[llvm] r204912 - InstCombine: merge constants in both operands of icmp.
Erik Verbruggen
erikjv at me.com
Thu Mar 27 12:23:21 PDT 2014
Actually, I'm wondering if it's better to revert my patch (and yours). The "nightly" test-suite pointed out a regression. If you have a tight for loop, like:
for (i = 0; i <= 39; i++)
doSomethingHere;
then before the patch it would be:
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
%exitcond = icmp eq i64 %indvars.iv.next, 40
while with the patch it's:
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
%exitcond = icmp eq i64 %indvars.iv, 39
That keeps the "old" %indvars.iv alive. This causes some performance tests to fail, like MultiSource/Benchmarks/MiBench/telecomm-gsm.
I Cc-ed d0k because he did the review, so what do you guys think?
-- Erik.
On 27 Mar 2014, at 18:56, Reid Kleckner <rnk at google.com> wrote:
> This transform was broken for unsigned comparisons. I disabled it in r204948, which should fix the problem. See the commit message for the counterexample.
>
>
> On Thu, Mar 27, 2014 at 4:16 AM, Erik Verbruggen <erikjv at me.com> wrote:
> Author: erikjv
> Date: Thu Mar 27 06:16:05 2014
> New Revision: 204912
>
> URL: http://llvm.org/viewvc/llvm-project?rev=204912&view=rev
> Log:
> InstCombine: merge constants in both operands of icmp.
>
> Transform:
> icmp X+Cst2, Cst
> into:
> icmp X, Cst-Cst2
> when Cst-Cst2 does not overflow, and the add has nsw.
>
> Modified:
> llvm/trunk/lib/Target/README.txt
> llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp
> llvm/trunk/test/Transforms/InstCombine/icmp.ll
>
> Modified: llvm/trunk/lib/Target/README.txt
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/README.txt?rev=204912&r1=204911&r2=204912&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Target/README.txt (original)
> +++ llvm/trunk/lib/Target/README.txt Thu Mar 27 06:16:05 2014
> @@ -930,18 +930,6 @@ optimized with "clang -emit-llvm-bc | op
>
> //===---------------------------------------------------------------------===//
>
> -int g(int x) { return (x - 10) < 0; }
> -Should combine to "x <= 9" (the sub has nsw). Currently not
> -optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
> -
> -//===---------------------------------------------------------------------===//
> -
> -int g(int x) { return (x + 10) < 0; }
> -Should combine to "x < -10" (the add has nsw). Currently not
> -optimized with "clang -emit-llvm-bc | opt -std-compile-opts".
> -
> -//===---------------------------------------------------------------------===//
> -
> int f(int i, int j) { return i < j + 1; }
> int g(int i, int j) { return j > i - 1; }
> Should combine to "i <= j" (the add/sub has nsw). Currently not
>
> Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp?rev=204912&r1=204911&r2=204912&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp (original)
> +++ llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp Thu Mar 27 06:16:05 2014
> @@ -3008,6 +3008,19 @@ Instruction *InstCombiner::visitICmpInst
> // icmp X, X+Cst
> if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
> return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate());
> +
> + ConstantInt *Cst2;
> + if (match(Op1, m_ConstantInt(Cst)) &&
> + match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst2))) &&
> + cast<BinaryOperator>(Op0)->hasNoSignedWrap()) {
> + // icmp X+Cst2, Cst --> icmp X, Cst-Cst2
> + // iff Cst-Cst2 does not overflow
> + bool Overflow;
> + APInt NewCst = Cst->getValue().ssub_ov(Cst2->getValue(), Overflow);
> + if (!Overflow)
> + return new ICmpInst(I.getPredicate(), X,
> + ConstantInt::get(Cst->getType(), NewCst));
> + }
> }
> return Changed ? &I : 0;
> }
>
> Modified: llvm/trunk/test/Transforms/InstCombine/icmp.ll
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/icmp.ll?rev=204912&r1=204911&r2=204912&view=diff
> ==============================================================================
> --- llvm/trunk/test/Transforms/InstCombine/icmp.ll (original)
> +++ llvm/trunk/test/Transforms/InstCombine/icmp.ll Thu Mar 27 06:16:05 2014
> @@ -1356,3 +1356,56 @@ define i1 @icmp_ashr_ashr_ne(i32 %a, i32
> %z = icmp ne i32 %x, %y
> ret i1 %z
> }
> +
> +; CHECK-LABEL: icmp_add_const_const1
> +; CHECK: %cmp = icmp slt i32 %x, -10
> +; CHECK-NOT: %add = add nsw i32 %x, 10
> +define i32 @icmp_add_const_const1(i32 %x) nounwind ssp uwtable {
> +entry:
> + %add = add nsw i32 %x, 10
> + %cmp = icmp slt i32 %add, 0
> + %conv = zext i1 %cmp to i32
> + ret i32 %conv
> +}
> +
> +; CHECK-LABEL: icmp_add_const_const2
> +; CHECK: %cmp = icmp slt i32 %x, -10
> +; CHECK-NOT: %add = add nsw i32 %x, 10
> +define i32 @icmp_add_const_const2(i32 %x) nounwind ssp uwtable {
> +entry:
> + %add = add nsw i32 10, %x
> + %cmp = icmp sgt i32 0, %add
> + %conv = zext i1 %cmp to i32
> + ret i32 %conv
> +}
> +
> +; CHECK-LABEL: icmp_add_const_const3
> +; CHECK: %cmp = icmp slt i32 %x, 20
> +; CHECK-NOT: %sub = add nsw i32 %x, -10
> +define i32 @icmp_add_const_const3(i32 %x) nounwind ssp uwtable {
> +entry:
> + %add = add nsw i32 -10, %x
> + %cmp = icmp sgt i32 10, %add
> + %conv = zext i1 %cmp to i32
> + ret i32 %conv
> +}
> +
> +; CHECK-LABEL: icmp_add_const_intmin
> +; CHECK: %cmp = icmp ne i32 %x, 2147483638
> +define i32 @icmp_add_const_intmin(i32 %x) nounwind ssp uwtable {
> +entry:
> + %add = add nsw i32 %x, 10
> + %cmp = icmp sgt i32 %add, -2147483648
> + %conv = zext i1 %cmp to i32
> + ret i32 %conv
> +}
> +
> +; CHECK-LABEL: icmp_add_const_intmax
> +; CHECK: %cmp = icmp ne i32 %x, 2147483637
> +define i32 @icmp_add_const_intmax(i32 %x) nounwind ssp uwtable {
> +entry:
> + %add = add nsw i32 %x, 10
> + %cmp = icmp slt i32 %add, 2147483647
> + %conv = zext i1 %cmp to i32
> + ret i32 %conv
> +}
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140327/cda7523b/attachment.html>
More information about the llvm-commits
mailing list