[X86][SLM] WriteVecIMul instructions only take 1uop (REAPPLIED)

The xmm variant have half the throughput (and +1cy latency) of the mmx variants, but are still 1uop.

I still need to do more thorough testing of SLM on test-suite before fixing the obvious bad numbers for WritePMULLD.

But this helps the D103695 helper script get to more accurate numbers for vXi32 multiplies of extended operands (i.e. we can use PMADDWD, PMULLW/PMULHW etc). Matches what Intel AoM / Agner / llvm-exegesis reports.

diff --git a/llvm/lib/Target/X86/X86ScheduleSLM.td b/llvm/lib/Target/X86/X86ScheduleSLM.td
index a545f3c..2bcb33e 100644 (file)
--- a/llvm/lib/Target/X86/X86ScheduleSLM.td
+++ b/llvm/lib/Target/X86/X86ScheduleSLM.td
@@ -372,8 +372,8 @@ defm : SLMWriteResPair<WriteVecALUX,  [SLM_FPC_RSV01],  1>;
 defm : SLMWriteResPair<WriteVecALUY,  [SLM_FPC_RSV01],  1>;
 defm : X86WriteResPairUnsupported<WriteVecALUZ>;
 defm : SLMWriteResPair<WriteVecIMul,  [SLM_FPC_RSV0],   4>;
-defm : SLMWriteResPair<WriteVecIMulX, [SLM_FPC_RSV0],   5, [2], 2>;
-defm : SLMWriteResPair<WriteVecIMulY, [SLM_FPC_RSV0],   5, [2], 2>;
+defm : SLMWriteResPair<WriteVecIMulX, [SLM_FPC_RSV0],   5, [2]>;
+defm : SLMWriteResPair<WriteVecIMulY, [SLM_FPC_RSV0],   5, [2]>;
 defm : X86WriteResPairUnsupported<WriteVecIMulZ>;
 // FIXME: The below is closer to correct, but caused some perf regressions.
 //defm : SLMWriteResPair<WritePMULLD,  [SLM_FPC_RSV0],   11, [11], 7>;

diff --git a/llvm/test/CodeGen/X86/slow-pmulld.ll b/llvm/test/CodeGen/X86/slow-pmulld.ll
index aeccae3..9a10a23 100644 (file)
--- a/llvm/test/CodeGen/X86/slow-pmulld.ll
+++ b/llvm/test/CodeGen/X86/slow-pmulld.ll
@@ -537,40 +537,40 @@ define <8 x i32> @test_mul_v8i32_v8i16(<8 x i16> %A) {
 define <16 x i32> @test_mul_v16i32_v16i16(<16 x i16> %A) {
 ; SLM32-LABEL: test_mul_v16i32_v16i16:
 ; SLM32:       # %bb.0:
-; SLM32-NEXT:    movdqa %xmm0, %xmm4
-; SLM32-NEXT:    movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
 ; SLM32-NEXT:    movdqa %xmm1, %xmm3
-; SLM32-NEXT:    movdqa %xmm4, %xmm2
-; SLM32-NEXT:    pmullw %xmm0, %xmm4
+; SLM32-NEXT:    movdqa %xmm0, %xmm1
+; SLM32-NEXT:    movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
+; SLM32-NEXT:    movdqa %xmm1, %xmm2
+; SLM32-NEXT:    movdqa %xmm3, %xmm4
+; SLM32-NEXT:    pmullw %xmm0, %xmm1
 ; SLM32-NEXT:    pmulhuw %xmm0, %xmm2
 ; SLM32-NEXT:    pmullw %xmm0, %xmm3
-; SLM32-NEXT:    pmulhuw %xmm0, %xmm1
-; SLM32-NEXT:    movdqa %xmm4, %xmm0
-; SLM32-NEXT:    punpckhwd {{.*#+}} xmm4 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
+; SLM32-NEXT:    pmulhuw %xmm0, %xmm4
+; SLM32-NEXT:    movdqa %xmm1, %xmm0
+; SLM32-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
 ; SLM32-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
 ; SLM32-NEXT:    movdqa %xmm3, %xmm2
-; SLM32-NEXT:    punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; SLM32-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; SLM32-NEXT:    movdqa %xmm4, %xmm1
+; SLM32-NEXT:    punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
+; SLM32-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
 ; SLM32-NEXT:    retl
 ;
 ; SLM64-LABEL: test_mul_v16i32_v16i16:
 ; SLM64:       # %bb.0:
-; SLM64-NEXT:    movdqa %xmm0, %xmm4
-; SLM64-NEXT:    movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
 ; SLM64-NEXT:    movdqa %xmm1, %xmm3
-; SLM64-NEXT:    movdqa %xmm4, %xmm2
-; SLM64-NEXT:    pmullw %xmm0, %xmm4
+; SLM64-NEXT:    movdqa %xmm0, %xmm1
+; SLM64-NEXT:    movdqa {{.*#+}} xmm0 = [18778,18778,18778,18778,18778,18778,18778,18778]
+; SLM64-NEXT:    movdqa %xmm1, %xmm2
+; SLM64-NEXT:    movdqa %xmm3, %xmm4
+; SLM64-NEXT:    pmullw %xmm0, %xmm1
 ; SLM64-NEXT:    pmulhuw %xmm0, %xmm2
 ; SLM64-NEXT:    pmullw %xmm0, %xmm3
-; SLM64-NEXT:    pmulhuw %xmm0, %xmm1
-; SLM64-NEXT:    movdqa %xmm4, %xmm0
-; SLM64-NEXT:    punpckhwd {{.*#+}} xmm4 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
+; SLM64-NEXT:    pmulhuw %xmm0, %xmm4
+; SLM64-NEXT:    movdqa %xmm1, %xmm0
+; SLM64-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
 ; SLM64-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
 ; SLM64-NEXT:    movdqa %xmm3, %xmm2
-; SLM64-NEXT:    punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; SLM64-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; SLM64-NEXT:    movdqa %xmm4, %xmm1
+; SLM64-NEXT:    punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7]
+; SLM64-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
 ; SLM64-NEXT:    retq
 ;
 ; SLOW32-LABEL: test_mul_v16i32_v16i16:

diff --git a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse2.s b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse2.s
index 8c5fff1..dfdfa13 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse2.s
+++ b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse2.s
@@ -563,8 +563,8 @@ xorpd       (%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        pextrw $1, %xmm0, %ecx
 # CHECK-NEXT:  1      1     1.00                        pinsrw $1, %eax, %xmm0
 # CHECK-NEXT:  1      4     1.00    *                   pinsrw $1, (%rax), %xmm0
-# CHECK-NEXT:  2      5     2.00                        pmaddwd        %xmm0, %xmm2
-# CHECK-NEXT:  2      8     2.00    *                   pmaddwd        (%rax), %xmm2
+# CHECK-NEXT:  1      5     2.00                        pmaddwd        %xmm0, %xmm2
+# CHECK-NEXT:  1      8     2.00    *                   pmaddwd        (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        pmaxsw %xmm0, %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   pmaxsw (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        pmaxub %xmm0, %xmm2
@@ -574,16 +574,16 @@ xorpd       (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        pminub %xmm0, %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   pminub (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        pmovmskb       %xmm0, %ecx
-# CHECK-NEXT:  2      5     2.00                        pmulhuw        %xmm0, %xmm2
-# CHECK-NEXT:  2      8     2.00    *                   pmulhuw        (%rax), %xmm2
-# CHECK-NEXT:  2      5     2.00                        pmulhw %xmm0, %xmm2
-# CHECK-NEXT:  2      8     2.00    *                   pmulhw (%rax), %xmm2
-# CHECK-NEXT:  2      5     2.00                        pmullw %xmm0, %xmm2
-# CHECK-NEXT:  2      8     2.00    *                   pmullw (%rax), %xmm2
+# CHECK-NEXT:  1      5     2.00                        pmulhuw        %xmm0, %xmm2
+# CHECK-NEXT:  1      8     2.00    *                   pmulhuw        (%rax), %xmm2
+# CHECK-NEXT:  1      5     2.00                        pmulhw %xmm0, %xmm2
+# CHECK-NEXT:  1      8     2.00    *                   pmulhw (%rax), %xmm2
+# CHECK-NEXT:  1      5     2.00                        pmullw %xmm0, %xmm2
+# CHECK-NEXT:  1      8     2.00    *                   pmullw (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        pmuludq        %mm0, %mm2
 # CHECK-NEXT:  1      7     1.00    *                   pmuludq        (%rax), %mm2
-# CHECK-NEXT:  2      5     2.00                        pmuludq        %xmm0, %xmm2
-# CHECK-NEXT:  2      8     2.00    *                   pmuludq        (%rax), %xmm2
+# CHECK-NEXT:  1      5     2.00                        pmuludq        %xmm0, %xmm2
+# CHECK-NEXT:  1      8     2.00    *                   pmuludq        (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        por    %xmm0, %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   por    (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        psadbw %xmm0, %xmm2

diff --git a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse41.s b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse41.s
index 2bcebea..e0e19e6 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/SLM/resources-sse41.s
+++ b/llvm/test/tools/llvm-mca/X86/SLM/resources-sse41.s
@@ -237,8 +237,8 @@ roundss     $1, (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   pmovzxwd       (%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        pmovzxwq       %xmm0, %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   pmovzxwq       (%rax), %xmm2
-# CHECK-NEXT:  2      5     2.00                        pmuldq %xmm0, %xmm2
-# CHECK-NEXT:  2      8     2.00    *                   pmuldq (%rax), %xmm2
+# CHECK-NEXT:  1      5     2.00                        pmuldq %xmm0, %xmm2
+# CHECK-NEXT:  1      8     2.00    *                   pmuldq (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        pmulld %xmm0, %xmm2
 # CHECK-NEXT:  1      7     1.00    *                   pmulld (%rax), %xmm2
 # CHECK-NEXT:  1      1     0.50                        ptest  %xmm0, %xmm1

diff --git a/llvm/test/tools/llvm-mca/X86/SLM/resources-ssse3.s b/llvm/test/tools/llvm-mca/X86/SLM/resources-ssse3.s
index f6c1bfe..3fb4878 100644 (file)
--- a/llvm/test/tools/llvm-mca/X86/SLM/resources-ssse3.s
+++ b/llvm/test/tools/llvm-mca/X86/SLM/resources-ssse3.s
@@ -148,12 +148,12 @@ psignw      (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00    *                   phsubw (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        pmaddubsw      %mm0, %mm2
 # CHECK-NEXT:  1      7     1.00    *                   pmaddubsw      (%rax), %mm2
-# CHECK-NEXT:  2      5     2.00                        pmaddubsw      %xmm0, %xmm2
-# CHECK-NEXT:  2      8     2.00    *                   pmaddubsw      (%rax), %xmm2
+# CHECK-NEXT:  1      5     2.00                        pmaddubsw      %xmm0, %xmm2
+# CHECK-NEXT:  1      8     2.00    *                   pmaddubsw      (%rax), %xmm2
 # CHECK-NEXT:  1      4     1.00                        pmulhrsw       %mm0, %mm2
 # CHECK-NEXT:  1      7     1.00    *                   pmulhrsw       (%rax), %mm2
-# CHECK-NEXT:  2      5     2.00                        pmulhrsw       %xmm0, %xmm2
-# CHECK-NEXT:  2      8     2.00    *                   pmulhrsw       (%rax), %xmm2
+# CHECK-NEXT:  1      5     2.00                        pmulhrsw       %xmm0, %xmm2
+# CHECK-NEXT:  1      8     2.00    *                   pmulhrsw       (%rax), %xmm2
 # CHECK-NEXT:  1      1     1.00                        pshufb %mm0, %mm2
 # CHECK-NEXT:  1      4     1.00    *                   pshufb (%rax), %mm2
 # CHECK-NEXT:  4      5     5.00                        pshufb %xmm0, %xmm2