PNaCl SIMD: allow element-aligned vector load/store

test/NaCl/PNaClABI/instructions.ll

 ; RUN: not pnacl-abicheck < %s | FileCheck %s
 ; Test instruction opcodes allowed by PNaCl ABI
 
 target datalayout = "e-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-p:32:32:32-v128:32:32"
 target triple = "le32-unknown-nacl"
 
 define internal void @terminators() {
 ; CHECK: ERROR: Function terminators
 ; Terminator instructions
 terminators:
@@ skipping 130 matching lines @@
 }
 
 define internal void @memory() {
 ; CHECK: ERROR: Function memory
 ; Memory operations
   %a1 = alloca i8, i32 4
   %ptr = inttoptr i32 0 to i32*
   %a2 = load i32* %ptr, align 1
   store i32 undef, i32* %ptr, align 1
  ; CHECK-NOT: disallowed
-  %vptr = inttoptr i32 0 to <4 x i32>* ; CHECK-NEXT: disallowed: bad result type: <4 x i32>*
-  %a3 = load <4 x i32>* %vptr, align 16 ; CHECK-NEXT: disallowed: bad pointer: {{.*}} load <4 x i32>*
-  store <4 x i32> undef, <4 x i32>* %vptr, align 16 ; CHECK-NEXT: disallowed: bad pointer: store <4 x i32>
   %a4 = getelementptr { i32, i32}* undef ; CHECK-NEXT: disallowed: bad instruction opcode: {{.*}} getelementptr
   ret void
 }
 
+define internal void @vector_memory() {
+; CHECK: ERROR: Function vector_memory
+; Vector memory operations.

[jvoung (off chromium), 2014/06/10 16:28:33]

Is there anywhere that tests alloca <vectorty>?

[JF, 2014/06/10 18:26:29]

Not that I know of. What are you suggesting? Maybe just an end-to-end test in
NaCl's tests/simd/?

[jvoung (off chromium), 2014/06/10 19:20:31]

Right now we only allow alloca i8, i32 <sz>, align <a>.

I think the replace ptrs w/ ints transformation pass converts the allocas. Does
that handle vector types okay (maybe add to
test/Transforms/NaCl/replace-ptrs-with-ints.ll)?

And yeah, since it all depends on the right transformation passes, pnacl-freeze
and pnacl-thaw being happy, maybe also test in tests/simd/ that the alloca works
(can you take the address of a vector-typed local variable, or something?).

[JF, 2014/06/11 16:04:51]

Done.
Will do, in a separate CL that I'm starting now (since it's in the NaCl repo).

+  %ptr16xi8 = inttoptr i32 0 to <16 x i8>*
+  %ptr8xi16 = inttoptr i32 0 to <8 x i16>*
+  %ptr4xi32 = inttoptr i32 0 to <4 x i32>*
+  %ptr4xfloat = inttoptr i32 0 to <4 x float>*
+
+  %l16xi8 = load <16 x i8>* %ptr16xi8, align 1
+  %l8xi16 = load <8 x i16>* %ptr8xi16, align 2
+  %l4xi32 = load <4 x i32>* %ptr4xi32, align 4
+  %l4xfloat = load <4 x float>* %ptr4xfloat, align 4
+
+  store <16 x i8> undef, <16 x i8>* %ptr16xi8, align 1
+  store <8 x i16> undef, <8 x i16>* %ptr8xi16, align 2
+  store <4 x i32> undef, <4 x i32>* %ptr4xi32, align 4
+  store <4 x float> undef, <4 x float>* %ptr4xfloat, align 4
+
+ ; CHECK-NOT: disallowed
+ ; The following vector types are disallowed.
+  %ptr2xi1 = inttoptr i32 0 to <2 x i1>*         ; CHECK-NEXT: disallowed: bad result type: <2 x i1>*
+  %ptr4xi1 = inttoptr i32 0 to <4 x i1>*         ; CHECK-NEXT: disallowed: bad result type: <4 x i1>*
+  %ptr8xi1 = inttoptr i32 0 to <8 x i1>*         ; CHECK-NEXT: disallowed: bad result type: <8 x i1>*
+  %ptr16xi1 = inttoptr i32 0 to <16 x i1>*       ; CHECK-NEXT: disallowed: bad result type: <16 x i1>*
+  %ptr32xi1 = inttoptr i32 0 to <32 x i1>*       ; CHECK-NEXT: disallowed: bad result type: <32 x i1>*
+  %ptr64xi1 = inttoptr i32 0 to <64 x i1>*       ; CHECK-NEXT: disallowed: bad result type: <64 x i1>*
+  %ptr2xi8 = inttoptr i32 0 to <2 x i8>*         ; CHECK-NEXT: disallowed: bad result type: <2 x i8>*
+  %ptr4xi8 = inttoptr i32 0 to <4 x i8>*         ; CHECK-NEXT: disallowed: bad result type: <4 x i8>*
+  %ptr32xi8 = inttoptr i32 0 to <32 x i8>*       ; CHECK-NEXT: disallowed: bad result type: <32 x i8>*
+  %ptr64xi8 = inttoptr i32 0 to <64 x i8>*       ; CHECK-NEXT: disallowed: bad result type: <64 x i8>*
+  %ptr2xi16 = inttoptr i32 0 to <2 x i16>*       ; CHECK-NEXT: disallowed: bad result type: <2 x i16>*
+  %ptr4xi16 = inttoptr i32 0 to <4 x i16>*       ; CHECK-NEXT: disallowed: bad result type: <4 x i16>*
+  %ptr16xi16 = inttoptr i32 0 to <16 x i16>*     ; CHECK-NEXT: disallowed: bad result type: <16 x i16>*
+  %ptr32xi16 = inttoptr i32 0 to <32 x i16>*     ; CHECK-NEXT: disallowed: bad result type: <32 x i16>*
+  %ptr2xi32 = inttoptr i32 0 to <2 x i32>*       ; CHECK-NEXT: disallowed: bad result type: <2 x i32>*
+  %ptr8xi32 = inttoptr i32 0 to <8 x i32>*       ; CHECK-NEXT: disallowed: bad result type: <8 x i32>*
+  %ptr16xi32 = inttoptr i32 0 to <16 x i32>*     ; CHECK-NEXT: disallowed: bad result type: <16 x i32>*
+  %ptr2xi64 = inttoptr i32 0 to <2 x i64>*       ; CHECK-NEXT: disallowed: bad result type: <2 x i64>*
+  %ptr4xi64 = inttoptr i32 0 to <4 x i64>*       ; CHECK-NEXT: disallowed: bad result type: <4 x i64>*
+  %ptr8xi64 = inttoptr i32 0 to <8 x i64>*       ; CHECK-NEXT: disallowed: bad result type: <8 x i64>*
+  %ptr2xfloat = inttoptr i32 0 to <2 x float>*   ; CHECK-NEXT: disallowed: bad result type: <2 x float>*
+  %ptr8xfloat = inttoptr i32 0 to <8 x float>*   ; CHECK-NEXT: disallowed: bad result type: <8 x float>*
+  %ptr16xfloat = inttoptr i32 0 to <16 x float>* ; CHECK-NEXT: disallowed: bad result type: <16 x float>*
+  %ptr2xdouble = inttoptr i32 0 to <2 x double>* ; CHECK-NEXT: disallowed: bad result type: <2 x double>*
+  %ptr4xdouble = inttoptr i32 0 to <4 x double>* ; CHECK-NEXT: disallowed: bad result type: <4 x double>*
+  %ptr8xdouble = inttoptr i32 0 to <8 x double>* ; CHECK-NEXT: disallowed: bad result type: <8 x double>*
+
+  ; i1 vector pointers are simply disallowed, their alignment is inconsequential.
+  %l4xi1 = load <4 x i1>* %ptr4xi1, align 1    ; CHECK-NEXT: disallowed: bad pointer: %l4xi1 = load <4 x i1>* %ptr4xi1, align 1
+  %l8xi1 = load <8 x i1>* %ptr8xi1, align 1    ; CHECK-NEXT: disallowed: bad pointer: %l8xi1 = load <8 x i1>* %ptr8xi1, align 1
+  %l16xi1 = load <16 x i1>* %ptr16xi1, align 1 ; CHECK-NEXT: disallowed: bad pointer: %l16xi1 = load <16 x i1>* %ptr16xi1, align 1
+
+  store <4 x i1> undef, <4 x i1>* %ptr4xi1, align 1    ; CHECK-NEXT: disallowed: bad pointer: store <4 x i1> undef, <4 x i1>* %ptr4xi1, align 1
+  store <8 x i1> undef, <8 x i1>* %ptr8xi1, align 1    ; CHECK-NEXT: disallowed: bad pointer: store <8 x i1> undef, <8 x i1>* %ptr8xi1, align 1
+  store <16 x i1> undef, <16 x i1>* %ptr16xi1, align 1 ; CHECK-NEXT: disallowed: bad pointer: store <16 x i1> undef, <16 x i1>* %ptr16xi1, align 1
+
+  ; Under- or over-aligned load/store are disallowed.
+  %a1_8xi16 = load <8 x i16>* %ptr8xi16, align 1       ; CHECK-NEXT: disallowed: bad alignment: %a1_8xi16 = load <8 x i16>* %ptr8xi16, align 1
+  %a1_4xi32 = load <4 x i32>* %ptr4xi32, align 1       ; CHECK-NEXT: disallowed: bad alignment: %a1_4xi32 = load <4 x i32>* %ptr4xi32, align 1
+  %a1_4xfloat = load <4 x float>* %ptr4xfloat, align 1 ; CHECK-NEXT: disallowed: bad alignment: %a1_4xfloat = load <4 x float>* %ptr4xfloat, align 1
+
+  %a16_16xi8 = load <16 x i8>* %ptr16xi8, align 16       ; CHECK-NEXT: disallowed: bad alignment: %a16_16xi8 = load <16 x i8>* %ptr16xi8, align 16
+  %a16_8xi16 = load <8 x i16>* %ptr8xi16, align 16       ; CHECK-NEXT: disallowed: bad alignment: %a16_8xi16 = load <8 x i16>* %ptr8xi16, align 16
+  %a16_4xi32 = load <4 x i32>* %ptr4xi32, align 16       ; CHECK-NEXT: disallowed: bad alignment: %a16_4xi32 = load <4 x i32>* %ptr4xi32, align 16
+  %a16_4xfloat = load <4 x float>* %ptr4xfloat, align 16 ; CHECK-NEXT: disallowed: bad alignment: %a16_4xfloat = load <4 x float>* %ptr4xfloat, align 16
+
+  store <8 x i16> undef, <8 x i16>* %ptr8xi16, align 1       ; CHECK-NEXT: disallowed: bad alignment: store <8 x i16> undef, <8 x i16>* %ptr8xi16, align 1
+  store <4 x i32> undef, <4 x i32>* %ptr4xi32, align 1       ; CHECK-NEXT: disallowed: bad alignment: store <4 x i32> undef, <4 x i32>* %ptr4xi32, align 1
+  store <4 x float> undef, <4 x float>* %ptr4xfloat, align 1 ; CHECK-NEXT: disallowed: bad alignment: store <4 x float> undef, <4 x float>* %ptr4xfloat, align 1
+
+  store <16 x i8> undef, <16 x i8>* %ptr16xi8, align 16       ; CHECK-NEXT: disallowed: bad alignment: store <16 x i8> undef, <16 x i8>* %ptr16xi8, align 16
+  store <8 x i16> undef, <8 x i16>* %ptr8xi16, align 16       ; CHECK-NEXT: disallowed: bad alignment: store <8 x i16> undef, <8 x i16>* %ptr8xi16, align 16
+  store <4 x i32> undef, <4 x i32>* %ptr4xi32, align 16       ; CHECK-NEXT: disallowed: bad alignment: store <4 x i32> undef, <4 x i32>* %ptr4xi32, align 16
+  store <4 x float> undef, <4 x float>* %ptr4xfloat, align 16 ; CHECK-NEXT: disallowed: bad alignment: store <4 x float> undef, <4 x float>* %ptr4xfloat, align 16
+
+ ret void
+}
+
 define internal void @atomic() {
 ; CHECK: ERROR: Function atomic
   %a1 = alloca i8, i32 4
   %ptr = inttoptr i32 0 to i32*
  ; CHECK-NOT: disallowed
  ; CHECK: disallowed: atomic load: {{.*}} load atomic
   %a2 = load atomic i32* %ptr seq_cst, align 4
 ; CHECK: disallowed: volatile load: {{.*}} load volatile
   %a3 = load volatile i32* %ptr, align 4
 ; CHECK: disallowed: atomic store: store atomic
@@ skipping 174 matching lines @@
 define internal void @inline_asm() {
 ; CHECK: ERROR: Function inline_asm
   call void asm "foo", ""()
   ret void
 }
 ; CHECK-NOT: disallowed
 ; CHECK: disallowed: inline assembly: call void asm "foo", ""()
 
 ; CHECK-NOT: disallowed
 ; If another check is added, there should be a check-not in between each check