[X86] Fix a bad bitcast in the load form of vXi16 uniform shift patterns for EVEX encoded instructions.

llvm-svn: 344563

4 files changed, 16 insertions, 21 deletions

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index b1cb1545ec4..158aba447ed 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -5826,7 +5826,7 @@ multiclass avx512_shift_rmbi<bits<8> opc, Format ImmFormM,
 
 multiclass avx512_shift_rrm<bits<8> opc, string OpcodeStr, SDNode OpNode,
                             X86FoldableSchedWrite sched, ValueType SrcVT,
-                            PatFrag bc_frag, X86VectorVTInfo _> {
+                            X86VectorVTInfo _> {
    // src2 is always 128-bit
   let ExeDomain = _.ExeDomain in {
   defm rr : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
@@ -5837,7 +5837,8 @@ multiclass avx512_shift_rrm<bits<8> opc, string OpcodeStr, SDNode OpNode,
   defm rm : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
                    (ins _.RC:$src1, i128mem:$src2), OpcodeStr,
                        "$src2, $src1", "$src1, $src2",
-                   (_.VT (OpNode _.RC:$src1, (bc_frag (loadv2i64 addr:$src2))))>,
+                   (_.VT (OpNode _.RC:$src1,
+                                 (SrcVT (bitconvert (loadv2i64 addr:$src2)))))>,
                    AVX512BIBase,
                    EVEX_4V, Sched<[sched.Folded, sched.ReadAfterFold]>;
   }
@@ -5845,18 +5846,18 @@ multiclass avx512_shift_rrm<bits<8> opc, string OpcodeStr, SDNode OpNode,
 
 multiclass avx512_shift_sizes<bits<8> opc, string OpcodeStr, SDNode OpNode,
                               X86SchedWriteWidths sched, ValueType SrcVT,
-                              PatFrag bc_frag, AVX512VLVectorVTInfo VTInfo,
+                              AVX512VLVectorVTInfo VTInfo,
                               Predicate prd> {
   let Predicates = [prd] in
   defm Z    : avx512_shift_rrm<opc, OpcodeStr, OpNode, sched.ZMM, SrcVT,
-                               bc_frag, VTInfo.info512>, EVEX_V512,
+                               VTInfo.info512>, EVEX_V512,
                                EVEX_CD8<VTInfo.info512.EltSize, CD8VQ> ;
   let Predicates = [prd, HasVLX] in {
   defm Z256 : avx512_shift_rrm<opc, OpcodeStr, OpNode, sched.YMM, SrcVT,
-                               bc_frag, VTInfo.info256>, EVEX_V256,
+                               VTInfo.info256>, EVEX_V256,
                                EVEX_CD8<VTInfo.info256.EltSize, CD8VH>;
   defm Z128 : avx512_shift_rrm<opc, OpcodeStr, OpNode, sched.XMM, SrcVT,
-                               bc_frag, VTInfo.info128>, EVEX_V128,
+                               VTInfo.info128>, EVEX_V128,
                                EVEX_CD8<VTInfo.info128.EltSize, CD8VF>;
   }
 }
@@ -5866,12 +5867,12 @@ multiclass avx512_shift_types<bits<8> opcd, bits<8> opcq, bits<8> opcw,
                               X86SchedWriteWidths sched,
                               bit NotEVEX2VEXConvertibleQ = 0> {
   defm D : avx512_shift_sizes<opcd, OpcodeStr#"d", OpNode, sched, v4i32,
-                              bc_v4i32, avx512vl_i32_info, HasAVX512>;
+                              avx512vl_i32_info, HasAVX512>;
   let notEVEX2VEXConvertible = NotEVEX2VEXConvertibleQ in
   defm Q : avx512_shift_sizes<opcq, OpcodeStr#"q", OpNode, sched, v2i64,
-                              bc_v2i64, avx512vl_i64_info, HasAVX512>, VEX_W;
+                              avx512vl_i64_info, HasAVX512>, VEX_W;
   defm W : avx512_shift_sizes<opcw, OpcodeStr#"w", OpNode, sched, v8i16,
-                              bc_v2i64, avx512vl_i16_info, HasBWI>;
+                              avx512vl_i16_info, HasBWI>;
 }
 
 multiclass avx512_shift_rmi_sizes<bits<8> opc, Format ImmFormR, Format ImmFormM,
diff --git a/llvm/test/CodeGen/X86/avx2-intrinsics-x86.ll b/llvm/test/CodeGen/X86/avx2-intrinsics-x86.ll
index bba70b139e2..101448e22ac 100644
--- a/llvm/test/CodeGen/X86/avx2-intrinsics-x86.ll
+++ b/llvm/test/CodeGen/X86/avx2-intrinsics-x86.ll
@@ -541,8 +541,7 @@ define <16 x i16> @test_x86_avx2_psrl_w_load(<16 x i16> %a0, <8 x i16>* %p) {
 ; X86-AVX512VL-LABEL: test_x86_avx2_psrl_w_load:
 ; X86-AVX512VL:       ## %bb.0:
 ; X86-AVX512VL-NEXT:    movl {{[0-9]+}}(%esp), %eax ## encoding: [0x8b,0x44,0x24,0x04]
-; X86-AVX512VL-NEXT:    vmovdqa (%eax), %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0x08]
-; X86-AVX512VL-NEXT:    vpsrlw %xmm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xd1,0xc1]
+; X86-AVX512VL-NEXT:    vpsrlw (%eax), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xd1,0x00]
 ; X86-AVX512VL-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X64-AVX-LABEL: test_x86_avx2_psrl_w_load:
@@ -552,8 +551,7 @@ define <16 x i16> @test_x86_avx2_psrl_w_load(<16 x i16> %a0, <8 x i16>* %p) {
 ;
 ; X64-AVX512VL-LABEL: test_x86_avx2_psrl_w_load:
 ; X64-AVX512VL:       ## %bb.0:
-; X64-AVX512VL-NEXT:    vmovdqa (%rdi), %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0x0f]
-; X64-AVX512VL-NEXT:    vpsrlw %xmm1, %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xd1,0xc1]
+; X64-AVX512VL-NEXT:    vpsrlw (%rdi), %ymm0, %ymm0 ## EVEX TO VEX Compression encoding: [0xc5,0xfd,0xd1,0x07]
 ; X64-AVX512VL-NEXT:    retq ## encoding: [0xc3]
   %a1 = load <8 x i16>, <8 x i16>* %p
   %res = call <16 x i16> @llvm.x86.avx2.psrl.w(<16 x i16> %a0, <8 x i16> %a1) ; <<16 x i16>> [#uses=1]
diff --git a/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll b/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll
index 650235d51b3..cf52746c3a5 100644
--- a/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll
@@ -1952,14 +1952,12 @@ define <32 x i16> @test_x86_avx512_psrl_w_512_load(<32 x i16> %a0, <8 x i16>* %p
 ; X86-LABEL: test_x86_avx512_psrl_w_512_load:
 ; X86:       # %bb.0:
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x04]
-; X86-NEXT:    vmovdqa (%eax), %xmm1 # encoding: [0xc5,0xf9,0x6f,0x08]
-; X86-NEXT:    vpsrlw %xmm1, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xd1,0xc1]
+; X86-NEXT:    vpsrlw (%eax), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xd1,0x00]
 ; X86-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-LABEL: test_x86_avx512_psrl_w_512_load:
 ; X64:       # %bb.0:
-; X64-NEXT:    vmovdqa (%rdi), %xmm1 # encoding: [0xc5,0xf9,0x6f,0x0f]
-; X64-NEXT:    vpsrlw %xmm1, %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xd1,0xc1]
+; X64-NEXT:    vpsrlw (%rdi), %zmm0, %zmm0 # encoding: [0x62,0xf1,0x7d,0x48,0xd1,0x07]
 ; X64-NEXT:    retq # encoding: [0xc3]
   %a1 = load <8 x i16>, <8 x i16>* %p
   %res = call <32 x i16> @llvm.x86.avx512.psrl.w.512(<32 x i16> %a0, <8 x i16> %a1) ; <<32 x i16>> [#uses=1]
diff --git a/llvm/test/CodeGen/X86/sse2-intrinsics-x86.ll b/llvm/test/CodeGen/X86/sse2-intrinsics-x86.ll
index 020c4985943..8dedce5fc8b 100644
--- a/llvm/test/CodeGen/X86/sse2-intrinsics-x86.ll
+++ b/llvm/test/CodeGen/X86/sse2-intrinsics-x86.ll
@@ -1434,8 +1434,7 @@ define <8 x i16> @test_x86_sse2_psrl_w_load(<8 x i16> %a0, <8 x i16>* %p) {
 ; X86-AVX512-LABEL: test_x86_sse2_psrl_w_load:
 ; X86-AVX512:       ## %bb.0:
 ; X86-AVX512-NEXT:    movl {{[0-9]+}}(%esp), %eax ## encoding: [0x8b,0x44,0x24,0x04]
-; X86-AVX512-NEXT:    vmovdqa (%eax), %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0x08]
-; X86-AVX512-NEXT:    vpsrlw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xd1,0xc1]
+; X86-AVX512-NEXT:    vpsrlw (%eax), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xd1,0x00]
 ; X86-AVX512-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X64-SSE-LABEL: test_x86_sse2_psrl_w_load:
@@ -1450,8 +1449,7 @@ define <8 x i16> @test_x86_sse2_psrl_w_load(<8 x i16> %a0, <8 x i16>* %p) {
 ;
 ; X64-AVX512-LABEL: test_x86_sse2_psrl_w_load:
 ; X64-AVX512:       ## %bb.0:
-; X64-AVX512-NEXT:    vmovdqa (%rdi), %xmm1 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0x6f,0x0f]
-; X64-AVX512-NEXT:    vpsrlw %xmm1, %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xd1,0xc1]
+; X64-AVX512-NEXT:    vpsrlw (%rdi), %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc5,0xf9,0xd1,0x07]
 ; X64-AVX512-NEXT:    retq ## encoding: [0xc3]
   %a1 = load <8 x i16>, <8 x i16>* %p
   %res = call <8 x i16> @llvm.x86.sse2.psrl.w(<8 x i16> %a0, <8 x i16> %a1) ; <<8 x i16>> [#uses=1]