[llvm-commits] [llvm] r129990 - in /llvm/trunk: lib/Target/X86/X86ISelDAGToDAG.cpp test/CodeGen/X86/narrow-shl-cst.ll
Benjamin Kramer
benny.kra at googlemail.com
Fri Apr 22 08:30:41 PDT 2011
Author: d0k
Date: Fri Apr 22 10:30:40 2011
New Revision: 129990
URL: http://llvm.org/viewvc/llvm-project?rev=129990&view=rev
Log:
X86: Try to use a smaller encoding by transforming (X << C1) & C2 into (X & (C2 >> C1)) & C1. (Part of PR5039)
This tends to happen a lot with bitfield code generated by clang. A simple example for x86_64 is
uint64_t foo(uint64_t x) { return (x&1) << 42; }
which used to compile into bloated code:
shlq $42, %rdi ## encoding: [0x48,0xc1,0xe7,0x2a]
movabsq $4398046511104, %rax ## encoding: [0x48,0xb8,0x00,0x00,0x00,0x00,0x00,0x04,0x00,0x00]
andq %rdi, %rax ## encoding: [0x48,0x21,0xf8]
ret ## encoding: [0xc3]
with this patch we can fold the immediate into the and:
andq $1, %rdi ## encoding: [0x48,0x83,0xe7,0x01]
movq %rdi, %rax ## encoding: [0x48,0x89,0xf8]
shlq $42, %rax ## encoding: [0x48,0xc1,0xe0,0x2a]
ret ## encoding: [0xc3]
It's possible to save another byte by using 'andl' instead of 'andq' but I currently see no way of doing
that without making this code even more complicated. See the TODOs in the code.
Added:
llvm/trunk/test/CodeGen/X86/narrow-shl-cst.ll
Modified:
llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp
Modified: llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp?rev=129990&r1=129989&r2=129990&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp Fri Apr 22 10:30:40 2011
@@ -1580,6 +1580,81 @@
return RetVal;
break;
}
+ case ISD::AND:
+ case ISD::OR:
+ case ISD::XOR: {
+ // For operations of the form (x << C1) op C2, check if we can use a smaller
+ // encoding for C2 by transforming it into (x op (C2>>C1)) << C1.
+ SDValue N0 = Node->getOperand(0);
+ SDValue N1 = Node->getOperand(1);
+
+ if (N0->getOpcode() != ISD::SHL || !N0->hasOneUse())
+ break;
+
+ // i8 is unshrinkable, i16 should be promoted to i32.
+ if (NVT != MVT::i32 && NVT != MVT::i64)
+ break;
+
+ ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(N1);
+ ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(N0->getOperand(1));
+ if (!Cst || !ShlCst)
+ break;
+
+ int64_t Val = Cst->getSExtValue();
+ uint64_t ShlVal = ShlCst->getZExtValue();
+
+ // Make sure that we don't change the operation by removing bits.
+ // This only matters for OR and XOR, AND is unaffected.
+ if (Opcode != ISD::AND && ((Val >> ShlVal) << ShlVal) != Val)
+ break;
+
+ unsigned ShlOp, Op;
+ EVT CstVT = NVT;
+
+ // Check the minimum bitwidth for the new constant.
+ // TODO: AND32ri is the same as AND64ri32 with zext imm.
+ // TODO: MOV32ri+OR64r is cheaper than MOV64ri64+OR64rr
+ // TODO: Using 16 and 8 bit operations is also possible for or32 & xor32.
+ if (!isInt<8>(Val) && isInt<8>(Val >> ShlVal))
+ CstVT = MVT::i8;
+ else if (!isInt<32>(Val) && isInt<32>(Val >> ShlVal))
+ CstVT = MVT::i32;
+
+ // Bail if there is no smaller encoding.
+ if (NVT == CstVT)
+ break;
+
+ switch (NVT.getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("Unsupported VT!");
+ case MVT::i32:
+ assert(CstVT == MVT::i8);
+ ShlOp = X86::SHL32ri;
+
+ switch (Opcode) {
+ case ISD::AND: Op = X86::AND32ri8; break;
+ case ISD::OR: Op = X86::OR32ri8; break;
+ case ISD::XOR: Op = X86::XOR32ri8; break;
+ }
+ break;
+ case MVT::i64:
+ assert(CstVT == MVT::i8 || CstVT == MVT::i32);
+ ShlOp = X86::SHL64ri;
+
+ switch (Opcode) {
+ case ISD::AND: Op = CstVT==MVT::i8? X86::AND64ri8 : X86::AND64ri32; break;
+ case ISD::OR: Op = CstVT==MVT::i8? X86::OR64ri8 : X86::OR64ri32; break;
+ case ISD::XOR: Op = CstVT==MVT::i8? X86::XOR64ri8 : X86::XOR64ri32; break;
+ }
+ break;
+ }
+
+ // Emit the smaller op and the shift.
+ SDValue NewCst = CurDAG->getTargetConstant(Val >> ShlVal, CstVT);
+ SDNode *New = CurDAG->getMachineNode(Op, dl, NVT, N0->getOperand(0),NewCst);
+ return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
+ getI8Imm(ShlVal));
+ break;
+ }
case X86ISD::UMUL: {
SDValue N0 = Node->getOperand(0);
SDValue N1 = Node->getOperand(1);
Added: llvm/trunk/test/CodeGen/X86/narrow-shl-cst.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/narrow-shl-cst.ll?rev=129990&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/narrow-shl-cst.ll (added)
+++ llvm/trunk/test/CodeGen/X86/narrow-shl-cst.ll Fri Apr 22 10:30:40 2011
@@ -0,0 +1,101 @@
+; RUN: llc < %s -march=x86-64 | FileCheck %s
+; PR5039
+
+define i32 @test1(i32 %x) nounwind {
+ %and = shl i32 %x, 10
+ %shl = and i32 %and, 31744
+ ret i32 %shl
+; CHECK: test1:
+; CHECK: andl $31
+; CHECK: shll $10
+}
+
+define i32 @test2(i32 %x) nounwind {
+ %or = shl i32 %x, 10
+ %shl = or i32 %or, 31744
+ ret i32 %shl
+; CHECK: test2:
+; CHECK: orl $31
+; CHECK: shll $10
+}
+
+define i32 @test3(i32 %x) nounwind {
+ %xor = shl i32 %x, 10
+ %shl = xor i32 %xor, 31744
+ ret i32 %shl
+; CHECK: test3:
+; CHECK: xorl $31
+; CHECK: shll $10
+}
+
+define i64 @test4(i64 %x) nounwind {
+ %and = shl i64 %x, 40
+ %shl = and i64 %and, 264982302294016
+ ret i64 %shl
+; CHECK: test4:
+; CHECK: andq $241
+; CHECK: shlq $40
+}
+
+define i64 @test5(i64 %x) nounwind {
+ %and = shl i64 %x, 40
+ %shl = and i64 %and, 34084860461056
+ ret i64 %shl
+; CHECK: test5:
+; CHECK: andq $31
+; CHECK: shlq $40
+}
+
+define i64 @test6(i64 %x) nounwind {
+ %and = shl i64 %x, 32
+ %shl = and i64 %and, -281474976710656
+ ret i64 %shl
+; CHECK: test6:
+; CHECK: andq $-65536
+; CHECK: shlq $32
+}
+
+define i64 @test7(i64 %x) nounwind {
+ %or = shl i64 %x, 40
+ %shl = or i64 %or, 264982302294016
+ ret i64 %shl
+; CHECK: test7:
+; CHECK: orq $241
+; CHECK: shlq $40
+}
+
+define i64 @test8(i64 %x) nounwind {
+ %or = shl i64 %x, 40
+ %shl = or i64 %or, 34084860461056
+ ret i64 %shl
+; CHECK: test8:
+; CHECK: orq $31
+; CHECK: shlq $40
+}
+
+define i64 @test9(i64 %x) nounwind {
+ %xor = shl i64 %x, 40
+ %shl = xor i64 %xor, 264982302294016
+ ret i64 %shl
+; CHECK: test9:
+; CHECK: orq $241
+; CHECK: shlq $40
+}
+
+define i64 @test10(i64 %x) nounwind {
+ %xor = shl i64 %x, 40
+ %shl = xor i64 %xor, 34084860461056
+ ret i64 %shl
+; CHECK: test10:
+; CHECK: xorq $31
+; CHECK: shlq $40
+}
+
+define i64 @test11(i64 %x) nounwind {
+ %xor = shl i64 %x, 33
+ %shl = xor i64 %xor, -562949953421312
+ ret i64 %shl
+; CHECK: test11:
+; CHECK: xorq $-65536
+; CHECK: shlq $33
+}
More information about the llvm-commits
mailing list