Add tests for full unroll heuristic: folding CFG, folding IVs.

[mzolotukhin]

Summary

These are the tests for D10205, D10206, and D10207.

[mzolotukhin]

• Remove a leftover from test.

[mzolotukhin]

• Add a test for visitICmp fix.
• Add a test for visitCast fix.

[mzolotukhin]

• Use new option names.
• Adjust tests according to the new naming scheme.
• Add a test to exercise simplifyUsingOffsets.

[mzolotukhin]

Ping!

[mzolotukhin]

Ping!

[mzolotukhin]

Ping!

[Gerolf]

test/Transforms/LoopUnroll/full-unroll-heuristics-cfg.ll

6

Please comment in more detail which cases can be simplified. Are there scenarios with constant loads that won't get simplified?

69

Tthis...

100

1024 (=2^10).
it's -> it is

[chandlerc]

I agree that these tests can use some more comments to clearly indicate what you expect.

Also, is it possible to isolate a test for the cast folding, put it in the patch with the cast folding? I think then that one would be ready to land. And the CFG tests I'd rather have as part of the CFG folding patch rather than separate.

test/Transforms/LoopUnroll/full-unroll-heuristics-cfg.ll

+; RUN: opt < %s -S -loop-unroll -unroll-max-iteration-count-to-analyze=100 -unroll-dynamic-cost-savings-discount=1000 -unroll-threshold=10 -unroll-percent-dynamic-cost-saved-threshold=50 | FileCheck %s
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+
+@known_constant = internal unnamed_addr constant [10 x i32] [i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1], align 16
+
+; If a load becomes a constant after loop unrolling, we sometimes can simplify
+; CFG. This test verifies that we handle such cases.
+; CHECK-LABEL: @branch
+; CHECK-NOT:   br i1 %exitcond, label %for.end, label %for.body
+define i32 @branch(i32* noalias nocapture readonly %b) {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.inc, %entry
+  %iv.0 = phi i64 [ 0, %entry ], [ %iv.1, %for.inc ]
+  %r.0 = phi i32 [ 0, %entry ], [ %r.1, %for.inc ]
+  %arrayidx1 = getelementptr inbounds [10 x i32], [10 x i32]* @known_constant, i64 0, i64 %iv.0
+  %x1 = load i32, i32* %arrayidx1, align 4
+  %cmp = icmp eq i32 %x1, 0
+  %iv.1 = add nuw nsw i64 %iv.0, 1
+  br i1 %cmp, label %if.then, label %for.inc
+
+if.then:                                          ; preds = %for.body
+  %arrayidx2 = getelementptr inbounds i32, i32* %b, i64 %iv.0
+  %x2 = load i32, i32* %arrayidx2, align 4
+  %add = add nsw i32 %x2, %r.0
+  br label %for.inc
+
+for.inc:                                          ; preds = %for.body, %if.then
+  %r.1 = phi i32 [ %add, %if.then ], [ %x1, %for.body ]
+  %exitcond = icmp eq i64 %iv.1, 10
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.inc
+  ret i32 %r.1
+}
+
+; A branch might also depend on an induction variable - in this case we can
+; evaluate it at every iteration, and thus simplify CFG.
+; CHECK-LABEL: @branch_iv
+; CHECK-NOT:   br i1 %exitcond, label %for.end, label %for.body
+define i64 @branch_iv(i64* noalias nocapture readonly %b) {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.inc, %entry
+  %indvars.iv = phi i64 [ 0, %entry ], [ %tmp3, %for.inc ]
+  %r.030 = phi i64 [ 0, %entry ], [ %r.1, %for.inc ]
+  %cmp3 = icmp eq i64 %indvars.iv, 5
+  %tmp3 = add nuw nsw i64 %indvars.iv, 1
+  br i1 %cmp3, label %if.then, label %for.inc
+
+if.then:                                          ; preds = %for.body
+  %arrayidx2 = getelementptr inbounds i64, i64* %b, i64 %tmp3
+  %tmp1 = load i64, i64* %arrayidx2, align 4
+  %add = add nsw i64 %tmp1, %r.030
+  br label %for.inc
+
+for.inc:                                          ; preds = %if.then, %for.body
+  %r.1 = phi i64 [ %add, %if.then ], [ %r.030, %for.body ]
+  %exitcond = icmp eq i64 %tmp3, 20
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.inc
+  ret i64 %r.1
+}
+
+; Induction variables are often casted to another type, and that shouldn't
+; prevent us from folding branches. Tthis test specifically checks if we can
+; handle this.
+; CHECK-LABEL: @branch_iv_trunc
+; CHECK-NOT:   br i1 %exitcond, label %for.end, label %for.body
+define i32 @branch_iv_trunc(i32* noalias nocapture readonly %b) {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.inc, %entry
+  %indvars.iv = phi i64 [ 0, %entry ], [ %tmp3, %for.inc ]
+  %r.030 = phi i32 [ 0, %entry ], [ %r.1, %for.inc ]
+  %tmp2 = trunc i64 %indvars.iv to i32
+  %cmp3 = icmp eq i32 %tmp2, 5
+  %tmp3 = add nuw nsw i64 %indvars.iv, 1
+  br i1 %cmp3, label %if.then, label %for.inc
+
+if.then:                                          ; preds = %for.body
+  %arrayidx2 = getelementptr inbounds i32, i32* %b, i64 %tmp3
+  %tmp1 = load i32, i32* %arrayidx2, align 4
+  %add = add nsw i32 %tmp1, %r.030
+  br label %for.inc
+
+for.inc:                                          ; preds = %if.then, %for.body
+  %r.1 = phi i32 [ %add, %if.then ], [ %r.030, %for.body ]
+  %exitcond = icmp eq i64 %tmp3, 10
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.inc
+  ret i32 %r.1
+}
+
+; In this test %i1024 has a step of 1024, and it's truncated to i8 on every
+; iteration. Obviously, result of this trunc is always 0, and we should be able
+; to take advantage of it.
+; CHECK-LABEL: @iv_trunc
+; CHECK:   ret i8 0
+define i8 @iv_trunc() {
+entry:
+  br label %for.body
+
+for.body:
+  %iv.0 = phi i64 [ 0, %entry ], [ %iv.1, %for.body ]
+  %i1024.0 = phi i64 [ 0, %entry ], [ %i1024.1, %for.body ]
+  %r.0 = phi i8 [ 1, %entry ], [ %r.1, %for.body ]
+  %iv.1 = add nuw nsw i64 %iv.0, 1
+  %i1024.1 = add nuw nsw i64 %i1024.0, 1024
+  %trunc = trunc i64 %i1024.1 to i8
+  %r.1 = mul i8 %r.0, %trunc
+  %exitcond = icmp eq i64 %iv.1, 10
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:
+  ret i8 %r.1
+}
+
+; Check that we don't crash when we analyze icmp with pointer-typed IV and a pointer.
+; CHECK-LABEL: @ptr_cmp
+define void @ptr_cmp() {
+entry:
+  br label %while.body
+
+while.body:
+  %iv.0 = phi i32* [ getelementptr inbounds ([10 x i32], [10 x i32]* @known_constant, i64 0, i64 0), %entry ], [ %iv.1, %while.body ]
+  %iv.1 = getelementptr inbounds i32, i32* %iv.0, i64 1
+  %exitcond = icmp eq i32* %iv.1, getelementptr inbounds ([10 x i32], [10 x i32]* @known_constant, i64 0, i64 9)
+  br i1 %exitcond, label %loop.exit, label %while.body
+
+loop.exit:
+  ret void
+}
+
+; Check that we don't crash when we analyze ptrtoint cast.
+; CHECK-LABEL: @ptrtoint_cast
+define void @ptrtoint_cast(i8 * %a) {
+entry:
+  %limit = getelementptr i8, i8* %a, i64 512
+  br label %loop.body
+
+loop.body:
+  %iv.0 = phi i8* [ %a, %entry ], [ %iv.1, %loop.body ]
+  %cast = ptrtoint i8* %iv.0 to i64
+  %iv.1 = getelementptr inbounds i8, i8* %iv.0, i64 1
+  %exitcond = icmp ne i8* %iv.1, %limit
+  br i1 %exitcond, label %loop.body, label %loop.exit
+
+loop.exit:
+  ret void
+}
+
+; Check that we handle pointer arithmetic.
+; CHECK-LABEL: @ptr_sub
+define i64 @ptr_sub(i8 * %a) {
+entry:
+  %limit = getelementptr i8, i8* %a, i64 40
+  %start.iv2 = getelementptr i8, i8* %a, i64 7
+  br label %loop.body
+
+loop.body:
+  %iv.0 = phi i8* [ %a, %entry ], [ %iv.1, %loop.body ]
+  %iv2.0 = phi i8* [ %start.iv2, %entry ], [ %iv2.1, %loop.body ]
+  %r.0 = phi i64 [ 0, %entry ], [ %r.1, %loop.body ]
+  %cast = ptrtoint i8* %iv.0 to i64
+  %cmp = icmp eq i8* %iv2.0, %iv.0
+  %sub = sext i1 %cmp to i64
+  %mul = mul i64 %sub, %cast
+  %r.1 = add i64 %r.0, %mul
+  %iv.1 = getelementptr inbounds i8, i8* %iv.0, i64 1
+  %iv2.1 = getelementptr inbounds i8, i8* %iv2.0, i64 1
+  %exitcond = icmp ne i8* %iv.1, %limit
+  br i1 %exitcond, label %loop.body, label %loop.exit
+
+loop.exit:
+  ret i64 %r.1
+}