[llvm] 04b1276 - LoopVectorize/iv-select-cmp: add tests for truncated IV
Ramkumar Ramachandra via llvm-commits
llvm-commits at lists.llvm.org
Wed Aug 30 05:10:22 PDT 2023
Author: Ramkumar Ramachandra
Date: 2023-08-30T13:09:37+01:00
New Revision: 04b1276ad3b8976241228be8a966b1557f63492f
URL: https://github.com/llvm/llvm-project/commit/04b1276ad3b8976241228be8a966b1557f63492f
DIFF: https://github.com/llvm/llvm-project/commit/04b1276ad3b8976241228be8a966b1557f63492f.diff
LOG: LoopVectorize/iv-select-cmp: add tests for truncated IV
The current tests in iv-select-cmp.ll are not representative of clang
output of common real-world C programs, which are often written with i32
induction vars, as opposed to i64 induction vars. Hence, add five tests
corresponding to the following programs:
int test(int *a, int n) {
int rdx = 331;
for (int i = 0; i < n; i++) {
if (a[i] > 3)
rdx = i;
}
return rdx;
}
int test(int *a) {
int rdx = 331;
for (int i = 0; i < 20000; i++) {
if (a[i] > 3)
rdx = i;
}
return rdx;
}
int test(int *a, long n) {
int rdx = 331;
for (int i = 0; i < n; i++) {
if (a[i] > 3)
rdx = i;
}
return rdx;
}
int test(int *a, unsigned n) {
int rdx = 331;
for (int i = 0; i < n; i++) {
if (a[i] > 3)
rdx = i;
}
return rdx;
}
int test(int *a) {
int rdx = 331;
for (long i = INT_MIN - 1; i < UINT_MAX; i++) {
if (a[i] > 3)
rdx = i;
}
return rdx;
}
The first two can theoretically be vectorized without a runtime-check,
while the third and fourth cannot. The fifth cannot be vectorized, even
with a runtime-check.
This issue was found while reviewing D150851.
Differential Revision: https://reviews.llvm.org/D156124
Added:
Modified:
llvm/test/Transforms/LoopVectorize/iv-select-cmp.ll
Removed:
################################################################################
diff --git a/llvm/test/Transforms/LoopVectorize/iv-select-cmp.ll b/llvm/test/Transforms/LoopVectorize/iv-select-cmp.ll
index cd176a3c949225..045c27b8f7e053 100644
--- a/llvm/test/Transforms/LoopVectorize/iv-select-cmp.ll
+++ b/llvm/test/Transforms/LoopVectorize/iv-select-cmp.ll
@@ -188,6 +188,172 @@ 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 nocapture readonly %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 nocapture readonly %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 nocapture readonly %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 nocapture readonly %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 nocapture readonly %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 nocapture readonly %a, ptr nocapture readonly %b, float %rdx.start, i64 %n) {
; CHECK-LABEL: @not_vectorized_select_float_induction_icmp
; CHECK-NOT: vector.body:
More information about the llvm-commits
mailing list