[SCCP] Use constant ranges for casts.

For casts with constant range operands, we can use
ConstantRange::castOp.

Reviewers: davide, efriedma, mssimpso

Reviewed By: efriedma

Differential Revision: https://reviews.llvm.org/D71938

diff --git a/llvm/lib/Transforms/Scalar/SCCP.cpp b/llvm/lib/Transforms/Scalar/SCCP.cpp
index ea25ac4..f1e97ce 100644 (file)
--- a/llvm/lib/Transforms/Scalar/SCCP.cpp
+++ b/llvm/lib/Transforms/Scalar/SCCP.cpp
@@ -793,8 +793,8 @@ void SCCPSolver::visitTerminator(Instruction &TI) {
 void SCCPSolver::visitCastInst(CastInst &I) {
   // ResolvedUndefsIn might mark I as overdefined. Bail out, even if we would
   // discover a concrete value later.
-  if (isOverdefined(ValueState[&I]))
-    return (void)markOverdefined(&I);
+  if (ValueState[&I].isOverdefined())
+    return;
 
   ValueLatticeElement OpSt = getValueState(I.getOperand(0));
   if (Constant *OpC = getConstant(OpSt)) {
@@ -804,6 +804,13 @@ void SCCPSolver::visitCastInst(CastInst &I) {
       return;
     // Propagate constant value
     markConstant(&I, C);
+  } else if (OpSt.isConstantRange() && I.getDestTy()->isIntegerTy()) {
+    auto &LV = getValueState(&I);
+    ConstantRange OpRange = OpSt.getConstantRange();
+    Type *DestTy = I.getDestTy();
+    ConstantRange Res =
+        OpRange.castOp(I.getOpcode(), DL.getTypeSizeInBits(DestTy));
+    mergeInValue(LV, &I, ValueLatticeElement::getRange(Res));
   } else if (!OpSt.isUnknownOrUndef())
     markOverdefined(&I);
 }

diff --git a/llvm/test/Transforms/SCCP/ip-ranges-casts.ll b/llvm/test/Transforms/SCCP/ip-ranges-casts.ll
new file mode 100644 (file)
index 0000000..6a4e576
--- /dev/null
+++ b/llvm/test/Transforms/SCCP/ip-ranges-casts.ll
@@ -0,0 +1,296 @@
+; RUN: opt < %s -ipsccp -S | FileCheck %s
+
+; x = [100, 301)
+define internal i1 @f.trunc(i32 %x) {
+; CHECK-LABEL: define internal i1 @f.trunc(i32 %x) {
+; CHECK-NEXT:    %t.1 = trunc i32 %x to i16
+; CHECK-NEXT:    %c.2 = icmp sgt i16 %t.1, 299
+; CHECK-NEXT:    %c.4 = icmp slt i16 %t.1, 101
+; CHECK-NEXT:    %res.1 = add i1 false, %c.2
+; CHECK-NEXT:    %res.2 = add i1 %res.1, false
+; CHECK-NEXT:    %res.3 = add i1 %res.2, %c.4
+; CHECK-NEXT:    %t.2 = trunc i32 %x to i8
+; CHECK-NEXT:    %c.5 = icmp sgt i8 %t.2, 44
+; CHECK-NEXT:    %c.6 = icmp sgt i8 %t.2, 43
+; CHECK-NEXT:    %c.7 = icmp slt i8 %t.2, 100
+; CHECK-NEXT:    %c.8 = icmp slt i8 %t.2, 101
+; CHECK-NEXT:    %res.4 = add i1 %res.3, %c.5
+; CHECK-NEXT:    %res.5 = add i1 %res.4, %c.6
+; CHECK-NEXT:    %res.6 = add i1 %res.5, %c.7
+; CHECK-NEXT:    %res.7 = add i1 %res.6, %c.8
+; CHECK-NEXT:    ret i1 %res.7
+
+  %t.1 = trunc i32 %x to i16
+  %c.1 = icmp sgt i16 %t.1, 300
+  %c.2 = icmp sgt i16 %t.1, 299
+  %c.3 = icmp slt i16 %t.1, 100
+  %c.4 = icmp slt i16 %t.1, 101
+  %res.1 = add i1 %c.1, %c.2
+  %res.2 = add i1 %res.1, %c.3
+  %res.3 = add i1 %res.2, %c.4
+  %t.2 = trunc i32 %x to i8
+  %c.5 = icmp sgt i8 %t.2, 300
+  %c.6 = icmp sgt i8 %t.2, 299
+  %c.7 = icmp slt i8 %t.2, 100
+  %c.8 = icmp slt i8 %t.2, 101
+  %res.4 = add i1 %res.3, %c.5
+  %res.5 = add i1 %res.4, %c.6
+  %res.6 = add i1 %res.5, %c.7
+  %res.7 = add i1 %res.6, %c.8
+  ret i1 %res.7
+}
+
+define i1 @caller1() {
+; CHECK-LABEL:  define i1 @caller1() {
+; CHECK-NEXT:    %call.1 = tail call i1 @f.trunc(i32 100)
+; CHECK-NEXT:    %call.2 = tail call i1 @f.trunc(i32 300)
+; CHECK-NEXT:    %res = and i1 %call.1, %call.2
+; CHECK-NEXT:    ret i1 %res
+;
+  %call.1 = tail call i1 @f.trunc(i32 100)
+  %call.2 = tail call i1 @f.trunc(i32 300)
+  %res = and i1 %call.1, %call.2
+  ret i1 %res
+}
+
+
+; x = [100, 301)
+define internal i1 @f.zext(i32 %x, i32 %y) {
+; CHECK-LABEL: define internal i1 @f.zext(i32 %x, i32 %y) {
+; CHECK-NEXT:    %t.1 = zext i32 %x to i64
+; CHECK-NEXT:    %c.2 = icmp sgt i64 %t.1, 299
+; CHECK-NEXT:    %c.4 = icmp slt i64 %t.1, 101
+; CHECK-NEXT:    %res.1 = add i1 false, %c.2
+; CHECK-NEXT:    %res.2 = add i1 %res.1, false
+; CHECK-NEXT:    %res.3 = add i1 %res.2, %c.4
+; CHECK-NEXT:    %t.2 = zext i32 %y to i64
+; CHECK-NEXT:    %c.5 = icmp sgt i64 %t.2, 300
+; CHECK-NEXT:    %c.6 = icmp sgt i64 %t.2, 299
+; CHECK-NEXT:    %c.8 = icmp slt i64 %t.2, 1
+; CHECK-NEXT:    %res.4 = add i1 %res.3, %c.5
+; CHECK-NEXT:    %res.5 = add i1 %res.4, %c.6
+; CHECK-NEXT:    %res.6 = add i1 %res.5, false
+; CHECK-NEXT:    %res.7 = add i1 %res.6, %c.8
+; CHECK-NEXT:    ret i1 %res.7
+
+  %t.1 = zext i32 %x to i64
+  %c.1 = icmp sgt i64 %t.1, 300
+  %c.2 = icmp sgt i64 %t.1, 299
+  %c.3 = icmp slt i64 %t.1, 100
+  %c.4 = icmp slt i64 %t.1, 101
+  %res.1 = add i1 %c.1, %c.2
+  %res.2 = add i1 %res.1, %c.3
+  %res.3 = add i1 %res.2, %c.4
+  %t.2 = zext i32 %y to i64
+  %c.5 = icmp sgt i64 %t.2, 300
+  %c.6 = icmp sgt i64 %t.2, 299
+  %c.7 = icmp slt i64 %t.2, 0
+  %c.8 = icmp slt i64 %t.2, 1
+  %res.4 = add i1 %res.3, %c.5
+  %res.5 = add i1 %res.4, %c.6
+  %res.6 = add i1 %res.5, %c.7
+  %res.7 = add i1 %res.6, %c.8
+  ret i1 %res.7
+}
+
+define i1 @caller.zext() {
+; CHECK-LABEL:  define i1 @caller.zext() {
+; CHECK-NEXT:    %call.1 = tail call i1 @f.zext(i32 100, i32 -120)
+; CHECK-NEXT:    %call.2 = tail call i1 @f.zext(i32 300, i32 900)
+; CHECK-NEXT:    %res = and i1 %call.1, %call.2
+; CHECK-NEXT:    ret i1 %res
+;
+  %call.1 = tail call i1 @f.zext(i32 100, i32 -120)
+  %call.2 = tail call i1 @f.zext(i32 300, i32 900)
+  %res = and i1 %call.1, %call.2
+  ret i1 %res
+}
+
+; x = [100, 301)
+define internal i1 @f.sext(i32 %x, i32 %y) {
+; CHECK-LABEL: define internal i1 @f.sext(i32 %x, i32 %y) {
+; CHECK-NEXT:    %t.1 = sext i32 %x to i64
+; CHECK-NEXT:    %c.2 = icmp sgt i64 %t.1, 299
+; CHECK-NEXT:    %c.4 = icmp slt i64 %t.1, 101
+; CHECK-NEXT:    %res.1 = add i1 false, %c.2
+; CHECK-NEXT:    %res.2 = add i1 %res.1, false
+; CHECK-NEXT:    %res.3 = add i1 %res.2, %c.4
+; CHECK-NEXT:    %t.2 = sext i32 %y to i64
+; CHECK-NEXT:    %c.6 = icmp sgt i64 %t.2, 899
+; CHECK-NEXT:    %c.8 = icmp slt i64 %t.2, -119
+; CHECK-NEXT:    %res.4 = add i1 %res.3, false
+; CHECK-NEXT:    %res.5 = add i1 %res.4, %c.6
+; CHECK-NEXT:    %res.6 = add i1 %res.5, false
+; CHECK-NEXT:    %res.7 = add i1 %res.6, %c.8
+; CHECK-NEXT:    ret i1 %res.7
+;
+  %t.1 = sext i32 %x to i64
+  %c.1 = icmp sgt i64 %t.1, 300
+  %c.2 = icmp sgt i64 %t.1, 299
+  %c.3 = icmp slt i64 %t.1, 100
+  %c.4 = icmp slt i64 %t.1, 101
+  %res.1 = add i1 %c.1, %c.2
+  %res.2 = add i1 %res.1, %c.3
+  %res.3 = add i1 %res.2, %c.4
+  %t.2 = sext i32 %y to i64
+  %c.5 = icmp sgt i64 %t.2, 900
+  %c.6 = icmp sgt i64 %t.2, 899
+  %c.7 = icmp slt i64 %t.2, -120
+  %c.8 = icmp slt i64 %t.2, -119
+  %res.4 = add i1 %res.3, %c.5
+  %res.5 = add i1 %res.4, %c.6
+  %res.6 = add i1 %res.5, %c.7
+  %res.7 = add i1 %res.6, %c.8
+  ret i1 %res.7
+}
+
+define i1 @caller.sext() {
+; CHECK-LABEL:  define i1 @caller.sext() {
+; CHECK-NEXT:    %call.1 = tail call i1 @f.sext(i32 100, i32 -120)
+; CHECK-NEXT:    %call.2 = tail call i1 @f.sext(i32 300, i32 900)
+; CHECK-NEXT:    %res = and i1 %call.1, %call.2
+; CHECK-NEXT:    ret i1 %res
+;
+  %call.1 = tail call i1 @f.sext(i32 100, i32 -120)
+  %call.2 = tail call i1 @f.sext(i32 300, i32 900)
+  %res = and i1 %call.1, %call.2
+  ret i1 %res
+}
+
+; There's nothing we can do besides going to the full range or overdefined.
+define internal i1 @f.fptosi(i32 %x) {
+; CHECK-LABEL: define internal i1 @f.fptosi(i32 %x) {
+; CHECK-NEXT:    %to.double = sitofp i32 %x to double
+; CHECK-NEXT:    %add = fadd double 0.000000e+00, %to.double
+; CHECK-NEXT:    %to.i32 = fptosi double %add to i32
+; CHECK-NEXT:    %c.1 = icmp sgt i32 %to.i32, 300
+; CHECK-NEXT:    %c.2 = icmp sgt i32 %to.i32, 299
+; CHECK-NEXT:    %c.3 = icmp slt i32 %to.i32, 100
+; CHECK-NEXT:    %c.4 = icmp slt i32 %to.i32, 101
+; CHECK-NEXT:    %res.1 = add i1 %c.1, %c.2
+; CHECK-NEXT:    %res.2 = add i1 %res.1, %c.3
+; CHECK-NEXT:    %res.3 = add i1 %res.2, %c.4
+; CHECK-NEXT:    ret i1 %res.3
+;
+  %to.double = sitofp i32 %x to double
+  %add = fadd double 0.000000e+00, %to.double
+  %to.i32 = fptosi double %add to i32
+  %c.1 = icmp sgt i32 %to.i32, 300
+  %c.2 = icmp sgt i32 %to.i32, 299
+  %c.3 = icmp slt i32 %to.i32, 100
+  %c.4 = icmp slt i32 %to.i32, 101
+  %res.1 = add i1 %c.1, %c.2
+  %res.2 = add i1 %res.1, %c.3
+  %res.3 = add i1 %res.2, %c.4
+  ret i1 %res.3
+}
+
+define i1 @caller.fptosi() {
+; CHECK-LABEL:  define i1 @caller.fptosi() {
+; CHECK-NEXT:    %call.1 = tail call i1 @f.fptosi(i32 100)
+; CHECK-NEXT:    %call.2 = tail call i1 @f.fptosi(i32 300)
+; CHECK-NEXT:    %res = and i1 %call.1, %call.2
+; CHECK-NEXT:    ret i1 %res
+;
+  %call.1 = tail call i1 @f.fptosi(i32 100)
+  %call.2 = tail call i1 @f.fptosi(i32 300)
+  %res = and i1 %call.1, %call.2
+  ret i1 %res
+}
+
+; There's nothing we can do besides going to the full range or overdefined.
+define internal i1 @f.fpext(i16 %x) {
+; CHECK-LABEL: define internal i1 @f.fpext(i16 %x) {
+; CHECK-NEXT:    %to.float = sitofp i16 %x to float
+; CHECK-NEXT:    %to.double = fpext float %to.float to double
+; CHECK-NEXT:    %to.i64 = fptoui float %to.float to i64
+; CHECK-NEXT:    %c.1 = icmp sgt i64 %to.i64, 300
+; CHECK-NEXT:    %c.2 = icmp sgt i64 %to.i64, 299
+; CHECK-NEXT:    %c.3 = icmp slt i64 %to.i64, 100
+; CHECK-NEXT:    %c.4 = icmp slt i64 %to.i64, 101
+; CHECK-NEXT:    %res.1 = add i1 %c.1, %c.2
+; CHECK-NEXT:    %res.2 = add i1 %res.1, %c.3
+; CHECK-NEXT:    %res.3 = add i1 %res.2, %c.4
+; CHECK-NEXT:    ret i1 %res.3
+;
+  %to.float = sitofp i16 %x to float
+  %to.double = fpext float %to.float  to double
+  %to.i64= fptoui float %to.float to i64
+  %c.1 = icmp sgt i64 %to.i64, 300
+  %c.2 = icmp sgt i64 %to.i64, 299
+  %c.3 = icmp slt i64 %to.i64, 100
+  %c.4 = icmp slt i64 %to.i64, 101
+  %res.1 = add i1 %c.1, %c.2
+  %res.2 = add i1 %res.1, %c.3
+  %res.3 = add i1 %res.2, %c.4
+  ret i1 %res.3
+}
+
+; There's nothing we can do besides going to the full range or overdefined.
+define i1 @caller.fpext() {
+; CHECK-LABEL:  define i1 @caller.fpext() {
+; CHECK-NEXT:    %call.1 = tail call i1 @f.fpext(i16 100)
+; CHECK-NEXT:    %call.2 = tail call i1 @f.fpext(i16 300)
+; CHECK-NEXT:    %res = and i1 %call.1, %call.2
+; CHECK-NEXT:    ret i1 %res
+;
+  %call.1 = tail call i1 @f.fpext(i16 100)
+  %call.2 = tail call i1 @f.fpext(i16 300)
+  %res = and i1 %call.1, %call.2
+  ret i1 %res
+}
+
+; There's nothing we can do besides going to the full range or overdefined.
+define internal i1 @f.inttoptr.ptrtoint(i64 %x) {
+; CHECK-LABEL: define internal i1 @f.inttoptr.ptrtoint(i64 %x) {
+; CHECK-NEXT:    %to.ptr = inttoptr i64 %x to i8*
+; CHECK-NEXT:    %to.i64 = ptrtoint i8* %to.ptr to i64
+; CHECK-NEXT:    %c.1 = icmp sgt i64 %to.i64, 300
+; CHECK-NEXT:    %c.2 = icmp sgt i64 %to.i64, 299
+; CHECK-NEXT:    %c.3 = icmp slt i64 %to.i64, 100
+; CHECK-NEXT:    %c.4 = icmp slt i64 %to.i64, 101
+; CHECK-NEXT:    %res.1 = add i1 %c.1, %c.2
+; CHECK-NEXT:    %res.2 = add i1 %res.1, %c.3
+; CHECK-NEXT:    %res.3 = add i1 %res.2, %c.4
+; CHECK-NEXT:    ret i1 %res.3
+;
+  %to.ptr = inttoptr i64 %x to i8*
+  %to.i64 = ptrtoint i8* %to.ptr to i64
+  %c.1 = icmp sgt i64 %to.i64, 300
+  %c.2 = icmp sgt i64 %to.i64, 299
+  %c.3 = icmp slt i64 %to.i64, 100
+  %c.4 = icmp slt i64 %to.i64, 101
+  %res.1 = add i1 %c.1, %c.2
+  %res.2 = add i1 %res.1, %c.3
+  %res.3 = add i1 %res.2, %c.4
+  ret i1 %res.3
+}
+
+define i1 @caller.inttoptr.ptrtoint() {
+; CHECK-LABEL:  define i1 @caller.inttoptr.ptrtoint() {
+; CHECK-NEXT:    %call.1 = tail call i1 @f.inttoptr.ptrtoint(i64 100)
+; CHECK-NEXT:    %call.2 = tail call i1 @f.inttoptr.ptrtoint(i64 300)
+; CHECK-NEXT:    %res = and i1 %call.1, %call.2
+; CHECK-NEXT:    ret i1 %res
+;
+  %call.1 = tail call i1 @f.inttoptr.ptrtoint(i64 100)
+  %call.2 = tail call i1 @f.inttoptr.ptrtoint(i64 300)
+  %res = and i1 %call.1, %call.2
+  ret i1 %res
+}
+
+; Make sure we do not create constant ranges for int to fp casts.
+define i1 @int_range_to_double_cast(i32 %a) {
+; CHECK-LABEL: define i1 @int_range_to_double_cast(i32 %a)
+; CHECK-NEXT:    %r = and i32 %a, 255
+; CHECK-NEXT:    %tmp4 = sitofp i32 %r to double
+; CHECK-NEXT:    %tmp10 = fadd double 0.000000e+00, %tmp4
+; CHECK-NEXT:    %tmp11 = fcmp olt double %tmp4, %tmp10
+; CHECK-NEXT:    ret i1 %tmp11
+;
+  %r = and i32 %a, 255
+  %tmp4 = sitofp i32 %r to double
+  %tmp10 = fadd double 0.000000e+00, %tmp4
+  %tmp11 = fcmp olt double %tmp4, %tmp10
+  ret i1 %tmp11
+}