[x86 fast-isel] Materialize allocas with the correct-sized lea for ILP32

[dschuff]

Summary

X86FastISel::fastMaterializeAlloca was incorrectly conditioning its
opcode selection on subtarget bitness rather than pointer size.

[jvoung]

test/CodeGen/X86/x32-alloca.ll

2 ↗

(On Diff #15817)

Add "-fast-isel-abort" to make sure it's actually using fast-isel for this codegen?

A bunch of the other fast-isel tests have a "fast-isel-" prefix for the filename. Perhaps it would be good to stick with that.

I couldn't find any existing test for this for non ILP32 x86-64 code, so might as well test the non-ILP32 x86-64 code path too. Maybe the test can be merged into fast-isel-x86-64.ll (but only have a check-prefix of this case for x32 for now)?

6 ↗

(On Diff #15817)

could probably just make bar a "declare" and not have to codegen it

[rnk]

lgtm

I suspect we'll see more issues like this.

test/CodeGen/X86/x32-alloca.ll

17 ↗

(On Diff #15817)

please fix

[dschuff]

• Review comments

test/CodeGen/X86/x32-alloca.ll

2 ↗

(On Diff #15817)

Added -fast-isel-abort and renamed; unfortunately fast-isel-x86_64.ll can't be codegen'd for x32/nacl currently (it throws an assert). Fixing that would be beyond the scope of this change but I added the same test case to that file for LP64.

6 ↗

(On Diff #15817)

Done.

17 ↗

(On Diff #15817)

Done.

[dschuff]

• really fix newline

[dschuff]

Closed by commit rL221386 (authored by @dschuff).

lib/Target/X86/X86FastISel.cpp

unsigned X86FastISel::fastMaterializeAlloca(const AllocaInst *C) {
   // X86SelectAddrss, and targetMaterializeAlloca.
   if (!FuncInfo.StaticAllocaMap.count(C))
     return 0;
   assert(C->isStaticAlloca() && "dynamic alloca in the static alloca map?");
 
   X86AddressMode AM;
   if (!X86SelectAddress(C, AM))
     return 0;
-  unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
+  unsigned Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r;
   const TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
   unsigned ResultReg = createResultReg(RC);
   addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                          TII.get(Opc), ResultReg), AM);
   return ResultReg;
 }
 
 unsigned X86FastISel::fastMaterializeFloatZero(const ConstantFP *CF) {

test/CodeGen/X86/fast-isel-x32.ll

+; RUN: llc < %s -mtriple=x86_64-linux-gnux32 -fast-isel -fast-isel-abort -regalloc=fast | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-nacl -fast-isel -fast-isel-abort -regalloc=fast | FileCheck %s
+
+; Test that alloca addresses are materialized with the right size instruction.
+
+declare void @bar(i32* %arg)
+
+; CHECK-LABEL: @foo
+define void @foo() {
+  %a = alloca i32
+; CHECK: leal {{.*}}, %edi
+  call void @bar(i32* %a)
+  ret void
+}

test/CodeGen/X86/fast-isel-x86-64.ll

define void @test23(i8* noalias sret %result) {
   ret void
 ; CHECK-LABEL: test23:
 ; CHECK: call
 ; CHECK: movq  %rdi, %rax
 ; CHECK: ret
 }
 
 declare i8* @foo23()
+
+declare void @takesi32ptr(i32* %arg)
+
+; CHECK-LABEL: allocamaterialize
+define void @allocamaterialize() {
+  %a = alloca i32
+; CHECK: leaq {{.*}}, %rdi
+  call void @takesi32ptr(i32* %a)
+  ret void
+}