[AVX-512] Teach fast isel to handle 512-bit vector bitcasts.

llvm-svn: 288641

2 files changed, 252 insertions, 2 deletions

diff --git a/llvm/lib/Target/X86/X86FastISel.cpp b/llvm/lib/Target/X86/X86FastISel.cpp
index 44d918f3786..53e6ab6e359 100644
--- a/llvm/lib/Target/X86/X86FastISel.cpp
+++ b/llvm/lib/Target/X86/X86FastISel.cpp
@@ -3563,8 +3563,14 @@ X86FastISel::fastSelectInstruction(const Instruction *I)  {
     if (!SrcVT.isSimple() || !DstVT.isSimple())
       return false;
 
-    if (!SrcVT.is128BitVector() &&
-        !(Subtarget->hasAVX() && SrcVT.is256BitVector()))
+    MVT SVT = SrcVT.getSimpleVT();
+    MVT DVT = DstVT.getSimpleVT();
+
+    if (!SVT.is128BitVector() &&
+        !(Subtarget->hasAVX() && SVT.is256BitVector()) &&
+        !(Subtarget->hasAVX512() && SVT.is512BitVector() &&
+          (Subtarget->hasBWI() || (SVT.getScalarSizeInBits() >= 32 &&
+                                   DVT.getScalarSizeInBits() >= 32))))
       return false;
 
     unsigned Reg = getRegForValue(I->getOperand(0));
diff --git a/llvm/test/CodeGen/X86/fast-isel-bitcasts-avx512.ll b/llvm/test/CodeGen/X86/fast-isel-bitcasts-avx512.ll
new file mode 100644
index 00000000000..7b81be3fc43
--- /dev/null
+++ b/llvm/test/CodeGen/X86/fast-isel-bitcasts-avx512.ll
@@ -0,0 +1,244 @@
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw -fast-isel -fast-isel-abort=1 -asm-verbose=0 | FileCheck %s
+;
+; Bitcasts between 512-bit vector types are no-ops since no instruction is
+; needed for the conversion.
+
+define <8 x i64> @v16i32_to_v8i64(<16 x i32> %a) {
+;CHECK-LABEL: v16i32_to_v8i64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x i32> %a to <8 x i64>
+  ret <8 x i64> %1
+}
+
+define <8 x i64> @v32i16_to_v8i64(<32 x i16> %a) {
+;CHECK-LABEL: v32i16_to_v8i64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <32 x i16> %a to <8 x i64>
+  ret <8 x i64> %1
+}
+
+define <8 x i64> @v64i8_to_v8i64(<64 x i8> %a) {
+;CHECK-LABEL: v64i8_to_v8i64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <64 x i8> %a to <8 x i64>
+  ret <8 x i64> %1
+}
+
+define <8 x i64> @v8f64_to_v8i64(<8 x double> %a) {
+;CHECK-LABEL: v8f64_to_v8i64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x double> %a to <8 x i64>
+  ret <8 x i64> %1
+}
+
+define <8 x i64> @v16f32_to_v8i64(<16 x float> %a) {
+;CHECK-LABEL: v16f32_to_v8i64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x float> %a to <8 x i64>
+  ret <8 x i64> %1
+}
+
+define <16 x i32> @v8i64_to_v16i32(<8 x i64> %a) {
+;CHECK-LABEL: v8i64_to_v16i32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x i64> %a to <16 x i32>
+  ret <16 x i32> %1
+}
+
+define <16 x i32> @v32i16_to_v16i32(<32 x i16> %a) {
+;CHECK-LABEL: v32i16_to_v16i32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <32 x i16> %a to <16 x i32>
+  ret <16 x i32> %1
+}
+
+define <16 x i32> @v64i8_to_v16i32(<64 x i8> %a) {
+;CHECK-LABEL: v64i8_to_v16i32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <64 x i8> %a to <16 x i32>
+  ret <16 x i32> %1
+}
+
+define <16 x i32> @v8f64_to_v16i32(<8 x double> %a) {
+;CHECK-LABEL: v8f64_to_v16i32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x double> %a to <16 x i32>
+  ret <16 x i32> %1
+}
+
+define <16 x i32> @v16f32_to_v16i32(<16 x float> %a) {
+;CHECK-LABEL: v16f32_to_v16i32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x float> %a to <16 x i32>
+  ret <16 x i32> %1
+}
+
+define <32 x i16> @v8i64_to_v32i16(<8 x i64> %a) {
+;CHECK-LABEL: v8i64_to_v32i16:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x i64> %a to <32 x i16>
+  ret <32 x i16> %1
+}
+
+define <32 x i16> @v16i32_to_v32i16(<16 x i32> %a) {
+;CHECK-LABEL: v16i32_to_v32i16:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x i32> %a to <32 x i16>
+  ret <32 x i16> %1
+}
+
+define <32 x i16> @v64i8_to_v32i16(<64 x i8> %a) {
+;CHECK-LABEL: v64i8_to_v32i16:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <64 x i8> %a to <32 x i16>
+  ret <32 x i16> %1
+}
+
+define <32 x i16> @v8f64_to_v32i16(<8 x double> %a) {
+;CHECK-LABEL: v8f64_to_v32i16:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x double> %a to <32 x i16>
+  ret <32 x i16> %1
+}
+
+define <32 x i16> @v16f32_to_v32i16(<16 x float> %a) {
+;CHECK-LABEL: v16f32_to_v32i16:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x float> %a to <32 x i16>
+  ret <32 x i16> %1
+}
+
+define <64 x i8> @v32i16_to_v64i8(<32 x i16> %a) {
+;CHECK-LABEL: v32i16_to_v64i8:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <32 x i16> %a to <64 x i8>
+  ret <64 x i8> %1
+}
+
+define <64 x i8> @v8i64_to_v64i8(<8 x i64> %a) {
+;CHECK-LABEL: v8i64_to_v64i8:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x i64> %a to <64 x i8>
+  ret <64 x i8> %1
+}
+
+define <64 x i8> @v16i32_to_v64i8(<16 x i32> %a) {
+;CHECK-LABEL: v16i32_to_v64i8:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x i32> %a to <64 x i8>
+  ret <64 x i8> %1
+}
+
+define <64 x i8> @v8f64_to_v64i8(<8 x double> %a) {
+;CHECK-LABEL: v8f64_to_v64i8:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x double> %a to <64 x i8>
+  ret <64 x i8> %1
+}
+
+define <64 x i8> @v16f32_to_v64i8(<16 x float> %a) {
+;CHECK-LABEL: v16f32_to_v64i8:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x float> %a to <64 x i8>
+  ret <64 x i8> %1
+}
+
+define <16 x float> @v64i8_to_v16f32(<64 x i8> %a) {
+;CHECK-LABEL: v64i8_to_v16f32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <64 x i8> %a to <16 x float>
+  ret <16 x float> %1
+}
+
+define <16 x float> @v32i16_to_v16f32(<32 x i16> %a) {
+;CHECK-LABEL: v32i16_to_v16f32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <32 x i16> %a to <16 x float>
+  ret <16 x float> %1
+}
+
+define <16 x float> @v8i64_to_v16f32(<8 x i64> %a) {
+;CHECK-LABEL: v8i64_to_v16f32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x i64> %a to <16 x float>
+  ret <16 x float> %1
+}
+
+define <16 x float> @v16i32_to_v16f32(<16 x i32> %a) {
+;CHECK-LABEL: v16i32_to_v16f32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x i32> %a to <16 x float>
+  ret <16 x float> %1
+}
+
+define <16 x float> @v8f64_to_v16f32(<8 x double> %a) {
+;CHECK-LABEL: v8f64_to_v16f32:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x double> %a to <16 x float>
+  ret <16 x float> %1
+}
+
+define <8 x double> @v16f32_to_v8f64(<16 x float> %a) {
+;CHECK-LABEL: v16f32_to_v8f64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x float> %a to <8 x double>
+  ret <8 x double> %1
+}
+
+define <8 x double> @v64i8_to_v8f64(<64 x i8> %a) {
+;CHECK-LABEL: v64i8_to_v8f64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <64 x i8> %a to <8 x double>
+  ret <8 x double> %1
+}
+
+define <8 x double> @v32i16_to_v8f64(<32 x i16> %a) {
+;CHECK-LABEL: v32i16_to_v8f64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <32 x i16> %a to <8 x double>
+  ret <8 x double> %1
+}
+
+define <8 x double> @v8i64_to_v8f64(<8 x i64> %a) {
+;CHECK-LABEL: v8i64_to_v8f64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <8 x i64> %a to <8 x double>
+  ret <8 x double> %1
+}
+
+define <8 x double> @v16i32_to_v8f64(<16 x i32> %a) {
+;CHECK-LABEL: v16i32_to_v8f64:
+;CHECK-NEXT: .cfi_startproc
+;CHECK-NEXT: ret
+  %1 = bitcast <16 x i32> %a to <8 x double>
+  ret <8 x double> %1
+}