diff options
Diffstat (limited to 'test/Transforms/PhaseOrdering')
-rw-r--r-- | test/Transforms/PhaseOrdering/PR6627.ll | 93 | ||||
-rw-r--r-- | test/Transforms/PhaseOrdering/basic.ll | 111 | ||||
-rw-r--r-- | test/Transforms/PhaseOrdering/scev.ll | 64 |
3 files changed, 179 insertions, 89 deletions
diff --git a/test/Transforms/PhaseOrdering/PR6627.ll b/test/Transforms/PhaseOrdering/PR6627.ll new file mode 100644 index 0000000..ef9947f --- /dev/null +++ b/test/Transforms/PhaseOrdering/PR6627.ll @@ -0,0 +1,93 @@ +; RUN: opt -O3 -S %s | FileCheck %s +; XFAIL: * + +declare i32 @doo(...) + +; PR6627 - This whole nasty sequence should be flattened down to a single +; 32-bit comparison. +define void @test2(i8* %arrayidx) nounwind ssp { +entry: + %xx = bitcast i8* %arrayidx to i32* + %x1 = load i32* %xx, align 4 + %tmp = trunc i32 %x1 to i8 + %conv = zext i8 %tmp to i32 + %cmp = icmp eq i32 %conv, 127 + br i1 %cmp, label %land.lhs.true, label %if.end + +land.lhs.true: ; preds = %entry + %arrayidx4 = getelementptr inbounds i8* %arrayidx, i64 1 + %tmp5 = load i8* %arrayidx4, align 1 + %conv6 = zext i8 %tmp5 to i32 + %cmp7 = icmp eq i32 %conv6, 69 + br i1 %cmp7, label %land.lhs.true9, label %if.end + +land.lhs.true9: ; preds = %land.lhs.true + %arrayidx12 = getelementptr inbounds i8* %arrayidx, i64 2 + %tmp13 = load i8* %arrayidx12, align 1 + %conv14 = zext i8 %tmp13 to i32 + %cmp15 = icmp eq i32 %conv14, 76 + br i1 %cmp15, label %land.lhs.true17, label %if.end + +land.lhs.true17: ; preds = %land.lhs.true9 + %arrayidx20 = getelementptr inbounds i8* %arrayidx, i64 3 + %tmp21 = load i8* %arrayidx20, align 1 + %conv22 = zext i8 %tmp21 to i32 + %cmp23 = icmp eq i32 %conv22, 70 + br i1 %cmp23, label %if.then, label %if.end + +if.then: ; preds = %land.lhs.true17 + %call25 = call i32 (...)* @doo() + br label %if.end + +if.end: + ret void + +; CHECK: @test2 +; CHECK: %x1 = load i32* %xx, align 4 +; CHECK-NEXT: icmp eq i32 %x1, 1179403647 +; CHECK-NEXT: br i1 {{.*}}, label %if.then, label %if.end +} + +; PR6627 - This should all be flattened down to one compare. This is the same +; as test2, except that the initial load is done as an i8 instead of i32, thus +; requiring widening. +define void @test2a(i8* %arrayidx) nounwind ssp { +entry: + %x1 = load i8* %arrayidx, align 4 + %conv = zext i8 %x1 to i32 + %cmp = icmp eq i32 %conv, 127 + br i1 %cmp, label %land.lhs.true, label %if.end + +land.lhs.true: ; preds = %entry + %arrayidx4 = getelementptr inbounds i8* %arrayidx, i64 1 + %tmp5 = load i8* %arrayidx4, align 1 + %conv6 = zext i8 %tmp5 to i32 + %cmp7 = icmp eq i32 %conv6, 69 + br i1 %cmp7, label %land.lhs.true9, label %if.end + +land.lhs.true9: ; preds = %land.lhs.true + %arrayidx12 = getelementptr inbounds i8* %arrayidx, i64 2 + %tmp13 = load i8* %arrayidx12, align 1 + %conv14 = zext i8 %tmp13 to i32 + %cmp15 = icmp eq i32 %conv14, 76 + br i1 %cmp15, label %land.lhs.true17, label %if.end + +land.lhs.true17: ; preds = %land.lhs.true9 + %arrayidx20 = getelementptr inbounds i8* %arrayidx, i64 3 + %tmp21 = load i8* %arrayidx20, align 1 + %conv22 = zext i8 %tmp21 to i32 + %cmp23 = icmp eq i32 %conv22, 70 + br i1 %cmp23, label %if.then, label %if.end + +if.then: ; preds = %land.lhs.true17 + %call25 = call i32 (...)* @doo() + br label %if.end + +if.end: + ret void + +; CHECK: @test2a +; CHECK: %x1 = load i32* {{.*}}, align 4 +; CHECK-NEXT: icmp eq i32 %x1, 1179403647 +; CHECK-NEXT: br i1 {{.*}}, label %if.then, label %if.end +} diff --git a/test/Transforms/PhaseOrdering/basic.ll b/test/Transforms/PhaseOrdering/basic.ll index e5b2ba4..88ebca0 100644 --- a/test/Transforms/PhaseOrdering/basic.ll +++ b/test/Transforms/PhaseOrdering/basic.ll @@ -1,5 +1,4 @@ ; RUN: opt -O3 -S %s | FileCheck %s -; XFAIL: * target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64" target triple = "x86_64-apple-macosx10.6.7" @@ -24,95 +23,29 @@ define void @test1() nounwind ssp { ; CHECK-NEXT: ret void } - -; PR6627 - This whole nasty sequence should be flattened down to a single -; 32-bit comparison. -define void @test2(i8* %arrayidx) nounwind ssp { +; This function exposes a phase ordering problem when InstCombine is +; turning %add into a bitmask, making it difficult to spot a 0 return value. +; +; It it also important that %add is expressed as a multiple of %div so scalar +; evolution can recognize it. +define i32 @test2(i32 %a, i32* %p) nounwind uwtable ssp { entry: - %xx = bitcast i8* %arrayidx to i32* - %x1 = load i32* %xx, align 4 - %tmp = trunc i32 %x1 to i8 - %conv = zext i8 %tmp to i32 - %cmp = icmp eq i32 %conv, 127 - br i1 %cmp, label %land.lhs.true, label %if.end - -land.lhs.true: ; preds = %entry - %arrayidx4 = getelementptr inbounds i8* %arrayidx, i64 1 - %tmp5 = load i8* %arrayidx4, align 1 - %conv6 = zext i8 %tmp5 to i32 - %cmp7 = icmp eq i32 %conv6, 69 - br i1 %cmp7, label %land.lhs.true9, label %if.end - -land.lhs.true9: ; preds = %land.lhs.true - %arrayidx12 = getelementptr inbounds i8* %arrayidx, i64 2 - %tmp13 = load i8* %arrayidx12, align 1 - %conv14 = zext i8 %tmp13 to i32 - %cmp15 = icmp eq i32 %conv14, 76 - br i1 %cmp15, label %land.lhs.true17, label %if.end - -land.lhs.true17: ; preds = %land.lhs.true9 - %arrayidx20 = getelementptr inbounds i8* %arrayidx, i64 3 - %tmp21 = load i8* %arrayidx20, align 1 - %conv22 = zext i8 %tmp21 to i32 - %cmp23 = icmp eq i32 %conv22, 70 - br i1 %cmp23, label %if.then, label %if.end - -if.then: ; preds = %land.lhs.true17 - %call25 = call i32 (...)* @doo() - br label %if.end - -if.end: - ret void + %div = udiv i32 %a, 4 + %arrayidx = getelementptr inbounds i32* %p, i64 0 + store i32 %div, i32* %arrayidx, align 4 + %add = add i32 %div, %div + %arrayidx1 = getelementptr inbounds i32* %p, i64 1 + store i32 %add, i32* %arrayidx1, align 4 + %arrayidx2 = getelementptr inbounds i32* %p, i64 1 + %0 = load i32* %arrayidx2, align 4 + %arrayidx3 = getelementptr inbounds i32* %p, i64 0 + %1 = load i32* %arrayidx3, align 4 + %mul = mul i32 2, %1 + %sub = sub i32 %0, %mul + ret i32 %sub ; CHECK: @test2 -; CHECK: %x1 = load i32* %xx, align 4 -; CHECK-NEXT: icmp eq i32 %x1, 1179403647 -; CHECK-NEXT: br i1 {{.*}}, label %if.then, label %if.end +; CHECK: %div = lshr i32 %a, 2 +; CHECK: %add = shl nuw nsw i32 %div, 1 +; CHECK: ret i32 0 } - -declare i32 @doo(...) - -; PR6627 - This should all be flattened down to one compare. This is the same -; as test2, except that the initial load is done as an i8 instead of i32, thus -; requiring widening. -define void @test2a(i8* %arrayidx) nounwind ssp { -entry: - %x1 = load i8* %arrayidx, align 4 - %conv = zext i8 %x1 to i32 - %cmp = icmp eq i32 %conv, 127 - br i1 %cmp, label %land.lhs.true, label %if.end - -land.lhs.true: ; preds = %entry - %arrayidx4 = getelementptr inbounds i8* %arrayidx, i64 1 - %tmp5 = load i8* %arrayidx4, align 1 - %conv6 = zext i8 %tmp5 to i32 - %cmp7 = icmp eq i32 %conv6, 69 - br i1 %cmp7, label %land.lhs.true9, label %if.end - -land.lhs.true9: ; preds = %land.lhs.true - %arrayidx12 = getelementptr inbounds i8* %arrayidx, i64 2 - %tmp13 = load i8* %arrayidx12, align 1 - %conv14 = zext i8 %tmp13 to i32 - %cmp15 = icmp eq i32 %conv14, 76 - br i1 %cmp15, label %land.lhs.true17, label %if.end - -land.lhs.true17: ; preds = %land.lhs.true9 - %arrayidx20 = getelementptr inbounds i8* %arrayidx, i64 3 - %tmp21 = load i8* %arrayidx20, align 1 - %conv22 = zext i8 %tmp21 to i32 - %cmp23 = icmp eq i32 %conv22, 70 - br i1 %cmp23, label %if.then, label %if.end - -if.then: ; preds = %land.lhs.true17 - %call25 = call i32 (...)* @doo() - br label %if.end - -if.end: - ret void - -; CHECK: @test2a -; CHECK: %x1 = load i32* {{.*}}, align 4 -; CHECK-NEXT: icmp eq i32 %x1, 1179403647 -; CHECK-NEXT: br i1 {{.*}}, label %if.then, label %if.end -} - diff --git a/test/Transforms/PhaseOrdering/scev.ll b/test/Transforms/PhaseOrdering/scev.ll new file mode 100644 index 0000000..c731280 --- /dev/null +++ b/test/Transforms/PhaseOrdering/scev.ll @@ -0,0 +1,64 @@ +; RUN: opt -O3 -S -analyze -scalar-evolution %s | FileCheck %s +; +; This file contains phase ordering tests for scalar evolution. +; Test that the standard passes don't obfuscate the IR so scalar evolution can't +; recognize expressions. + +; CHECK: test1 +; The loop body contains two increments by %div. +; Make sure that 2*%div is recognizable, and not expressed as a bit mask of %d. +; CHECK: --> {%p,+,(2 * (%d /u 4) * sizeof(i32))} +define void @test1(i64 %d, i32* %p) nounwind uwtable ssp { +entry: + %div = udiv i64 %d, 4 + br label %for.cond + +for.cond: ; preds = %for.inc, %entry + %p.addr.0 = phi i32* [ %p, %entry ], [ %add.ptr1, %for.inc ] + %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.inc ] + %cmp = icmp ne i32 %i.0, 64 + br i1 %cmp, label %for.body, label %for.end + +for.body: ; preds = %for.cond + store i32 0, i32* %p.addr.0, align 4 + %add.ptr = getelementptr inbounds i32* %p.addr.0, i64 %div + store i32 1, i32* %add.ptr, align 4 + %add.ptr1 = getelementptr inbounds i32* %add.ptr, i64 %div + br label %for.inc + +for.inc: ; preds = %for.body + %inc = add i32 %i.0, 1 + br label %for.cond + +for.end: ; preds = %for.cond + ret void +} + +; CHECK: test1a +; Same thing as test1, but it is even more tempting to fold 2 * (%d /u 2) +; CHECK: --> {%p,+,(2 * (%d /u 2) * sizeof(i32))} +define void @test1a(i64 %d, i32* %p) nounwind uwtable ssp { +entry: + %div = udiv i64 %d, 2 + br label %for.cond + +for.cond: ; preds = %for.inc, %entry + %p.addr.0 = phi i32* [ %p, %entry ], [ %add.ptr1, %for.inc ] + %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.inc ] + %cmp = icmp ne i32 %i.0, 64 + br i1 %cmp, label %for.body, label %for.end + +for.body: ; preds = %for.cond + store i32 0, i32* %p.addr.0, align 4 + %add.ptr = getelementptr inbounds i32* %p.addr.0, i64 %div + store i32 1, i32* %add.ptr, align 4 + %add.ptr1 = getelementptr inbounds i32* %add.ptr, i64 %div + br label %for.inc + +for.inc: ; preds = %for.body + %inc = add i32 %i.0, 1 + br label %for.cond + +for.end: ; preds = %for.cond + ret void +} |