[llvm-commits] [llvm] r107373 - in /llvm/trunk: lib/Target/X86/X86FastISel.cpp test/CodeGen/X86/fast-isel-gep.ll

[Dan Gohman]

Author: djg
Date: Wed Jun 30 21:27:15 2010
New Revision: 107373

URL: http://llvm.org/viewvc/llvm-project?rev=107373&view=rev
Log:
Teach X86FastISel to fold constant offsets and scaled indices in
the same address.

Modified:
    llvm/trunk/lib/Target/X86/X86FastISel.cpp
    llvm/trunk/test/CodeGen/X86/fast-isel-gep.ll

Modified: llvm/trunk/lib/Target/X86/X86FastISel.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86FastISel.cpp?rev=107373&r1=107372&r2=107373&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86FastISel.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86FastISel.cpp Wed Jun 30 21:27:15 2010
@@ -423,20 +423,29 @@
         Disp += SL->getElementOffset(Idx);
       } else {
         uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
-        if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
-          // Constant-offset addressing.
-          Disp += CI->getSExtValue() * S;
-        } else if (IndexReg == 0 &&
-                   (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
-                   (S == 1 || S == 2 || S == 4 || S == 8)) {
-          // Scaled-index addressing.
-          Scale = S;
-          IndexReg = getRegForGEPIndex(Op).first;
-          if (IndexReg == 0)
-            return false;
-        } else
-          // Unsupported.
-          goto unsupported_gep;
+        SmallVector<const Value *, 4> Worklist;
+        Worklist.push_back(Op);
+        do {
+          Op = Worklist.pop_back_val();
+          if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
+            // Constant-offset addressing.
+            Disp += CI->getSExtValue() * S;
+          } else if (IndexReg == 0 &&
+                     (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
+                     (S == 1 || S == 2 || S == 4 || S == 8)) {
+            // Scaled-index addressing.
+            Scale = S;
+            IndexReg = getRegForGEPIndex(Op).first;
+            if (IndexReg == 0)
+              return false;
+          } else if (const AddOperator *Add = dyn_cast<AddOperator>(Op)) {
+            // An add. Try to fold both operands.
+            Worklist.push_back(Add->getOperand(0));
+            Worklist.push_back(Add->getOperand(1));
+          } else
+            // Unsupported.
+            goto unsupported_gep;
+        } while (!Worklist.empty());
       }
     }
     // Check for displacement overflow.

Modified: llvm/trunk/test/CodeGen/X86/fast-isel-gep.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/fast-isel-gep.ll?rev=107373&r1=107372&r2=107373&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/fast-isel-gep.ll (original)
+++ llvm/trunk/test/CodeGen/X86/fast-isel-gep.ll Wed Jun 30 21:27:15 2010
@@ -51,3 +51,22 @@
 ; X64:  	ret
 
 }
+
+define double @test4(i64 %x, double* %p) nounwind {
+entry:
+  %x.addr = alloca i64, align 8                   ; <i64*> [#uses=2]
+  %p.addr = alloca double*, align 8               ; <double**> [#uses=2]
+  store i64 %x, i64* %x.addr
+  store double* %p, double** %p.addr
+  %tmp = load i64* %x.addr                        ; <i64> [#uses=1]
+  %add = add nsw i64 %tmp, 16                     ; <i64> [#uses=1]
+  %tmp1 = load double** %p.addr                   ; <double*> [#uses=1]
+  %arrayidx = getelementptr inbounds double* %tmp1, i64 %add ; <double*> [#uses=1]
+  %tmp2 = load double* %arrayidx                  ; <double> [#uses=1]
+  ret double %tmp2
+
+; X32: test4:
+; X32: 128(%e{{.*}},%e{{.*}},8)
+; X64: test4:
+; X64: 128(%r{{.*}},%r{{.*}},8)
+}