[X86] Model MXCSR for AVX instructions other than AVX512

[pengfei]

Summary

Model MXCSR for AVX instructions other than AVX512

[LuoYuanke]

llvm/lib/Target/X86/X86InstrSSE.td

7329–7330

Is there any exception when convert from half to single float?

[craig.topper]

llvm/lib/Target/X86/X86InstrSSE.td

7329–7330

SNAN input signals an invalid operation exception

[craig.topper]

LGTM

llvm/lib/Target/X86/X86InstrFMA.td

def m     : FMA3<opc, MRMSrcMem, (outs RC:$dst),
                    (ins RC:$src1, RC:$src2, x86memop:$src3),
                    !strconcat(OpcodeStr,
                               "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
                    [(set RC:$dst, (VT (Op (MemFrag addr:$src3), RC:$src1,
                                           RC:$src2)))]>,
                    Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
 }
 
-let Constraints = "$src1 = $dst", hasSideEffects = 0, isCommutable = 1 in
+let Constraints = "$src1 = $dst", hasSideEffects = 0, isCommutable = 1,
+    Uses = [MXCSR], mayRaiseFPException = 1 in
 multiclass fma3p_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
                        string OpcodeStr, string PackTy, string Suff,
                        PatFrag MemFrag128, PatFrag MemFrag256,
                        SDNode Op, ValueType OpTy128, ValueType OpTy256,
                        X86SchedWriteWidths sched> {
   defm NAME#213#Suff : fma3p_rm_213<opc213, !strconcat(OpcodeStr, "213", PackTy),
                                     VR128, OpTy128, f128mem, MemFrag128, Op, sched.XMM>;
   defm NAME#231#Suff : fma3p_rm_231<opc231, !strconcat(OpcodeStr, "231", PackTy),
def m : FMA3S<opc, MRMSrcMem, (outs RC:$dst),
                 !strconcat(OpcodeStr,
                            "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
                 [(set RC:$dst,
                   (OpNode (load addr:$src3), RC:$src1, RC:$src2))]>,
                 Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
 }
 
 let Constraints = "$src1 = $dst", isCommutable = 1, isCodeGenOnly = 1,
-    hasSideEffects = 0 in
+    hasSideEffects = 0, Uses = [MXCSR], mayRaiseFPException = 1 in
 multiclass fma3s_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
                        string OpStr, string PackTy, string Suff,
                        SDNode OpNode, RegisterClass RC,
                        X86MemOperand x86memop, X86FoldableSchedWrite sched> {
   defm NAME#213#Suff : fma3s_rm_213<opc213, !strconcat(OpStr, "213", PackTy),
                                     x86memop, RC, OpNode, sched>;
   defm NAME#231#Suff : fma3s_rm_231<opc231, !strconcat(OpStr, "231", PackTy),
                                     x86memop, RC, OpNode, sched>;

 // All of the FMA*_Int opcodes are defined as commutable here.
 // Commuting the 2nd and 3rd source register operands of FMAs is quite trivial
 // and the corresponding optimizations have been developed.
 // Commuting the 1st operand of FMA*_Int requires some additional analysis,
 // the commute optimization is legal only if all users of FMA*_Int use only
 // the lowest element of the FMA*_Int instruction. Even though such analysis
 // may be not implemented yet we allow the routines doing the actual commute
 // transformation to decide if one or another instruction is commutable or not.
-let Constraints = "$src1 = $dst", isCommutable = 1, hasSideEffects = 0 in
+let Constraints = "$src1 = $dst", isCommutable = 1, hasSideEffects = 0,
+    Uses = [MXCSR], mayRaiseFPException = 1 in
 multiclass fma3s_rm_int<bits<8> opc, string OpcodeStr,
                         Operand memopr, RegisterClass RC,
                         X86FoldableSchedWrite sched> {
   def r_Int : FMA3S_Int<opc, MRMSrcReg, (outs RC:$dst),
                         (ins RC:$src1, RC:$src2, RC:$src3),
                         !strconcat(OpcodeStr,
                                    "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
                         []>, Sched<[sched]>;

 defm : scalar_fma_patterns<X86Fmsub, "VFMSUB", "SD", X86Movsd, v2f64, f64, FR64, loadf64>;
 defm : scalar_fma_patterns<X86Fnmadd, "VFNMADD", "SD", X86Movsd, v2f64, f64, FR64, loadf64>;
 defm : scalar_fma_patterns<X86Fnmsub, "VFNMSUB", "SD", X86Movsd, v2f64, f64, FR64, loadf64>;
 
 //===----------------------------------------------------------------------===//
 // FMA4 - AMD 4 operand Fused Multiply-Add instructions
 //===----------------------------------------------------------------------===//
 
+let Uses = [MXCSR], mayRaiseFPException = 1 in
 multiclass fma4s<bits<8> opc, string OpcodeStr, RegisterClass RC,
                  X86MemOperand x86memop, ValueType OpVT, SDNode OpNode,
                  PatFrag mem_frag, X86FoldableSchedWrite sched> {
   let isCommutable = 1 in
   def rr : FMA4S<opc, MRMSrcRegOp4, (outs RC:$dst),
            (ins RC:$src1, RC:$src2, RC:$src3),
            !strconcat(OpcodeStr,
            "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
def rr_REV : FMA4S<opc, MRMSrcReg, (outs RC:$dst),
                (ins RC:$src1, RC:$src2, RC:$src3),
                !strconcat(OpcodeStr,
                "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>,
                VEX_LIG, FoldGenData<NAME#rr>, Sched<[sched]>;
 }
 
 multiclass fma4s_int<bits<8> opc, string OpcodeStr, Operand memop,
                      ValueType VT, X86FoldableSchedWrite sched> {
-let isCodeGenOnly = 1, hasSideEffects = 0 in {
+let isCodeGenOnly = 1, hasSideEffects = 0,
+    Uses = [MXCSR], mayRaiseFPException = 1 in {
   def rr_Int : FMA4S_Int<opc, MRMSrcRegOp4, (outs VR128:$dst),
                (ins VR128:$src1, VR128:$src2, VR128:$src3),
                !strconcat(OpcodeStr,
                "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
                []>, VEX_W, VEX_LIG, Sched<[sched]>;
   let mayLoad = 1 in
   def rm_Int : FMA4S_Int<opc, MRMSrcMemOp4, (outs VR128:$dst),
                (ins VR128:$src1, VR128:$src2, memop:$src3),
let isCodeGenOnly = 1, hasSideEffects = 0,
   def rr_Int_REV : FMA4S_Int<opc, MRMSrcReg, (outs VR128:$dst),
                (ins VR128:$src1, VR128:$src2, VR128:$src3),
                !strconcat(OpcodeStr,
                "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
                []>, VEX_LIG, FoldGenData<NAME#rr_Int>, Sched<[sched]>;
 } // isCodeGenOnly = 1
 }
 
+let Uses = [MXCSR], mayRaiseFPException = 1 in
 multiclass fma4p<bits<8> opc, string OpcodeStr, SDNode OpNode,
                  ValueType OpVT128, ValueType OpVT256,
                  PatFrag ld_frag128, PatFrag ld_frag256,
                  X86SchedWriteWidths sched> {
   let isCommutable = 1 in
   def rr : FMA4<opc, MRMSrcRegOp4, (outs VR128:$dst),
            (ins VR128:$src1, VR128:$src2, VR128:$src3),
            !strconcat(OpcodeStr,

llvm/lib/Target/X86/X86InstrSSE.td

def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
               (OpNode VR128:$src1, sse_load_f64:$src2, timm:$src3))]>,
         Sched<[sched.Folded, sched.ReadAfterFold]>;
 } // ExeDomain = SSEPackedDouble, isCodeGenOnly = 1
 }
 }
 
 // FP round - roundss, roundps, roundsd, roundpd
 let Predicates = [HasAVX, NoVLX] in {
-  let ExeDomain = SSEPackedSingle in {
+  let ExeDomain = SSEPackedSingle, Uses = [MXCSR], mayRaiseFPException = 1 in {
     // Intrinsic form
     defm VROUNDPS  : sse41_fp_unop_p<0x08, "vroundps", f128mem, VR128, v4f32,
                                      loadv4f32, X86VRndScale, SchedWriteFRnd.XMM>,
                                    VEX, VEX_WIG;
     defm VROUNDPSY : sse41_fp_unop_p<0x08, "vroundps", f256mem, VR256, v8f32,
                                      loadv8f32, X86VRndScale, SchedWriteFRnd.YMM>,
                                    VEX, VEX_L, VEX_WIG;
   }
 
-  let ExeDomain = SSEPackedDouble in {
+  let ExeDomain = SSEPackedDouble, Uses = [MXCSR], mayRaiseFPException = 1 in {
     defm VROUNDPD  : sse41_fp_unop_p<0x09, "vroundpd", f128mem, VR128, v2f64,
                                      loadv2f64, X86VRndScale, SchedWriteFRnd.XMM>,
                                    VEX, VEX_WIG;
     defm VROUNDPDY : sse41_fp_unop_p<0x09, "vroundpd", f256mem, VR256, v4f64,
                                      loadv4f64, X86VRndScale, SchedWriteFRnd.YMM>,
                                    VEX, VEX_L, VEX_WIG;
   }
 }
 let Predicates = [UseAVX] in {
   defm VROUND  : sse41_fp_binop_s<0x0A, 0x0B, "vround", SchedWriteFRnd.Scl,
                                   v4f32, v2f64, X86RndScales, 0>,
-                                  VEX_4V, VEX_LIG, VEX_WIG;
+                                  VEX_4V, VEX_LIG, VEX_WIG, SIMD_EXC;
   defm VROUND  : avx_fp_unop_rm<0x0A, 0x0B, "vround", SchedWriteFRnd.Scl>,
-                                VEX_4V, VEX_LIG, VEX_WIG;
+                                VEX_4V, VEX_LIG, VEX_WIG, SIMD_EXC;
 }
 
 let Predicates = [UseAVX] in {
   def : Pat<(X86VRndScale FR32:$src1, timm:$src2),
             (VROUNDSSr (f32 (IMPLICIT_DEF)), FR32:$src1, timm:$src2)>;
   def : Pat<(X86VRndScale FR64:$src1, timm:$src2),
             (VROUNDSDr (f64 (IMPLICIT_DEF)), FR64:$src1, timm:$src2)>;
 }
multiclass f16c_ps2ph<RegisterClass RC, X86MemOperand x86memop,
   let hasSideEffects = 0, mayStore = 1 in
   def mr : Ii8<0x1D, MRMDestMem, (outs),
                (ins x86memop:$dst, RC:$src1, i32u8imm:$src2),
                "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
                TAPD, VEX, Sched<[MR]>;
 }
 
 let Predicates = [HasF16C, NoVLX] in {
-  defm VCVTPH2PS  : f16c_ph2ps<VR128, f64mem, WriteCvtPH2PS>;
-  defm VCVTPH2PSY : f16c_ph2ps<VR256, f128mem, WriteCvtPH2PSY>, VEX_L;
+  defm VCVTPH2PS  : f16c_ph2ps<VR128, f64mem, WriteCvtPH2PS>, SIMD_EXC;
+  defm VCVTPH2PSY : f16c_ph2ps<VR256, f128mem, WriteCvtPH2PSY>, VEX_L, SIMD_EXC;
   defm VCVTPS2PH  : f16c_ps2ph<VR128, f64mem, WriteCvtPS2PH,
-                               WriteCvtPS2PHSt>;
+                               WriteCvtPS2PHSt>, SIMD_EXC;
   defm VCVTPS2PHY : f16c_ps2ph<VR256, f128mem, WriteCvtPS2PHY,
-                               WriteCvtPS2PHYSt>, VEX_L;
+                               WriteCvtPS2PHYSt>, VEX_L, SIMD_EXC;
 
   // Pattern match vcvtph2ps of a scalar i64 load.
   def : Pat<(v4f32 (X86cvtph2ps (bc_v8i16 (v2i64 (X86vzload64 addr:$src))))),
             (VCVTPH2PSrm addr:$src)>;
   def : Pat<(v4f32 (X86cvtph2ps (bc_v8i16
               (v2i64 (scalar_to_vector (loadi64 addr:$src)))))),
             (VCVTPH2PSrm addr:$src)>;
 

llvm/test/CodeGen/X86/mxcsr-reg-usage.ll

-; RUN: llc -march=x86-64 -mattr=+mmx -stop-after finalize-isel -o - %s | FileCheck %s
+; RUN: llc -march=x86-64 -mattr=+mmx,+fma,+f16c -stop-after finalize-isel -o - %s | FileCheck %s
 ; This test ensures that the MXCSR is implicitly used by MMX FP instructions.
 
 define x86_mmx @mxcsr_mmx(<4 x float> %a0) {
 ; CHECK: MMX_CVTPS2PIirr %{{[0-9]}}, implicit $mxcsr
 ; CHECK: MMX_CVTPI2PSirr %{{[0-9]}}, killed %{{[0-9]}}, implicit $mxcsr
 ; CHECK: MMX_CVTTPS2PIirr killed %{{[0-9]}}, implicit $mxcsr
 ; CHECK: MMX_CVTPI2PDirr killed %{{[0-9]$}}
 ; CHECK: MMX_CVTPD2PIirr killed %{{[0-9]}}, implicit $mxcsr
   %1 = call x86_mmx @llvm.x86.sse.cvtps2pi(<4 x float> %a0)
   %2 = call <4 x float> @llvm.x86.sse.cvtpi2ps(<4 x float> %a0, x86_mmx %1)
   %3 = call x86_mmx @llvm.x86.sse.cvttps2pi(<4 x float> %2)
   %4 = call <2 x double> @llvm.x86.sse.cvtpi2pd(x86_mmx %3)
   %5 = call x86_mmx @llvm.x86.sse.cvtpd2pi(<2 x double> %4)
   ret x86_mmx %5
 }
 
+define half @mxcsr_f16c(float %a) {
+; CHECK: VCVTPS2PH{{.*}}mxcsr
+; CHECK: VCVTPH2PS{{.*}}mxcsr
+  %res = fptrunc float %a to half
+  ret half %res
+}
+
+define <4 x float> @mxcsr_fma_ss(<4 x float> %a, <4 x float> %b) {
+; CHECK: VFMADD{{.*}}mxcsr
+  %res = call <4 x float> @llvm.x86.fma.vfmadd.ss(<4 x float> %b, <4 x float> %a, <4 x float>
+%a)
+  ret <4 x float> %res
+}
+
+define <4 x float> @mxcsr_fma_ps(<4 x float> %a, <4 x float> %b) {
+; CHECK: VFMADD{{.*}}mxcsr
+  %res = call <4 x float> @llvm.x86.fma.vfmadd.ps(<4 x float> %b, <4 x float> %a, <4 x float>
+%a)
+  ret <4 x float> %res
+}
+
 declare x86_mmx @llvm.x86.sse.cvtps2pi(<4 x float>)
 declare<4 x float> @llvm.x86.sse.cvtpi2ps(<4 x float>, x86_mmx)
 declare x86_mmx @llvm.x86.sse.cvttps2pi(<4 x float>)
 declare <2 x double> @llvm.x86.sse.cvtpi2pd(x86_mmx)
 declare x86_mmx @llvm.x86.sse.cvtpd2pi(<2 x double>)
+declare <4 x float> @llvm.x86.fma.vfmadd.ss(<4 x float>, <4 x float>, <4 x float>)
+declare <4 x float> @llvm.x86.fma.vfmadd.ps(<4 x float>, <4 x float>, <4 x float>)

llvm/test/tools/llvm-exegesis/X86/uops-VFMADDSS4rm.s

 # RUN: llvm-exegesis -mode=uops -opcode-name=VFMADDSS4rm -repetition-mode=duplicate | FileCheck %s
 # RUN: llvm-exegesis -mode=uops -opcode-name=VFMADDSS4rm -repetition-mode=loop | FileCheck %s
 
 CHECK:      mode:            uops
 CHECK-NEXT: key:
 CHECK-NEXT:   instructions:
 CHECK-NEXT:     VFMADDSS4rm
 CHECK:      register_initial_values:
-# FIXME: This will be changed to CHECK by the following patch that modeling MXCSR to VFMADDSS.
-CHECK-NOT:      MXCSR
+CHECK:      MXCSR