PNaCl: Add support for GCC/LLVM vector extensions

test/Transforms/NaCl/globalize-constant-vectors.ll

+; RUN: opt -globalize-constant-vectors %s -S | FileCheck %s
+; RUN: opt -globalize-constant-vectors %s -S | FileCheck -check-prefix=C4xi1 %s
+; RUN: opt -globalize-constant-vectors %s -S | FileCheck -check-prefix=C8xi1 %s
+; RUN: opt -globalize-constant-vectors %s -S | FileCheck -check-prefix=C16xi1 %s
+; RUN: opt -globalize-constant-vectors %s -S | FileCheck -check-prefix=C16xi8 %s
+; RUN: opt -globalize-constant-vectors %s -S | FileCheck -check-prefix=C8xi16 %s
+; RUN: opt -globalize-constant-vectors %s -S | FileCheck -check-prefix=C4xi32 %s
+; RUN: opt -globalize-constant-vectors %s -S | FileCheck -check-prefix=C4xfloat %s
+; RUN: opt -globalize-constant-vectors %s -S | FileCheck -check-prefix=Cbranch %s
+; RUN: opt -globalize-constant-vectors %s -S | FileCheck -check-prefix=Cduplicate %s
+
+; Run the test once per function so that each check can look at its
+; globals as well as its function.
+
+; The datalayout is needed to determine the alignment of the globals.
+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"
+
+; Globals shouldn't get globalized.
+; CHECK: @global_should_stay_untouched = internal constant <4 x i32> <i32 1337, i32 0, i32 0, i32 0>
+@global_should_stay_untouched = internal constant <4 x i32> <i32 1337, i32 0, i32 0, i32 0>
+
+; 4xi1 vectors should get globalized.
+define void @test4xi1(<4 x i1> %in) {
+  %ft0 = and <4 x i1> %in, <i1 false, i1 true, i1 false, i1 true>
+  %ft1 = and <4 x i1> <i1 true, i1 false, i1 true, i1 false>, %in
+  ret void
+}
+; C4xi1: @[[C1:[_a-z0-9]+]] = internal constant <4 x i1> <i1 false, i1 true, i1 false, i1 true>, align 4
+; C4xi1: @[[C2:[_a-z0-9]+]] = internal constant <4 x i1> <i1 true, i1 false, i1 true, i1 false>, align 4
+; C4xi1: define void @test4xi1(<4 x i1> %in) {
+; C4xi1-NEXT: %[[M1:[_a-z0-9]+]] = load <4 x i1>* @[[C1]], align 4
+; C4xi1-NEXT: %[[M2:[_a-z0-9]+]] = load <4 x i1>* @[[C2]], align 4
+; C4xi1-NEXT: %ft0 = and <4 x i1> %in, %[[M1]]
+; C4xi1-NEXT: %ft1 = and <4 x i1> %[[M2]], %in
+; C4xi1-NEXT: ret void
+
+; 8xi1 vectors should get globalized.
+define void @test8xi1(<8 x i1> %in) {
+  %ft0 = and <8 x i1> %in, <i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true>
+  %ft1 = and <8 x i1> <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false>, %in
+  ret void
+}
+; C8xi1: @[[C1:[_a-z0-9]+]] = internal constant <8 x i1> <i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true>, align 8
+; C8xi1: @[[C2:[_a-z0-9]+]] = internal constant <8 x i1> <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false>, align 8
+; C8xi1: define void @test8xi1(<8 x i1> %in) {
+; C8xi1-NEXT: %[[M1:[_a-z0-9]+]] = load <8 x i1>* @[[C1]], align 8
+; C8xi1-NEXT: %[[M2:[_a-z0-9]+]] = load <8 x i1>* @[[C2]], align 8
+; C8xi1-NEXT: %ft0 = and <8 x i1> %in, %[[M1]]
+; C8xi1-NEXT: %ft1 = and <8 x i1> %[[M2]], %in
+; C8xi1-NEXT: ret void
+
+; 16xi1 vectors should get globalized.
+define void @test16xi1(<16 x i1> %in) {
+  %ft0 = and <16 x i1> %in, <i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true>
+  %ft1 = and <16 x i1> <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false>, %in
+  ret void
+}
+; C16xi1: @[[C1:[_a-z0-9]+]] = internal constant <16 x i1> <i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true>, align 16
+; C16xi1: @[[C2:[_a-z0-9]+]] = internal constant <16 x i1> <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false>, align 16
+; C16xi1: define void @test16xi1(<16 x i1> %in) {
+; C16xi1-NEXT: %[[M1:[_a-z0-9]+]] = load <16 x i1>* @[[C1]], align 16
+; C16xi1-NEXT: %[[M2:[_a-z0-9]+]] = load <16 x i1>* @[[C2]], align 16
+; C16xi1-NEXT: %ft0 = and <16 x i1> %in, %[[M1]]
+; C16xi1-NEXT: %ft1 = and <16 x i1> %[[M2]], %in
+; C16xi1-NEXT: ret void
+
+; 16xi8 vectors should get globalized.
+define void @test16xi8(<16 x i8> %in) {
+  %nonsquares = add <16 x i8> %in, <i8 2, i8 3, i8 5, i8 6, i8 7, i8 8, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15, i8 17, i8 18, i8 19, i8 20>
+  %sort = add <16 x i8> <i8 0, i8 1, i8 3, i8 5, i8 9, i8 11, i8 14, i8 17, i8 25, i8 27, i8 30, i8 33, i8 38, i8 41, i8 45, i8 49>, %in
+  ret void
+}
+; C16xi8: @[[C1:[_a-z0-9]+]] = internal constant <16 x i8> <i8 2, i8 3, i8 5, i8 6, i8 7, i8 8, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15, i8 17, i8 18, i8 19, i8 20>, align 4
+; C16xi8: @[[C2:[_a-z0-9]+]] = internal constant <16 x i8> <i8 0, i8 1, i8 3, i8 5, i8 9, i8 11, i8 14, i8 17, i8 25, i8 27, i8 30, i8 33, i8 38, i8 41, i8 45, i8 49>, align 4
+; C16xi8: define void @test16xi8(<16 x i8> %in) {
+; C16xi8-NEXT: %[[M1:[_a-z0-9]+]] = load <16 x i8>* @[[C1]], align 4
+; C16xi8-NEXT: %[[M2:[_a-z0-9]+]] = load <16 x i8>* @[[C2]], align 4
+; C16xi8-NEXT: %nonsquares = add <16 x i8> %in, %[[M1]]
+; C16xi8-NEXT: %sort = add <16 x i8> %[[M2]], %in
+; C16xi8-NEXT: ret void
+
+; 8xi16 vectors should get globalized.
+define void @test8xi16(<8 x i16> %in) {
+  %fib = add <8 x i16> %in, <i16 0, i16 1, i16 1, i16 2, i16 3, i16 5, i16 8, i16 13>
+  %answer = add <8 x i16> <i16 42, i16 42, i16 42, i16 42, i16 42, i16 42, i16 42, i16 42>, %in
+  ret void
+}
+; C8xi16: @[[C1:[_a-z0-9]+]] = internal constant <8 x i16> <i16 0, i16 1, i16 1, i16 2, i16 3, i16 5, i16 8, i16 13>, align 4
+; C8xi16: @[[C2:[_a-z0-9]+]] = internal constant <8 x i16> <i16 42, i16 42, i16 42, i16 42, i16 42, i16 42, i16 42, i16 42>, align 4
+; C8xi16: define void @test8xi16(<8 x i16> %in) {
+; C8xi16-NEXT: %[[M1:[_a-z0-9]+]] = load <8 x i16>* @[[C1]], align 4
+; C8xi16-NEXT: %[[M2:[_a-z0-9]+]] = load <8 x i16>* @[[C2]], align 4
+; C8xi16-NEXT: %fib = add <8 x i16> %in, %[[M1]]
+; C8xi16-NEXT: %answer = add <8 x i16> %[[M2]], %in
+; C8xi16-NEXT: ret void
+
+; 4xi32 vectors should get globalized.
+define void @test4xi32(<4 x i32> %in) {
+  %tetrahedral = add <4 x i32> %in, <i32 1, i32 4, i32 10, i32 20>
+  %serauqs = add <4 x i32> <i32 1, i32 4, i32 9, i32 61>, %in
+  ret void
+}
+; C4xi32: @[[C1:[_a-z0-9]+]] = internal constant <4 x i32> <i32 1, i32 4, i32 10, i32 20>, align 4
+; C4xi32: @[[C2:[_a-z0-9]+]] = internal constant <4 x i32> <i32 1, i32 4, i32 9, i32 61>, align 4
+; C4xi32: define void @test4xi32(<4 x i32> %in) {
+; C4xi32-NEXT: %[[M1:[_a-z0-9]+]] = load <4 x i32>* @[[C1]], align 4
+; C4xi32-NEXT: %[[M2:[_a-z0-9]+]] = load <4 x i32>* @[[C2]], align 4
+; C4xi32-NEXT: %tetrahedral = add <4 x i32> %in, %[[M1]]
+; C4xi32-NEXT: %serauqs = add <4 x i32> %[[M2]], %in
+; C4xi32-NEXT: ret void
+
+; 4xfloat vectors should get globalized.
+define void @test4xfloat(<4 x float> %in) {
+  %polyhex = fadd <4 x float> %in, <float 1., float 1., float 3., float 7.>
+  %poset = fadd <4 x float> <float 1., float 1., float 3., float 19.>, %in
+  ret void
+}
+; C4xfloat: @[[C1:[_a-z0-9]+]] = internal constant <4 x float> <float 1.000000e+00, float 1.000000e+00, float 3.000000e+00, float 7.000000e+00>, align 4
+; C4xfloat: @[[C2:[_a-z0-9]+]] = internal constant <4 x float> <float 1.000000e+00, float 1.000000e+00, float 3.000000e+00, float 1.900000e+01>, align 4
+; C4xfloat: define void @test4xfloat(<4 x float> %in) {
+; C4xfloat-NEXT: %[[M1:[_a-z0-9]+]] = load <4 x float>* @[[C1]], align 4
+; C4xfloat-NEXT: %[[M2:[_a-z0-9]+]] = load <4 x float>* @[[C2]], align 4
+; C4xfloat-NEXT: %polyhex = fadd <4 x float> %in, %[[M1]]
+; C4xfloat-NEXT: %poset = fadd <4 x float> %[[M2]], %in
+; C4xfloat-NEXT: ret void
+
+; Globalized constant loads have to dominate their use.
+define void @testbranch(i1 %cond, <4 x i32> %in) {
+  br i1 %cond, label %lhs, label %rhs
+lhs:
+  %from_lhs = add <4 x i32> %in, <i32 1, i32 1, i32 2, i32 2>
+  br label %done
+rhs:
+  %from_rhs = add <4 x i32> <i32 2, i32 2, i32 1, i32 1>, %in
+  br label %done
+done:
+  %merged = phi <4 x i32> [ %from_lhs, %lhs ], [ %from_rhs, %rhs ]
+  ret void
+}
+; Cbranch: @[[C1:[_a-z0-9]+]] = internal constant <4 x i32> <i32 1, i32 1, i32 2, i32 2>, align 4
+; Cbranch: @[[C2:[_a-z0-9]+]] = internal constant <4 x i32> <i32 2, i32 2, i32 1, i32 1>, align 4
+; Cbranch: define void @testbranch(i1 %cond, <4 x i32> %in) {
+; Cbranch-NEXT: %[[M1:[_a-z0-9]+]] = load <4 x i32>* @[[C1]], align 4
+; Cbranch-NEXT: %[[M2:[_a-z0-9]+]] = load <4 x i32>* @[[C2]], align 4
+; Cbranch-NEXT: br i1 %cond, label %lhs, label %rhs
+; Cbranch: lhs:
+; Cbranch-NEXT: %from_lhs = add <4 x i32> %in, %[[M1]]
+; Cbranch-NEXT: br label %done
+; Cbranch: rhs:
+; Cbranch-NEXT: %from_rhs = add <4 x i32> %[[M2]], %in
+; Cbranch-NEXT: br label %done
+; Cbranch: done:
+; Cbranch-NEXT: %merged = phi <4 x i32> [ %from_lhs, %lhs ], [ %from_rhs, %rhs ]
+; Cbranch-NEXT: ret void
+
+; Globalizing redundant constants between functions should materialize
+; them in each function.
+define void @testduplicate1() {
+  %foo = add <4 x i32> <i32 1, i32 1, i32 1, i32 1>, <i32 0, i32 0, i32 0, i32 0>
+  ret void
+}
+define void @testduplicate2() {
+  %foo = add <4 x i32> <i32 1, i32 1, i32 1, i32 1>, <i32 0, i32 0, i32 0, i32 0>
+  ret void
+}
+; Cduplicate: @[[C1:[_a-z0-9]+]] = internal constant <4 x i32> <i32 1, i32 1, i32 1, i32 1>, align 4
+; Cduplicate: @[[C2:[_a-z0-9]+]] = internal constant <4 x i32> <i32 1, i32 1, i32 1, i32 1>, align 4
+; Cduplicate: define void @testduplicate1() {
+; Cduplicate-NEXT: %[[M1:[_a-z0-9]+]] = load <4 x i32>* @[[C1]], align 4
+; Cduplicate-NEXT: %foo = add <4 x i32> %[[M1]], zeroinitializer

[jvoung (off chromium), 2014/04/16 22:03:12]

Didn't realize that "zeroinitializer" could be an operand...

Just to check, the zeroinitializer is also a globalized constant?

[JF, 2014/04/16 22:35:36]

That's not guaranteed, it could be a regular immediate too, but in this case
it's simple enough that it shouldn't be an issue.

+; Cduplicate-NEXT: ret void
+; Cduplicate: define void @testduplicate2() {
+; Cduplicate-NEXT: %[[M1:[_a-z0-9]+]] = load <4 x i32>* @[[C2]], align 4
+; Cduplicate-NEXT: %foo = add <4 x i32> %[[M1]], zeroinitializer
+; Cduplicate-NEXT: ret void