diff options
author | Dan Gohman <gohman@apple.com> | 2011-07-01 22:05:19 +0000 |
---|---|---|
committer | Dan Gohman <gohman@apple.com> | 2011-07-01 22:05:19 +0000 |
commit | 71997f303e47343e994fba53814455753eeb7e05 (patch) | |
tree | ab5eb06761b724574b21464f7e2cf0bb2bbe52a2 | |
parent | f1ac465b67d5fc11a0d9cd09b98ceb4ffa75dd97 (diff) | |
download | llvm-71997f303e47343e994fba53814455753eeb7e05.tar.gz llvm-71997f303e47343e994fba53814455753eeb7e05.tar.bz2 llvm-71997f303e47343e994fba53814455753eeb7e05.tar.xz |
Teach IVUsers to stop at non-affine expressions unless they are both
outside the loop and reducible.
This more completely hides them from LSR, which isn't usually able to
do anything meaningful with non-affine expressions anyway, and this
consequently hides them from SCEVExpander, which is acutely unprepared
for non-affine expressions.
Replace test/CodeGen/X86/lsr-nonaffine.ll with a new test that tests
the new behavior.
This works around the bug in PR10117 / rdar://problem/9633149, and is
generally an improvement besides.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@134268 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r-- | lib/Analysis/IVUsers.cpp | 17 | ||||
-rw-r--r-- | test/CodeGen/X86/lsr-nonaffine.ll | 43 |
2 files changed, 35 insertions, 25 deletions
diff --git a/lib/Analysis/IVUsers.cpp b/lib/Analysis/IVUsers.cpp index 7a9dc0f14c..e5f0a77ab6 100644 --- a/lib/Analysis/IVUsers.cpp +++ b/lib/Analysis/IVUsers.cpp @@ -46,17 +46,20 @@ Pass *llvm::createIVUsersPass() { /// used by the given expression, within the context of analyzing the /// given loop. static bool isInteresting(const SCEV *S, const Instruction *I, const Loop *L, - ScalarEvolution *SE) { + ScalarEvolution *SE, LoopInfo *LI) { // An addrec is interesting if it's affine or if it has an interesting start. if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { - // Keep things simple. Don't touch loop-variant strides. + // Keep things simple. Don't touch loop-variant strides unless they're + // only used outside the loop and we can simplify them. if (AR->getLoop() == L) - return AR->isAffine() || !L->contains(I); + return AR->isAffine() || + (!L->contains(I) && + SE->getSCEVAtScope(AR, LI->getLoopFor(I->getParent())) != AR); // Otherwise recurse to see if the start value is interesting, and that // the step value is not interesting, since we don't yet know how to // do effective SCEV expansions for addrecs with interesting steps. - return isInteresting(AR->getStart(), I, L, SE) && - !isInteresting(AR->getStepRecurrence(*SE), I, L, SE); + return isInteresting(AR->getStart(), I, L, SE, LI) && + !isInteresting(AR->getStepRecurrence(*SE), I, L, SE, LI); } // An add is interesting if exactly one of its operands is interesting. @@ -64,7 +67,7 @@ static bool isInteresting(const SCEV *S, const Instruction *I, const Loop *L, bool AnyInterestingYet = false; for (SCEVAddExpr::op_iterator OI = Add->op_begin(), OE = Add->op_end(); OI != OE; ++OI) - if (isInteresting(*OI, I, L, SE)) { + if (isInteresting(*OI, I, L, SE, LI)) { if (AnyInterestingYet) return false; AnyInterestingYet = true; @@ -98,7 +101,7 @@ bool IVUsers::AddUsersIfInteresting(Instruction *I) { // If we've come to an uninteresting expression, stop the traversal and // call this a user. - if (!isInteresting(ISE, I, L, SE)) + if (!isInteresting(ISE, I, L, SE, LI)) return false; SmallPtrSet<Instruction *, 4> UniqueUsers; diff --git a/test/CodeGen/X86/lsr-nonaffine.ll b/test/CodeGen/X86/lsr-nonaffine.ll index b0d30641dd..4771646f41 100644 --- a/test/CodeGen/X86/lsr-nonaffine.ll +++ b/test/CodeGen/X86/lsr-nonaffine.ll @@ -1,23 +1,30 @@ -; RUN: llc -march=x86-64 < %s | FileCheck %s +; RUN: llc -asm-verbose=false -march=x86-64 -o - < %s | FileCheck %s -; LSR should compute the correct starting values for this loop. Note that -; it's not necessarily LSR's job to compute loop exit expressions; that's -; indvars' job. -; CHECK: movl $12 -; CHECK: movl $42 +; LSR should leave non-affine expressions alone because it currently +; doesn't know how to do anything with them, and when it tries, it +; gets SCEVExpander's current expansion for them, which is suboptimal. -define i32 @real_symmetric_eigen(i32 %n) nounwind { -while.body127: ; preds = %while.cond122 - br label %while.cond141 +; CHECK: xorl %eax, %eax +; CHECK-NEXT: align +; CHECK-NEXT: BB0_1: +; CHECK-NEXT: movq %rax, (%rdx) +; CHECK-NEXT: addq %rsi, %rax +; CHECK-NEXT: cmpq %rdi, %rax +; CHECK-NEXT: jl +; CHECK-NEXT: imulq %rax, %rax +; CHECK-NEXT: ret +define i64 @foo(i64 %n, i64 %s, i64* %p) nounwind { +entry: + br label %loop -while.cond141: ; preds = %while.cond141, %while.body127 - %0 = phi i32 [ 7, %while.body127 ], [ %indvar.next67, %while.cond141 ] ; <i32> [#uses=3] - %indvar.next67 = add i32 %0, 1 ; <i32> [#uses=1] - %t = icmp slt i32 %indvar.next67, %n - br i1 %t, label %if.then171, label %while.cond141 +loop: + %i = phi i64 [ 0, %entry ], [ %i.next, %loop ] + volatile store i64 %i, i64* %p + %i.next = add i64 %i, %s + %c = icmp slt i64 %i.next, %n + br i1 %c, label %loop, label %exit -if.then171: ; preds = %while.cond141 - %mul150 = mul i32 %0, %0 ; <i32> [#uses=1] - %add174 = add i32 %mul150, %0 ; <i32> [#uses=1] - ret i32 %add174 +exit: + %mul = mul i64 %i.next, %i.next + ret i64 %mul } |