[llvm] [LV][NFC] Move and add truncated-related FindLastIV reduction test cases. (PR #67674)
Mel Chen via llvm-commits
llvm-commits at lists.llvm.org
Fri Sep 29 02:02:15 PDT 2023
https://github.com/Mel-Chen updated https://github.com/llvm/llvm-project/pull/67674
>From dc69da35d0cd2fa3042408bdf5b5bd8c04d187a0 Mon Sep 17 00:00:00 2001
From: Mel Chen <mel.chen at sifive.com>
Date: Wed, 27 Sep 2023 21:12:23 -0700
Subject: [PATCH 1/2] [LV][NFC] Move and add truncated-related FindLastIV
reduction test cases.
---
.../LoopVectorize/iv-select-cmp-trunc.ll | 274 ++++++++++++++++++
.../Transforms/LoopVectorize/iv-select-cmp.ll | 166 -----------
2 files changed, 274 insertions(+), 166 deletions(-)
create mode 100644 llvm/test/Transforms/LoopVectorize/iv-select-cmp-trunc.ll
diff --git a/llvm/test/Transforms/LoopVectorize/iv-select-cmp-trunc.ll b/llvm/test/Transforms/LoopVectorize/iv-select-cmp-trunc.ll
new file mode 100644
index 000000000000000..f2d39d7a402fb29
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/iv-select-cmp-trunc.ll
@@ -0,0 +1,274 @@
+; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -S < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -passes=loop-vectorize -force-vector-interleave=4 -force-vector-width=4 -S < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -passes=loop-vectorize -force-vector-interleave=4 -force-vector-width=1 -S < %s | FileCheck %s --check-prefix=CHECK
+
+; This test can theoretically be vectorized without a runtime-check, by
+; pattern-matching on the constructs that are introduced by IndVarSimplify.
+; We can check two things:
+; %1 = trunc i64 %iv to i32
+; This indicates that the %iv is truncated to i32. We can then check the loop
+; guard is a signed i32:
+; %cmp.sgt = icmp sgt i32 %n, 0
+; and successfully vectorize the case without a runtime-check.
+define i32 @select_icmp_const_truncated_iv_widened_exit(ptr %a, i32 %n) {
+; CHECK-LABEL: define i32 @select_icmp_const_truncated_iv_widened_exit
+; CHECK-NOT: vector.body:
+;
+entry:
+ %cmp.sgt = icmp sgt i32 %n, 0
+ br i1 %cmp.sgt, label %for.body.preheader, label %exit
+
+for.body.preheader: ; preds = %entry
+ %wide.trip.count = zext i32 %n to i64
+ br label %for.body
+
+for.body: ; preds = %for.body.preheader, %for.body
+ %iv = phi i64 [ 0, %for.body.preheader ], [ %inc, %for.body ]
+ %rdx = phi i32 [ 331, %for.body.preheader ], [ %spec.select, %for.body ]
+ %arrayidx = getelementptr inbounds i64, ptr %a, i64 %iv
+ %0 = load i64, ptr %arrayidx, align 8
+ %cmp = icmp sgt i64 %0, 3
+ %1 = trunc i64 %iv to i32
+ %spec.select = select i1 %cmp, i32 %1, i32 %rdx
+ %inc = add nuw nsw i64 %iv, 1
+ %exitcond.not = icmp eq i64 %inc, %wide.trip.count
+ br i1 %exitcond.not, label %exit, label %for.body
+
+exit: ; preds = %for.body, %entry
+ %rdx.lcssa = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
+ ret i32 %rdx.lcssa
+}
+
+; This test can theoretically be vectorized without a runtime-check, by
+; pattern-matching on the constructs that are introduced by IndVarSimplify.
+; We can check two things:
+; %1 = trunc i64 %iv to i32
+; This indicates that the %iv is truncated to i32. We can then check the loop
+; exit condition, which compares to a constant that fits within i32:
+; %exitcond.not = icmp eq i64 %inc, 20000
+; and successfully vectorize the case without a runtime-check.
+define i32 @select_icmp_const_truncated_iv_const_exit(ptr %a) {
+; CHECK-LABEL: define i32 @select_icmp_const_truncated_iv_const_exit
+; CHECK-NOT: vector.body:
+;
+entry:
+ br label %for.body
+
+for.body: ; preds = %entry, %for.body
+ %iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+ %rdx = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
+ %arrayidx = getelementptr inbounds i64, ptr %a, i64 %iv
+ %0 = load i64, ptr %arrayidx, align 8
+ %cmp = icmp sgt i64 %0, 3
+ %1 = trunc i64 %iv to i32
+ %spec.select = select i1 %cmp, i32 %1, i32 %rdx
+ %inc = add nuw nsw i64 %iv, 1
+ %exitcond.not = icmp eq i64 %inc, 20000
+ br i1 %exitcond.not, label %exit, label %for.body
+
+exit: ; preds = %for.body
+ ret i32 %spec.select
+}
+
+define i32 @select_fcmp_max_valid_const_ub(ptr %a) {
+; CHECK-LABEL: define i32 @select_fcmp_max_valid_const_ub
+; CHECK-NOT: vector.body:
+;
+entry:
+ br label %for.body
+
+for.body: ; preds = %entry, %for.body
+ %iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+ %rdx = phi i32 [ -1, %entry ], [ %spec.select, %for.body ]
+ %arrayidx = getelementptr inbounds float, ptr %a, i64 %iv
+ %0 = load float, ptr %arrayidx, align 4
+ %cmp = fcmp fast olt float %0, 0.000000e+00
+ %1 = trunc i64 %iv to i32
+ %spec.select = select i1 %cmp, i32 %1, i32 %rdx
+ %inc = add nuw nsw i64 %iv, 1
+ %exitcond.not = icmp eq i64 %inc, 2147483648
+ br i1 %exitcond.not, label %exit, label %for.body
+
+exit: ; preds = %for.body
+ ret i32 %spec.select
+}
+
+; Negative tests
+
+; This test can theoretically be vectorized, but only with a runtime-check.
+; The construct that are introduced by IndVarSimplify is:
+; %1 = trunc i64 %iv to i32
+; However, the loop guard is an i64:
+; %cmp.sgt = icmp sgt i64 %n, 0
+; We cannot guarantee that %iv won't overflow an i32 value (and hence hit the
+; sentinel value), and need a runtime-check to vectorize this case.
+define i32 @not_vectorized_select_icmp_const_truncated_iv_unwidened_exit(ptr %a, i64 %n) {
+; CHECK-LABEL: define i32 @not_vectorized_select_icmp_const_truncated_iv_unwidened_exit
+; CHECK-NOT: vector.body:
+;
+entry:
+ %cmp.sgt = icmp sgt i64 %n, 0
+ br i1 %cmp.sgt, label %for.body, label %exit
+
+for.body: ; preds = %entry, %for.body
+ %iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+ %rdx = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
+ %arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
+ %0 = load i32, ptr %arrayidx, align 4
+ %cmp = icmp sgt i32 %0, 3
+ %1 = trunc i64 %iv to i32
+ %spec.select = select i1 %cmp, i32 %1, i32 %rdx
+ %inc = add nuw nsw i64 %iv, 1
+ %exitcond.not = icmp eq i64 %inc, %n
+ br i1 %exitcond.not, label %exit, label %for.body
+
+exit: ; preds = %for.body, %entry
+ %rdx.lcssa = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
+ ret i32 %rdx.lcssa
+}
+
+; This test can theoretically be vectorized, but only with a runtime-check.
+; The construct that are introduced by IndVarSimplify is:
+; %1 = trunc i64 %iv to i32
+; However, the loop guard is unsigned:
+; %cmp.not = icmp eq i32 %n, 0
+; We cannot guarantee that %iv won't overflow an i32 value (and hence hit the
+; sentinel value), and need a runtime-check to vectorize this case.
+define i32 @not_vectorized_select_icmp_const_truncated_iv_unsigned_loop_guard(ptr %a, i32 %n) {
+; CHECK-LABEL: define i32 @not_vectorized_select_icmp_const_truncated_iv_unsigned_loop_guard
+; CHECK-NOT: vector.body:
+;
+entry:
+ %cmp.not = icmp eq i32 %n, 0
+ br i1 %cmp.not, label %exit, label %for.body.preheader
+
+for.body.preheader: ; preds = %entry
+ %wide.trip.count = zext i32 %n to i64
+ br label %for.body
+
+for.body: ; preds = %for.body.preheader, %for.body
+ %iv = phi i64 [ 0, %for.body.preheader ], [ %inc, %for.body ]
+ %rdx = phi i32 [ 331, %for.body.preheader ], [ %spec.select, %for.body ]
+ %arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
+ %0 = load i32, ptr %arrayidx, align 4
+ %cmp1 = icmp sgt i32 %0, 3
+ %1 = trunc i64 %iv to i32
+ %spec.select = select i1 %cmp1, i32 %1, i32 %rdx
+ %inc = add nuw nsw i64 %iv, 1
+ %exitcond.not = icmp eq i64 %inc, %wide.trip.count
+ br i1 %exitcond.not, label %exit, label %for.body
+
+exit: ; preds = %for.body, %entry
+ %rdx.lcssa = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
+ ret i32 %rdx.lcssa
+}
+
+; This test cannot be vectorized, even with a runtime check.
+; The construct that are introduced by IndVarSimplify is:
+; %1 = trunc i64 %iv to i32
+; However, the loop exit condition is a constant that overflows i32:
+; %exitcond.not = icmp eq i64 %inc, 4294967294
+; Hence, the i32 will most certainly wrap and hit the sentinel value, and we
+; cannot vectorize this case.
+define i32 @not_vectorized_select_icmp_truncated_iv_out_of_bound(ptr %a) {
+; CHECK-LABEL: define i32 @not_vectorized_select_icmp_truncated_iv_out_of_bound
+; CHECK-NOT: vector.body:
+;
+entry:
+ br label %for.body
+
+for.body: ; preds = %entry, %for.body
+ %iv = phi i64 [ 2147483646, %entry ], [ %inc, %for.body ]
+ %rdx = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
+ %arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
+ %0 = load i32, ptr %arrayidx, align 4
+ %cmp = icmp sgt i32 %0, 3
+ %conv = trunc i64 %iv to i32
+ %spec.select = select i1 %cmp, i32 %conv, i32 %rdx
+ %inc = add nuw nsw i64 %iv, 1
+ %exitcond.not = icmp eq i64 %inc, 4294967294
+ br i1 %exitcond.not, label %exit, label %for.body
+
+exit: ; preds = %for.body
+ ret i32 %spec.select
+}
+
+define i32 @not_vectorized_select_iv_icmp_no_guard(ptr %a, ptr %b, i32 %start, i32 %n) {
+; CHECK-LABEL: define i32 @not_vectorized_select_iv_icmp_no_guard
+; CHECK-NOT: vector.body:
+;
+entry:
+ %wide.trip.count = zext i32 %n to i64
+ br label %for.body
+
+for.body: ; preds = %entry, %for.body
+ %iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+ %rdx = phi i32 [ %start, %entry ], [ %cond, %for.body ]
+ %arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
+ %0 = load i32, ptr %arrayidx, align 4
+ %arrayidx2 = getelementptr inbounds i32, ptr %b, i64 %iv
+ %1 = load i32, ptr %arrayidx2, align 4
+ %cmp = icmp sgt i32 %0, %1
+ %2 = trunc i64 %iv to i32
+ %cond = select i1 %cmp, i32 %2, i32 %rdx
+ %inc = add nuw nsw i64 %iv, 1
+ %exitcond.not = icmp eq i64 %inc, %wide.trip.count
+ br i1 %exitcond.not, label %exit, label %for.body
+
+exit: ; preds = %for.body
+ ret i32 %cond
+}
+
+define i32 @not_vectorized_select_fcmp_invalid_const_ub(ptr %a) {
+; CHECK-LABEL: define i32 @not_vectorized_select_fcmp_invalid_const_ub
+; CHECK-NOT: vector.body:
+;
+entry:
+ br label %for.body
+
+for.body: ; preds = %entry, %for.body
+ %iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+ %rdx = phi i32 [ -1, %entry ], [ %spec.select, %for.body ]
+ %arrayidx = getelementptr inbounds float, ptr %a, i64 %iv
+ %0 = load float, ptr %arrayidx, align 4
+ %cmp = fcmp fast olt float %0, 0.000000e+00
+ %1 = trunc i64 %iv to i32
+ %spec.select = select i1 %cmp, i32 %1, i32 %rdx
+ %inc = add nuw nsw i64 %iv, 1
+ %exitcond.not = icmp eq i64 %inc, 2147483649
+ br i1 %exitcond.not, label %exit, label %for.body
+
+exit: ; preds = %for.body
+ ret i32 %spec.select
+}
+
+define i16 @not_vectorized_select_iv_icmp_overflow_unwidened_tripcount(ptr %a, ptr %b, i16 %start, i32 %n) {
+; CHECK-LABEL: define i16 @not_vectorized_select_iv_icmp_overflow_unwidened_tripcount
+; CHECK-NOT: vector.body:
+;
+entry:
+ %cmp9 = icmp sgt i32 %n, 0
+ br i1 %cmp9, label %for.body.preheader, label %exit
+
+for.body.preheader: ; preds = %entry
+ %wide.trip.count = zext i32 %n to i64
+ br label %for.body
+
+for.body: ; preds = %for.body.preheader, %for.body
+ %iv = phi i64 [ 0, %for.body.preheader ], [ %inc, %for.body ]
+ %rdx = phi i16 [ %start, %for.body.preheader ], [ %cond, %for.body ]
+ %arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
+ %0 = load i32, ptr %arrayidx, align 4
+ %arrayidx2 = getelementptr inbounds i32, ptr %b, i64 %iv
+ %1 = load i32, ptr %arrayidx2, align 4
+ %cmp3 = icmp sgt i32 %0, %1
+ %2 = trunc i64 %iv to i16
+ %cond = select i1 %cmp3, i16 %2, i16 %rdx
+ %inc = add nuw nsw i64 %iv, 1
+ %exitcond.not = icmp eq i64 %inc, %wide.trip.count
+ br i1 %exitcond.not, label %exit, label %for.body
+
+exit: ; preds = %for.body, %entry
+ %rdx.0.lcssa = phi i16 [ %start, %entry ], [ %cond, %for.body ]
+ ret i16 %rdx.0.lcssa
+}
diff --git a/llvm/test/Transforms/LoopVectorize/iv-select-cmp.ll b/llvm/test/Transforms/LoopVectorize/iv-select-cmp.ll
index f69e50ab74442bd..6108dbd0b191e67 100644
--- a/llvm/test/Transforms/LoopVectorize/iv-select-cmp.ll
+++ b/llvm/test/Transforms/LoopVectorize/iv-select-cmp.ll
@@ -188,172 +188,6 @@ exit: ; preds = %for.body
; Negative tests
-; This test can theoretically be vectorized without a runtime-check, by
-; pattern-matching on the constructs that are introduced by IndVarSimplify.
-; We can check two things:
-; %1 = trunc i64 %iv to i32
-; This indicates that the %iv is truncated to i32. We can then check the loop
-; guard is a signed i32:
-; %cmp.sgt = icmp sgt i32 %n, 0
-; and successfully vectorize the case without a runtime-check.
-define i32 @not_vectorized_select_icmp_const_truncated_iv_widened_exit(ptr %a, i32 %n) {
-; CHECK-LABEL: define i32 @not_vectorized_select_icmp_const_truncated_iv_widened_exit
-; CHECK-NOT: vector.body:
-;
-entry:
- %cmp.sgt = icmp sgt i32 %n, 0
- br i1 %cmp.sgt, label %for.body.preheader, label %exit
-
-for.body.preheader: ; preds = %entry
- %wide.trip.count = zext i32 %n to i64
- br label %for.body
-
-for.body: ; preds = %for.body.preheader, %for.body
- %iv = phi i64 [ 0, %for.body.preheader ], [ %inc, %for.body ]
- %rdx = phi i32 [ 331, %for.body.preheader ], [ %spec.select, %for.body ]
- %arrayidx = getelementptr inbounds i64, ptr %a, i64 %iv
- %0 = load i64, ptr %arrayidx, align 8
- %cmp = icmp sgt i64 %0, 3
- %1 = trunc i64 %iv to i32
- %spec.select = select i1 %cmp, i32 %1, i32 %rdx
- %inc = add nuw nsw i64 %iv, 1
- %exitcond.not = icmp eq i64 %inc, %wide.trip.count
- br i1 %exitcond.not, label %exit, label %for.body
-
-exit: ; preds = %for.body, %entry
- %rdx.lcssa = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
- ret i32 %rdx.lcssa
-}
-
-; This test can theoretically be vectorized without a runtime-check, by
-; pattern-matching on the constructs that are introduced by IndVarSimplify.
-; We can check two things:
-; %1 = trunc i64 %iv to i32
-; This indicates that the %iv is truncated to i32. We can then check the loop
-; exit condition, which compares to a constant that fits within i32:
-; %exitcond.not = icmp eq i64 %inc, 20000
-; and successfully vectorize the case without a runtime-check.
-define i32 @not_vectorized_select_icmp_const_truncated_iv_const_exit(ptr %a) {
-; CHECK-LABEL: define i32 @not_vectorized_select_icmp_const_truncated_iv_const_exit
-; CHECK-NOT: vector.body:
-;
-entry:
- br label %for.body
-
-for.body: ; preds = %entry, %for.body
- %iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
- %rdx = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
- %arrayidx = getelementptr inbounds i64, ptr %a, i64 %iv
- %0 = load i64, ptr %arrayidx, align 8
- %cmp = icmp sgt i64 %0, 3
- %1 = trunc i64 %iv to i32
- %spec.select = select i1 %cmp, i32 %1, i32 %rdx
- %inc = add nuw nsw i64 %iv, 1
- %exitcond.not = icmp eq i64 %inc, 20000
- br i1 %exitcond.not, label %exit, label %for.body
-
-exit: ; preds = %for.body
- ret i32 %spec.select
-}
-
-; This test can theoretically be vectorized, but only with a runtime-check.
-; The construct that are introduced by IndVarSimplify is:
-; %1 = trunc i64 %iv to i32
-; However, the loop guard is an i64:
-; %cmp.sgt = icmp sgt i64 %n, 0
-; We cannot guarantee that %iv won't overflow an i32 value (and hence hit the
-; sentinel value), and need a runtime-check to vectorize this case.
-define i32 @not_vectorized_select_icmp_const_truncated_iv_unwidened_exit(ptr %a, i64 %n) {
-; CHECK-LABEL: define i32 @not_vectorized_select_icmp_const_truncated_iv_unwidened_exit
-; CHECK-NOT: vector.body:
-;
-entry:
- %cmp.sgt = icmp sgt i64 %n, 0
- br i1 %cmp.sgt, label %for.body, label %exit
-
-for.body: ; preds = %entry, %for.body
- %iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
- %rdx = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
- %arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
- %0 = load i32, ptr %arrayidx, align 4
- %cmp = icmp sgt i32 %0, 3
- %1 = trunc i64 %iv to i32
- %spec.select = select i1 %cmp, i32 %1, i32 %rdx
- %inc = add nuw nsw i64 %iv, 1
- %exitcond.not = icmp eq i64 %inc, %n
- br i1 %exitcond.not, label %exit, label %for.body
-
-exit: ; preds = %for.body, %entry
- %rdx.lcssa = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
- ret i32 %rdx.lcssa
-}
-
-; This test can theoretically be vectorized, but only with a runtime-check.
-; The construct that are introduced by IndVarSimplify is:
-; %1 = trunc i64 %iv to i32
-; However, the loop guard is unsigned:
-; %cmp.not = icmp eq i32 %n, 0
-; We cannot guarantee that %iv won't overflow an i32 value (and hence hit the
-; sentinel value), and need a runtime-check to vectorize this case.
-define i32 @not_vectorized_select_icmp_const_truncated_iv_unsigned_loop_guard(ptr %a, i32 %n) {
-; CHECK-LABEL: define i32 @not_vectorized_select_icmp_const_truncated_iv_unsigned_loop_guard
-; CHECK-NOT: vector.body:
-;
-entry:
- %cmp.not = icmp eq i32 %n, 0
- br i1 %cmp.not, label %exit, label %for.body.preheader
-
-for.body.preheader: ; preds = %entry
- %wide.trip.count = zext i32 %n to i64
- br label %for.body
-
-for.body: ; preds = %for.body.preheader, %for.body
- %iv = phi i64 [ 0, %for.body.preheader ], [ %inc, %for.body ]
- %rdx = phi i32 [ 331, %for.body.preheader ], [ %spec.select, %for.body ]
- %arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
- %0 = load i32, ptr %arrayidx, align 4
- %cmp1 = icmp sgt i32 %0, 3
- %1 = trunc i64 %iv to i32
- %spec.select = select i1 %cmp1, i32 %1, i32 %rdx
- %inc = add nuw nsw i64 %iv, 1
- %exitcond.not = icmp eq i64 %inc, %wide.trip.count
- br i1 %exitcond.not, label %exit, label %for.body
-
-exit: ; preds = %for.body, %entry
- %rdx.lcssa = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
- ret i32 %rdx.lcssa
-}
-
-; This test cannot be vectorized, even with a runtime check.
-; The construct that are introduced by IndVarSimplify is:
-; %1 = trunc i64 %iv to i32
-; However, the loop exit condition is a constant that overflows i32:
-; %exitcond.not = icmp eq i64 %inc, 4294967294
-; Hence, the i32 will most certainly wrap and hit the sentinel value, and we
-; cannot vectorize this case.
-define i32 @not_vectorized_select_icmp_truncated_iv_out_of_bound(ptr %a) {
-; CHECK-LABEL: define i32 @not_vectorized_select_icmp_truncated_iv_out_of_bound
-; CHECK-NOT: vector.body:
-;
-entry:
- br label %for.body
-
-for.body: ; preds = %entry, %for.body
- %iv = phi i64 [ 2147483646, %entry ], [ %inc, %for.body ]
- %rdx = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
- %arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
- %0 = load i32, ptr %arrayidx, align 4
- %cmp = icmp sgt i32 %0, 3
- %conv = trunc i64 %iv to i32
- %spec.select = select i1 %cmp, i32 %conv, i32 %rdx
- %inc = add nuw nsw i64 %iv, 1
- %exitcond.not = icmp eq i64 %inc, 4294967294
- br i1 %exitcond.not, label %exit, label %for.body
-
-exit: ; preds = %for.body
- ret i32 %spec.select
-}
-
define float @not_vectorized_select_float_induction_icmp(ptr %a, ptr %b, float %rdx.start, i64 %n) {
; CHECK-LABEL: @not_vectorized_select_float_induction_icmp
; CHECK-NOT: vector.body:
>From f1eac6eb631c4cce3bdfa986759206ad8e63a0af Mon Sep 17 00:00:00 2001
From: Mel Chen <mel.chen at sifive.com>
Date: Fri, 29 Sep 2023 02:01:51 -0700
Subject: [PATCH 2/2] Add the comment for test cases.
---
.../Transforms/LoopVectorize/iv-select-cmp-trunc.ll | 11 +++++++++++
1 file changed, 11 insertions(+)
diff --git a/llvm/test/Transforms/LoopVectorize/iv-select-cmp-trunc.ll b/llvm/test/Transforms/LoopVectorize/iv-select-cmp-trunc.ll
index f2d39d7a402fb29..b94e9f99868efdd 100644
--- a/llvm/test/Transforms/LoopVectorize/iv-select-cmp-trunc.ll
+++ b/llvm/test/Transforms/LoopVectorize/iv-select-cmp-trunc.ll
@@ -70,6 +70,8 @@ exit: ; preds = %for.body
ret i32 %spec.select
}
+; Without loop guard, the maximum constant trip count that can be vectorized is
+; the signed maximum value of reduction type.
define i32 @select_fcmp_max_valid_const_ub(ptr %a) {
; CHECK-LABEL: define i32 @select_fcmp_max_valid_const_ub
; CHECK-NOT: vector.body:
@@ -193,6 +195,8 @@ exit: ; preds = %for.body
ret i32 %spec.select
}
+; Forbidding vectorization of the FindLastIV pattern involving a truncated
+; induction variable in the absence of any loop guard.
define i32 @not_vectorized_select_iv_icmp_no_guard(ptr %a, ptr %b, i32 %start, i32 %n) {
; CHECK-LABEL: define i32 @not_vectorized_select_iv_icmp_no_guard
; CHECK-NOT: vector.body:
@@ -219,6 +223,10 @@ exit: ; preds = %for.body
ret i32 %cond
}
+; Without loop guard, when the constant trip count exceeds the maximum signed
+; value of the reduction type, truncation may cause overflow. Therefore,
+; vectorizer is unable to guarantee that the induction variable is monotonic
+; increasing.
define i32 @not_vectorized_select_fcmp_invalid_const_ub(ptr %a) {
; CHECK-LABEL: define i32 @not_vectorized_select_fcmp_invalid_const_ub
; CHECK-NOT: vector.body:
@@ -242,6 +250,9 @@ exit: ; preds = %for.body
ret i32 %spec.select
}
+; Even with loop guard protection, if the destination type of the truncation
+; instruction is smaller than the trip count type before extension, overflow
+; could still occur.
define i16 @not_vectorized_select_iv_icmp_overflow_unwidened_tripcount(ptr %a, ptr %b, i16 %start, i32 %n) {
; CHECK-LABEL: define i16 @not_vectorized_select_iv_icmp_overflow_unwidened_tripcount
; CHECK-NOT: vector.body:
More information about the llvm-commits
mailing list