WIDEN:     EMIT vp<[[CAN_IV:%.+]]> = CANONICAL-INDUCTION
+; WIDEN:     EMIT ir<%gep> = getelementptr
+; WIDEN:     EMIT ir<%val> = load ir<%gep>
+; WIDEN:     WIDEN ir<%add> = add ir<%val>, ir<1>
+; WIDEN:     EMIT store ir<%add>, ir<%gep>
+; WIDEN:     EMIT branch-on-count
+; WIDEN:   No successors
+; WIDEN: }
+
+; Test replicate conversion: VPInstruction with !vplan.widen !{!"replicate"}
+; should be converted to VPReplicateRecipe (REPLICATE instead of EMIT).
+; REPLICATE-LABEL: VPlan ' for UF>=1' {
+; REPLICATE: <x1> vector loop: {
+; REPLICATE:   vector.body:
+; REPLICATE:     EMIT vp<[[CAN_IV:%.+]]> = CANONICAL-INDUCTION
+; REPLICATE:     EMIT ir<%gep> = getelementptr
+; REPLICATE:     EMIT ir<%val> = load ir<%gep>
+; REPLICATE:     REPLICATE ir<%add> = add ir<%val>, ir<1>
+; REPLICATE:     EMIT store ir<%add>, ir<%gep>
+; REPLICATE:     EMIT branch-on-count
+; REPLICATE:   No successors
+; REPLICATE: }
+
+define void @test_widen(ptr %A, i64 %N) {
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %gep = getelementptr i32, ptr %A, i64 %iv
+  %val = load i32, ptr %gep
+  %add = add i32 %val, 1, !vplan.widen !0
+  store i32 %add, ptr %gep
+  %iv.next = add i64 %iv, 1
+  %cmp = icmp slt i64 %iv.next, %N
+  br i1 %cmp, label %loop, label %exit
+
+exit:
+  ret void
+}
+
+define void @test_replicate(ptr %A, i64 %N) {
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %gep = getelementptr i32, ptr %A, i64 %iv
+  %val = load i32, ptr %gep
+  %add = add i32 %val, 1, !vplan.widen !1
----------------
eas wrote:

This might be confusing to read. I think it better be either `!vplan.replicate !1` or `!vplan.explicit.recipe !1` I think the former would be more readable.

https://github.com/llvm/llvm-project/pull/178522