transform: allocate the correct amount of bytes in an alloca

When I wrote the code originally, I didn't know about SetAlignment so I
hacked a way around it by allocating [...]uintptr types. However, this
allocates a few too many bytes in some cases.
This commit changes this to only allocate the space that we actually
need.

The code size effect is mixed, but generally positive. The combined
average is reduced by 0.27% with more programs being reduced in size
than are increasing in size.

transform/allocs.go

@@ -95,19 +95,36 @@ func OptimizeAllocs(mod llvm.Module, printAllocs *regexp.Regexp, logger func(tok
 		}
 		// The pointer value does not escape.
 
+		// Determine the appropriate alignment of the alloca. The size of the
+		// allocation gives us a hint what the alignment should be.
+		var alignment int
+		if size%2 != 0 {
+			alignment = 1
+		} else if size%4 != 0 {
+			alignment = 2
+		} else if size%8 != 0 {
+			alignment = 4
+		} else {
+			alignment = 8
+		}
+		if pointerAlignment := targetData.ABITypeAlignment(i8ptrType); pointerAlignment < alignment {
+			// Use min(alignment, alignof(void*)) as the alignment.
+			alignment = pointerAlignment
+		}
+
 		// Insert alloca in the entry block. Do it here so that mem2reg can
 		// promote it to a SSA value.
 		fn := bitcast.InstructionParent().Parent()
 		builder.SetInsertPointBefore(fn.EntryBasicBlock().FirstInstruction())
-		alignment := targetData.ABITypeAlignment(i8ptrType)
+		allocaType := llvm.ArrayType(mod.Context().Int8Type(), int(size))
-		sizeInWords := (size + uint64(alignment) - 1) / uint64(alignment)
-		allocaType := llvm.ArrayType(mod.Context().IntType(alignment*8), int(sizeInWords))
 		alloca := builder.CreateAlloca(allocaType, "stackalloc.alloca")
+		alloca.SetAlignment(alignment)
 
 		// Zero the allocation inside the block where the value was originally allocated.
 		zero := llvm.ConstNull(alloca.Type().ElementType())
 		builder.SetInsertPointBefore(bitcast)
-		builder.CreateStore(zero, alloca)
+		store := builder.CreateStore(zero, alloca)
+		store.SetAlignment(alignment)
 
 		// Replace heap alloc bitcast with stack alloc bitcast.
 		stackalloc := builder.CreateBitCast(alloca, bitcast.Type(), "stackalloc")

transform/testdata/allocs.out.ll

@@ -6,17 +6,17 @@ target triple = "armv7m-none-eabi"
 declare nonnull i8* @runtime.alloc(i32, i8*)
 
 define void @testInt() {
-  %stackalloc.alloca = alloca [1 x i32], align 4
+  %stackalloc.alloca = alloca [4 x i8], align 4
-  store [1 x i32] zeroinitializer, [1 x i32]* %stackalloc.alloca, align 4
+  store [4 x i8] zeroinitializer, [4 x i8]* %stackalloc.alloca, align 4
-  %stackalloc = bitcast [1 x i32]* %stackalloc.alloca to i32*
+  %stackalloc = bitcast [4 x i8]* %stackalloc.alloca to i32*
   store i32 5, i32* %stackalloc, align 4
   ret void
 }
 
 define i16 @testArray() {
-  %stackalloc.alloca = alloca [2 x i32], align 4
+  %stackalloc.alloca = alloca [6 x i8], align 2
-  store [2 x i32] zeroinitializer, [2 x i32]* %stackalloc.alloca, align 4
+  store [6 x i8] zeroinitializer, [6 x i8]* %stackalloc.alloca, align 2
-  %stackalloc = bitcast [2 x i32]* %stackalloc.alloca to i16*
+  %stackalloc = bitcast [6 x i8]* %stackalloc.alloca to i16*
   %1 = getelementptr i16, i16* %stackalloc, i32 1
   store i16 5, i16* %1, align 2
   %2 = getelementptr i16, i16* %stackalloc, i32 2
@@ -39,9 +39,9 @@ define void @testEscapingCall2() {
 }
 
 define void @testNonEscapingCall() {
-  %stackalloc.alloca = alloca [1 x i32], align 4
+  %stackalloc.alloca = alloca [4 x i8], align 4
-  store [1 x i32] zeroinitializer, [1 x i32]* %stackalloc.alloca, align 4
+  store [4 x i8] zeroinitializer, [4 x i8]* %stackalloc.alloca, align 4
-  %stackalloc = bitcast [1 x i32]* %stackalloc.alloca to i32*
+  %stackalloc = bitcast [4 x i8]* %stackalloc.alloca to i32*
   %1 = call i32* @noescapeIntPtr(i32* %stackalloc)
   ret void
 }
@@ -54,12 +54,12 @@ define i32* @testEscapingReturn() {
 
 define void @testNonEscapingLoop() {
 entry:
-  %stackalloc.alloca = alloca [1 x i32], align 4
+  %stackalloc.alloca = alloca [4 x i8], align 4
   br label %loop
 
 loop:                                             ; preds = %loop, %entry
-  store [1 x i32] zeroinitializer, [1 x i32]* %stackalloc.alloca, align 4
+  store [4 x i8] zeroinitializer, [4 x i8]* %stackalloc.alloca, align 4
-  %stackalloc = bitcast [1 x i32]* %stackalloc.alloca to i32*
+  %stackalloc = bitcast [4 x i8]* %stackalloc.alloca to i32*
   %0 = call i32* @noescapeIntPtr(i32* %stackalloc)
   %1 = icmp eq i32* null, %0
   br i1 %1, label %loop, label %end