[X86] WriteFShuffle256 shuffles aren't microcoded in the llvm sense

znver1/2 might have poor throughput for crosslane shuffles but they don't consume 100 cycles of resources

I think there was a misunderstanding between the AMD definition of microcoding (more than 2-3 uops) and LLVM (here be dragons - impossible to approximately model the instruction)

This is more yak shaving to come from D103695 - this time working out why codegen involving broadcasts gives such weird numbers

diff --git a/llvm/lib/Target/X86/X86ScheduleZnver1.td b/llvm/lib/Target/X86/X86ScheduleZnver1.td
index 79beb3e..01deab3 100644 (file)
--- a/llvm/lib/Target/X86/X86ScheduleZnver1.td
+++ b/llvm/lib/Target/X86/X86ScheduleZnver1.td
@@ -372,6 +372,8 @@ defm : ZnWriteResFpuPair<WriteFSqrt64X,  [ZnFPU3], 20, [8]>;
 defm : ZnWriteResFpuPair<WriteFSqrt64Y,  [ZnFPU3], 20, [16], 1>;
 defm : X86WriteResPairUnsupported<WriteFSqrt64Z>;
 defm : ZnWriteResFpuPair<WriteFSqrt80,   [ZnFPU3], 20, [20]>;
+defm : ZnWriteResFpuPair<WriteFShuffle256, [ZnFPU], 2>;
+defm : ZnWriteResFpuPair<WriteFVarShuffle256, [ZnFPU], 2>;
 
 // Vector integer operations which uses FPU units
 defm : X86WriteRes<WriteVecLoad,         [ZnAGU], 8, [1], 1>;
@@ -479,11 +481,6 @@ defm : ZnWriteResFpuPair<WriteAESKeyGen, [ZnFPU01], 4>;
 def : WriteRes<WriteFence,  [ZnAGU]>;
 def : WriteRes<WriteNop, []>;
 
-// Following instructions with latency=100 are microcoded.
-// We set long latency so as to block the entire pipeline.
-defm : ZnWriteResFpuPair<WriteFShuffle256, [ZnFPU], 100>;
-defm : ZnWriteResFpuPair<WriteFVarShuffle256, [ZnFPU], 100>;
-
 // Microcoded Instructions
 def ZnWriteMicrocoded : SchedWriteRes<[]> {
   let Latency = 100;

diff --git a/llvm/lib/Target/X86/X86ScheduleZnver2.td b/llvm/lib/Target/X86/X86ScheduleZnver2.td
index b4f72a9..788b71e 100644 (file)
--- a/llvm/lib/Target/X86/X86ScheduleZnver2.td
+++ b/llvm/lib/Target/X86/X86ScheduleZnver2.td
@@ -371,6 +371,8 @@ defm : Zn2WriteResFpuPair<WriteFSqrt64X,  [Zn2FPU3], 20, [10]>;
 defm : Zn2WriteResFpuPair<WriteFSqrt64Y,  [Zn2FPU3], 20, [10]>;
 defm : X86WriteResPairUnsupported<WriteFSqrt64Z>;
 defm : Zn2WriteResFpuPair<WriteFSqrt80,   [Zn2FPU3], 20, [20]>;
+defm : Zn2WriteResFpuPair<WriteFShuffle256, [Zn2FPU], 2>;
+defm : Zn2WriteResFpuPair<WriteFVarShuffle256, [Zn2FPU], 2>;
 
 // Vector integer operations which uses FPU units
 defm : X86WriteRes<WriteVecLoad,         [Zn2AGU], 8, [1], 1>;
@@ -478,11 +480,6 @@ defm : Zn2WriteResFpuPair<WriteAESKeyGen, [Zn2FPU01], 4>;
 def : WriteRes<WriteFence,  [Zn2AGU]>;
 def : WriteRes<WriteNop, []>;
 
-// Following instructions with latency=100 are microcoded.
-// We set long latency so as to block the entire pipeline.
-defm : Zn2WriteResFpuPair<WriteFShuffle256, [Zn2FPU], 100>;
-defm : Zn2WriteResFpuPair<WriteFVarShuffle256, [Zn2FPU], 100>;
-
 // Microcoded Instructions
 def Zn2WriteMicrocoded : SchedWriteRes<[]> {
   let Latency = 100;

diff --git a/llvm/test/tools/llvm-mca/X86/Znver1/resources-avx2.s b/llvm/test/tools/llvm-mca/X86/Znver1/resources-avx2.s
index 0648c31..179e6a4 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/Znver1/resources-avx2.s
+++ b/llvm/test/tools/llvm-mca/X86/Znver1/resources-avx2.s
@@ -461,8 +461,8 @@ vpxor           (%rax), %ymm1, %ymm2
 
 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 # CHECK-NEXT:  1      8     0.50    *                   vbroadcasti128 (%rax), %ymm0
-# CHECK-NEXT:  1      100   0.25                        vbroadcastsd   %xmm0, %ymm0
-# CHECK-NEXT:  1      100   0.25                        vbroadcastss   %xmm0, %ymm0
+# CHECK-NEXT:  1      2     0.25                        vbroadcastsd   %xmm0, %ymm0
+# CHECK-NEXT:  1      2     0.25                        vbroadcastss   %xmm0, %ymm0
 # CHECK-NEXT:  1      2     0.25                        vextracti128   $1, %ymm0, %xmm2
 # CHECK-NEXT:  1      1     0.50           *            vextracti128   $1, %ymm0, (%rax)
 # CHECK-NEXT:  1      100   0.25    *                   vgatherdpd     %xmm0, (%rax,%xmm1,2), %xmm2
@@ -562,10 +562,10 @@ vpxor           (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      9     0.50    *                   vperm2i128     $1, (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      2     0.25                        vpermd %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      9     0.50    *                   vpermd (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vpermpd        $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      107   0.50    *                   vpermpd        $1, (%rax), %ymm2
-# CHECK-NEXT:  1      100   0.25                        vpermps        %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      107   0.50    *                   vpermps        (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      2     0.25                        vpermpd        $1, %ymm0, %ymm2
+# CHECK-NEXT:  1      9     0.50    *                   vpermpd        $1, (%rax), %ymm2
+# CHECK-NEXT:  1      2     0.25                        vpermps        %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      9     0.50    *                   vpermps        (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      2     0.25                        vpermq $1, %ymm0, %ymm2
 # CHECK-NEXT:  1      9     0.50    *                   vpermq $1, (%rax), %ymm2
 # CHECK-NEXT:  1      100   0.25    *                   vpgatherdd     %xmm0, (%rax,%xmm1,2), %xmm2

diff --git a/llvm/test/tools/llvm-mca/X86/Znver2/resources-avx2.s b/llvm/test/tools/llvm-mca/X86/Znver2/resources-avx2.s
index 93eb00e..e77f2f7 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/Znver2/resources-avx2.s
+++ b/llvm/test/tools/llvm-mca/X86/Znver2/resources-avx2.s
@@ -461,8 +461,8 @@ vpxor           (%rax), %ymm1, %ymm2
 
 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 # CHECK-NEXT:  1      8     0.33    *                   vbroadcasti128 (%rax), %ymm0
-# CHECK-NEXT:  1      100   0.25                        vbroadcastsd   %xmm0, %ymm0
-# CHECK-NEXT:  1      100   0.25                        vbroadcastss   %xmm0, %ymm0
+# CHECK-NEXT:  1      2     0.25                        vbroadcastsd   %xmm0, %ymm0
+# CHECK-NEXT:  1      2     0.25                        vbroadcastss   %xmm0, %ymm0
 # CHECK-NEXT:  1      2     0.25                        vextracti128   $1, %ymm0, %xmm2
 # CHECK-NEXT:  1      1     0.33           *            vextracti128   $1, %ymm0, (%rax)
 # CHECK-NEXT:  1      100   0.25    *                   vgatherdpd     %xmm0, (%rax,%xmm1,2), %xmm2
@@ -562,10 +562,10 @@ vpxor           (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      9     0.33    *                   vperm2i128     $1, (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      2     0.25                        vpermd %ymm0, %ymm1, %ymm2
 # CHECK-NEXT:  1      9     0.33    *                   vpermd (%rax), %ymm1, %ymm2
-# CHECK-NEXT:  1      100   0.25                        vpermpd        $1, %ymm0, %ymm2
-# CHECK-NEXT:  1      107   0.33    *                   vpermpd        $1, (%rax), %ymm2
-# CHECK-NEXT:  1      100   0.25                        vpermps        %ymm0, %ymm1, %ymm2
-# CHECK-NEXT:  1      107   0.33    *                   vpermps        (%rax), %ymm1, %ymm2
+# CHECK-NEXT:  1      2     0.25                        vpermpd        $1, %ymm0, %ymm2
+# CHECK-NEXT:  1      9     0.33    *                   vpermpd        $1, (%rax), %ymm2
+# CHECK-NEXT:  1      2     0.25                        vpermps        %ymm0, %ymm1, %ymm2
+# CHECK-NEXT:  1      9     0.33    *                   vpermps        (%rax), %ymm1, %ymm2
 # CHECK-NEXT:  1      2     0.25                        vpermq $1, %ymm0, %ymm2
 # CHECK-NEXT:  1      9     0.33    *                   vpermq $1, (%rax), %ymm2
 # CHECK-NEXT:  1      100   0.25    *                   vpgatherdd     %xmm0, (%rax,%xmm1,2), %xmm2