From 014c000f6aa84528dfe492bfb02a04ef2993b98c Mon Sep 17 00:00:00 2001 From: Sanjay Patel Date: Mon, 12 Feb 2018 21:37:27 +0000 Subject: [PATCH] [DAG] make binops with undef operands consistent with IR This started by noticing that scalar and vector types were producing different results with div ops in PR36305: https://bugs.llvm.org/show_bug.cgi?id=36305 ...but the problem is bigger. I couldn't keep it straight without a table, so I'm attaching that as a PDF to the review. The x86 tests in undef-ops.ll correspond to that table. Green means that instsimplify and the DAG agree on the result for all types. Red means the DAG was returning undef when IR was not. Yellow means the DAG was returning a non-undef result when IR returned undef. This patch assumes that we're currently doing the right thing in IR. Note: I couldn't find any problems with lowering vector constants as the code comments were warning, but those comments were written long ago in rL36413 . Differential Revision: https://reviews.llvm.org/D43141 llvm-svn: 324941 --- .../lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 27 +++++-------------- .../CodeGen/Hexagon/tail-dup-subreg-map.ll | 8 +++--- llvm/test/CodeGen/X86/pr13577.ll | 1 - llvm/test/CodeGen/X86/pr33960.ll | 4 +-- llvm/test/CodeGen/X86/undef-ops.ll | 17 +++++++++--- 5 files changed, 27 insertions(+), 30 deletions(-) diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp index 68502fd23780..2ece6a82f6c1 100644 --- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp @@ -4667,19 +4667,15 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, case ISD::FSUB: case ISD::FDIV: case ISD::FREM: - case ISD::SRA: return N1; // fold op(undef, arg2) -> undef case ISD::UDIV: case ISD::SDIV: case ISD::UREM: case ISD::SREM: + case ISD::SRA: case ISD::SRL: case ISD::SHL: - if (!VT.isVector()) - return getConstant(0, DL, VT); // fold op(undef, arg2) -> 0 - // For vectors, we can't easily build an all zero vector, just return - // the LHS. - return N2; + return getConstant(0, DL, VT); // fold op(undef, arg2) -> 0 } } } @@ -4701,6 +4697,9 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, case ISD::SDIV: case ISD::UREM: case ISD::SREM: + case ISD::SRA: + case ISD::SRL: + case ISD::SHL: return N2; // fold op(arg1, undef) -> undef case ISD::FADD: case ISD::FSUB: @@ -4712,21 +4711,9 @@ SDValue SelectionDAG::getNode(unsigned Opcode, const SDLoc &DL, EVT VT, break; case ISD::MUL: case ISD::AND: - case ISD::SRL: - case ISD::SHL: - if (!VT.isVector()) - return getConstant(0, DL, VT); // fold op(arg1, undef) -> 0 - // For vectors, we can't easily build an all zero vector, just return - // the LHS. - return N1; + return getConstant(0, DL, VT); // fold op(arg1, undef) -> 0 case ISD::OR: - if (!VT.isVector()) - return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT); - // For vectors, we can't easily build an all one vector, just return - // the LHS. - return N1; - case ISD::SRA: - return N1; + return getAllOnesConstant(DL, VT); } } diff --git a/llvm/test/CodeGen/Hexagon/tail-dup-subreg-map.ll b/llvm/test/CodeGen/Hexagon/tail-dup-subreg-map.ll index 1b11d087832a..148d018b01ec 100644 --- a/llvm/test/CodeGen/Hexagon/tail-dup-subreg-map.ll +++ b/llvm/test/CodeGen/Hexagon/tail-dup-subreg-map.ll @@ -5,7 +5,7 @@ ; subregisters were dropped by the tail duplicator, resulting in invalid ; COPY instructions being generated. -; CHECK: = extractu(r{{[0-9]+}},#15,#17) +; CHECK: = asl(r{{[0-9]+}}:{{[0-9]+}},#15) target triple = "hexagon" @@ -36,20 +36,20 @@ if.then5.i: ; preds = %if.then.i br label %if.end.i if.else.i: ; preds = %if.then.i - %shl12.i = shl i64 %0, undef + %shl12.i = shl i64 %0, 7 br label %if.end.i if.end.i: ; preds = %if.else.i, %if.then5.i %aSig0.0 = phi i64 [ undef, %if.then5.i ], [ %shl12.i, %if.else.i ] %storemerge43.i = phi i64 [ %shl.i21, %if.then5.i ], [ 0, %if.else.i ] - %sub15.i = sub nsw i32 -63, undef + %sub15.i = sub nsw i32 -63, 8 br label %if.end13 if.else16.i: ; preds = %if.then7 br label %if.end13 if.else: ; preds = %entry - %or12 = or i64 undef, 281474976710656 + %or12 = or i64 9, 281474976710656 br label %if.end13 if.end13: ; preds = %if.else, %if.else16.i, %if.end.i diff --git a/llvm/test/CodeGen/X86/pr13577.ll b/llvm/test/CodeGen/X86/pr13577.ll index 66bbf4531e5c..8c0752b127a1 100644 --- a/llvm/test/CodeGen/X86/pr13577.ll +++ b/llvm/test/CodeGen/X86/pr13577.ll @@ -31,7 +31,6 @@ define float @pr26070() { ; CHECK: ## %bb.0: ; CHECK-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,0,0,0] -; CHECK-NEXT: orps {{.*}}(%rip), %xmm0 ; CHECK-NEXT: retq %c = call float @copysignf(float 1.0, float undef) readnone ret float %c diff --git a/llvm/test/CodeGen/X86/pr33960.ll b/llvm/test/CodeGen/X86/pr33960.ll index 34af4df94551..fd5a9a400503 100644 --- a/llvm/test/CodeGen/X86/pr33960.ll +++ b/llvm/test/CodeGen/X86/pr33960.ll @@ -7,12 +7,12 @@ define void @PR33960() { ; X86-LABEL: PR33960: ; X86: # %bb.0: # %entry -; X86-NEXT: movl $0, b +; X86-NEXT: movl $-1, b ; X86-NEXT: retl ; ; X64-LABEL: PR33960: ; X64: # %bb.0: # %entry -; X64-NEXT: movl $0, {{.*}}(%rip) +; X64-NEXT: movl $-1, {{.*}}(%rip) ; X64-NEXT: retq entry: %tmp = insertelement <4 x i32> , i32 -2, i32 3 diff --git a/llvm/test/CodeGen/X86/undef-ops.ll b/llvm/test/CodeGen/X86/undef-ops.ll index 18a408a7fc5f..1ed1ec18e1d5 100644 --- a/llvm/test/CodeGen/X86/undef-ops.ll +++ b/llvm/test/CodeGen/X86/undef-ops.ll @@ -77,6 +77,7 @@ define i32 @mul_undef_rhs(i32 %x) { define <4 x i32> @mul_undef_rhs_vec(<4 x i32> %x) { ; CHECK-LABEL: mul_undef_rhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = mul <4 x i32> %x, undef ret <4 x i32> %r @@ -94,6 +95,7 @@ define i32 @mul_undef_lhs(i32 %x) { define <4 x i32> @mul_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: mul_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = mul <4 x i32> undef, %x ret <4 x i32> %r @@ -127,6 +129,7 @@ define i32 @sdiv_undef_lhs(i32 %x) { define <4 x i32> @sdiv_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: sdiv_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = sdiv <4 x i32> undef, %x ret <4 x i32> %r @@ -160,6 +163,7 @@ define i32 @udiv_undef_lhs(i32 %x) { define <4 x i32> @udiv_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: udiv_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = udiv <4 x i32> undef, %x ret <4 x i32> %r @@ -193,6 +197,7 @@ define i32 @srem_undef_lhs(i32 %x) { define <4 x i32> @srem_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: srem_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = srem <4 x i32> undef, %x ret <4 x i32> %r @@ -226,6 +231,7 @@ define i32 @urem_undef_lhs(i32 %x) { define <4 x i32> @urem_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: urem_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = urem <4 x i32> undef, %x ret <4 x i32> %r @@ -234,7 +240,6 @@ define <4 x i32> @urem_undef_lhs_vec(<4 x i32> %x) { define i32 @ashr_undef_rhs(i32 %x) { ; CHECK-LABEL: ashr_undef_rhs: ; CHECK: # %bb.0: -; CHECK-NEXT: movl %edi, %eax ; CHECK-NEXT: retq %r = ashr i32 %x, undef ret i32 %r @@ -251,6 +256,7 @@ define <4 x i32> @ashr_undef_rhs_vec(<4 x i32> %x) { define i32 @ashr_undef_lhs(i32 %x) { ; CHECK-LABEL: ashr_undef_lhs: ; CHECK: # %bb.0: +; CHECK-NEXT: xorl %eax, %eax ; CHECK-NEXT: retq %r = ashr i32 undef, %x ret i32 %r @@ -259,6 +265,7 @@ define i32 @ashr_undef_lhs(i32 %x) { define <4 x i32> @ashr_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: ashr_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = ashr <4 x i32> undef, %x ret <4 x i32> %r @@ -267,7 +274,6 @@ define <4 x i32> @ashr_undef_lhs_vec(<4 x i32> %x) { define i32 @lshr_undef_rhs(i32 %x) { ; CHECK-LABEL: lshr_undef_rhs: ; CHECK: # %bb.0: -; CHECK-NEXT: xorl %eax, %eax ; CHECK-NEXT: retq %r = lshr i32 %x, undef ret i32 %r @@ -293,6 +299,7 @@ define i32 @lshr_undef_lhs(i32 %x) { define <4 x i32> @lshr_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: lshr_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = lshr <4 x i32> undef, %x ret <4 x i32> %r @@ -301,7 +308,6 @@ define <4 x i32> @lshr_undef_lhs_vec(<4 x i32> %x) { define i32 @shl_undef_rhs(i32 %x) { ; CHECK-LABEL: shl_undef_rhs: ; CHECK: # %bb.0: -; CHECK-NEXT: xorl %eax, %eax ; CHECK-NEXT: retq %r = shl i32 %x, undef ret i32 %r @@ -327,6 +333,7 @@ define i32 @shl_undef_lhs(i32 %x) { define <4 x i32> @shl_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: shl_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = shl <4 x i32> undef, %x ret <4 x i32> %r @@ -344,6 +351,7 @@ define i32 @and_undef_rhs(i32 %x) { define <4 x i32> @and_undef_rhs_vec(<4 x i32> %x) { ; CHECK-LABEL: and_undef_rhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = and <4 x i32> %x, undef ret <4 x i32> %r @@ -361,6 +369,7 @@ define i32 @and_undef_lhs(i32 %x) { define <4 x i32> @and_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: and_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: xorps %xmm0, %xmm0 ; CHECK-NEXT: retq %r = and <4 x i32> undef, %x ret <4 x i32> %r @@ -378,6 +387,7 @@ define i32 @or_undef_rhs(i32 %x) { define <4 x i32> @or_undef_rhs_vec(<4 x i32> %x) { ; CHECK-LABEL: or_undef_rhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: pcmpeqd %xmm0, %xmm0 ; CHECK-NEXT: retq %r = or <4 x i32> %x, undef ret <4 x i32> %r @@ -395,6 +405,7 @@ define i32 @or_undef_lhs(i32 %x) { define <4 x i32> @or_undef_lhs_vec(<4 x i32> %x) { ; CHECK-LABEL: or_undef_lhs_vec: ; CHECK: # %bb.0: +; CHECK-NEXT: pcmpeqd %xmm0, %xmm0 ; CHECK-NEXT: retq %r = or <4 x i32> undef, %x ret <4 x i32> %r