From 5d8c2e460cac05dedf3466d630995f1475317fc9 Mon Sep 17 00:00:00 2001 From: Tim Northover Date: Tue, 27 Aug 2013 13:46:45 +0000 Subject: [PATCH] DAGCombiner: make sure or/shl/srl really has zero high bits before forming bswap We want to convert code like (or (srl N, 8), (shl N, 8)) into (srl (bswap N), const), but this is only valid if the bits above 16 on the source pattern are 0, the checks we were doing on this were slightly wrong before. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189348 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 30 +++++++-- test/CodeGen/X86/bswap.ll | 107 +++++++++++++++++++++++++++++-- 2 files changed, 125 insertions(+), 12 deletions(-) diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index bd127b5a646..47c332a5be4 100644 --- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -2861,6 +2861,14 @@ SDValue DAGCombiner::visitAND(SDNode *N) { } } + // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const) + if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) { + SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), + N0.getOperand(1), false); + if (BSwap.getNode()) + return BSwap; + } + return SDValue(); } @@ -2945,13 +2953,23 @@ SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, if (N00 != N10) return SDValue(); - // Make sure everything beyond the low halfword is zero since the SRL 16 - // will clear the top bits. + // Make sure everything beyond the low halfword gets set to zero since the SRL + // 16 will clear the top bits. unsigned OpSizeInBits = VT.getSizeInBits(); - if (DemandHighBits && OpSizeInBits > 16 && - (!LookPassAnd0 || !LookPassAnd1) && - !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16))) - return SDValue(); + if (DemandHighBits && OpSizeInBits > 16) { + // If the left-shift isn't masked out then the only way this is a bswap is + // if all bits beyond the low 8 are 0. In that case the entire pattern + // reduces to a left shift anyway: leave it for other parts of the combiner. + if (!LookPassAnd0) + return SDValue(); + + // However, if the right shift isn't masked out then it might be because + // it's not needed. See if we can spot that too. + if (!LookPassAnd1 && + !DAG.MaskedValueIsZero( + N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16))) + return SDValue(); + } SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00); if (OpSizeInBits > 16) diff --git a/test/CodeGen/X86/bswap.ll b/test/CodeGen/X86/bswap.ll index 9e46592a4ac..e6a456c39dd 100644 --- a/test/CodeGen/X86/bswap.ll +++ b/test/CodeGen/X86/bswap.ll @@ -1,6 +1,7 @@ ; bswap should be constant folded when it is passed a constant argument ; RUN: llc < %s -march=x86 -mcpu=i686 | FileCheck %s +; RUN: llc < %s -march=x86-64 | FileCheck %s --check-prefix=CHECK64 declare i16 @llvm.bswap.i16(i16) @@ -11,6 +12,9 @@ declare i64 @llvm.bswap.i64(i64) define i16 @W(i16 %A) { ; CHECK-LABEL: W: ; CHECK: rolw $8, %ax + +; CHECK64-LABEL: W: +; CHECK64: rolw $8, % %Z = call i16 @llvm.bswap.i16( i16 %A ) ; [#uses=1] ret i16 %Z } @@ -18,6 +22,9 @@ define i16 @W(i16 %A) { define i32 @X(i32 %A) { ; CHECK-LABEL: X: ; CHECK: bswapl %eax + +; CHECK64-LABEL: X: +; CHECK64: bswapl % %Z = call i32 @llvm.bswap.i32( i32 %A ) ; [#uses=1] ret i32 %Z } @@ -26,6 +33,9 @@ define i64 @Y(i64 %A) { ; CHECK-LABEL: Y: ; CHECK: bswapl %eax ; CHECK: bswapl %edx + +; CHECK64-LABEL: Y: +; CHECK64: bswapq % %Z = call i64 @llvm.bswap.i64( i64 %A ) ; [#uses=1] ret i64 %Z } @@ -33,9 +43,13 @@ define i64 @Y(i64 %A) { ; rdar://9164521 define i32 @test1(i32 %a) nounwind readnone { entry: -; CHECK: test1 -; CHECK: bswapl %eax -; CHECK: shrl $16, %eax +; CHECK-LABEL: test1: +; CHECK: bswapl [[REG:%.*]] +; CHECK: shrl $16, [[REG]] + +; CHECK64-LABEL: test1: +; CHECK64: bswapl [[REG:%.*]] +; CHECK64: shrl $16, [[REG]] %and = lshr i32 %a, 8 %shr3 = and i32 %and, 255 %and2 = shl i32 %a, 8 @@ -46,9 +60,13 @@ entry: define i32 @test2(i32 %a) nounwind readnone { entry: -; CHECK: test2 -; CHECK: bswapl %eax -; CHECK: sarl $16, %eax +; CHECK-LABEL: test2: +; CHECK: bswapl [[REG:%.*]] +; CHECK: sarl $16, [[REG]] + +; CHECK64-LABEL: test2: +; CHECK64: bswapl [[REG:%.*]] +; CHECK64: sarl $16, [[REG]] %and = lshr i32 %a, 8 %shr4 = and i32 %and, 255 %and2 = shl i32 %a, 8 @@ -57,3 +75,80 @@ entry: %conv3 = ashr exact i32 %sext, 16 ret i32 %conv3 } + +@var8 = global i8 0 +@var16 = global i16 0 + +; The "shl" below can move bits into the high parts of the value, so the +; operation is not a "bswap, shr" pair. + +; rdar://problem/14814049 +define i64 @not_bswap() { +; CHECK-LABEL: not_bswap: +; CHECK-NOT: bswapl +; CHECK: ret + +; CHECK64-LABEL: not_bswap: +; CHECK64-NOT: bswapq +; CHECK64: ret + %init = load i16* @var16 + %big = zext i16 %init to i64 + + %hishifted = lshr i64 %big, 8 + %loshifted = shl i64 %big, 8 + + %notswapped = or i64 %hishifted, %loshifted + + ret i64 %notswapped +} + +; This time, the lshr (and subsequent or) is completely useless. While it's +; technically correct to convert this into a "bswap, shr", it's suboptimal. A +; simple shl works better. + +define i64 @not_useful_bswap() { +; CHECK-LABEL: not_useful_bswap: +; CHECK-NOT: bswapl +; CHECK: ret + +; CHECK64-LABEL: not_useful_bswap: +; CHECK64-NOT: bswapq +; CHECK64: ret + + %init = load i8* @var8 + %big = zext i8 %init to i64 + + %hishifted = lshr i64 %big, 8 + %loshifted = shl i64 %big, 8 + + %notswapped = or i64 %hishifted, %loshifted + + ret i64 %notswapped +} + +; Finally, it *is* OK to just mask off the shl if we know that the value is zero +; beyond 16 bits anyway. This is a legitimate bswap. + +define i64 @finally_useful_bswap() { +; CHECK-LABEL: finally_useful_bswap: +; CHECK: bswapl [[REG:%.*]] +; CHECK: shrl $16, [[REG]] +; CHECK: ret + +; CHECK64-LABEL: finally_useful_bswap: +; CHECK64: bswapq [[REG:%.*]] +; CHECK64: shrq $48, [[REG]] +; CHECK64: ret + + %init = load i16* @var16 + %big = zext i16 %init to i64 + + %hishifted = lshr i64 %big, 8 + %lomasked = and i64 %big, 255 + %loshifted = shl i64 %lomasked, 8 + + %swapped = or i64 %hishifted, %loshifted + + ret i64 %swapped +} + -- 2.11.0