[llvm] da965a7 - [X86][SLM] Fix MUL uops, latency and throughput
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Sat Sep 4 05:21:55 PDT 2021
Author: Simon Pilgrim
Date: 2021-09-04T13:21:34+01:00
New Revision: da965a77d566b9295a5928ca4c989650131bfc0b
URL: https://github.com/llvm/llvm-project/commit/da965a77d566b9295a5928ca4c989650131bfc0b
DIFF: https://github.com/llvm/llvm-project/commit/da965a77d566b9295a5928ca4c989650131bfc0b.diff
LOG: [X86][SLM] Fix MUL uops, latency and throughput
These were all set to the same best case mul i32 values (which seems to be the only version of MUL that SLM actually performs well with).
Noticed while trying to improve multiplication costs for vectorization via the D103695 helper script. Confirmed with Intel AoM / Agner / InstLatX64.
Added:
Modified:
llvm/lib/Target/X86/X86ScheduleSLM.td
llvm/test/tools/llvm-mca/X86/SLM/resources-x86_64.s
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ScheduleSLM.td b/llvm/lib/Target/X86/X86ScheduleSLM.td
index 2e58b1cb4249e..252f51228f4f0 100644
--- a/llvm/lib/Target/X86/X86ScheduleSLM.td
+++ b/llvm/lib/Target/X86/X86ScheduleSLM.td
@@ -101,16 +101,16 @@ def : InstRW<[WriteMove], (instrs COPY)>;
defm : SLMWriteResPair<WriteALU, [SLM_IEC_RSV01], 1>;
defm : SLMWriteResPair<WriteADC, [SLM_IEC_RSV01], 1>;
-defm : SLMWriteResPair<WriteIMul8, [SLM_IEC_RSV1], 3>;
-defm : SLMWriteResPair<WriteIMul16, [SLM_IEC_RSV1], 3>;
-defm : SLMWriteResPair<WriteIMul16Imm, [SLM_IEC_RSV1], 3>;
-defm : SLMWriteResPair<WriteIMul16Reg, [SLM_IEC_RSV1], 3>;
-defm : SLMWriteResPair<WriteIMul32, [SLM_IEC_RSV1], 3>;
+defm : SLMWriteResPair<WriteIMul8, [SLM_IEC_RSV1], 5, [5], 3>;
+defm : SLMWriteResPair<WriteIMul16, [SLM_IEC_RSV1], 5, [5], 4, 3, 1>;
+defm : SLMWriteResPair<WriteIMul16Imm, [SLM_IEC_RSV1], 4, [4], 2, 3, 1>;
+defm : SLMWriteResPair<WriteIMul16Reg, [SLM_IEC_RSV1], 4, [4], 2, 3, 1>;
+defm : SLMWriteResPair<WriteIMul32, [SLM_IEC_RSV1], 5, [5], 3, 3, 1>;
defm : SLMWriteResPair<WriteIMul32Imm, [SLM_IEC_RSV1], 3>;
defm : SLMWriteResPair<WriteIMul32Reg, [SLM_IEC_RSV1], 3>;
-defm : SLMWriteResPair<WriteIMul64, [SLM_IEC_RSV1], 3>;
-defm : SLMWriteResPair<WriteIMul64Imm, [SLM_IEC_RSV1], 3>;
-defm : SLMWriteResPair<WriteIMul64Reg, [SLM_IEC_RSV1], 3>;
+defm : SLMWriteResPair<WriteIMul64, [SLM_IEC_RSV1], 7, [7], 3>;
+defm : SLMWriteResPair<WriteIMul64Imm, [SLM_IEC_RSV1], 5, [2]>;
+defm : SLMWriteResPair<WriteIMul64Reg, [SLM_IEC_RSV1], 5, [2]>;
defm : X86WriteResUnsupported<WriteIMulH>;
defm : X86WriteResUnsupported<WriteIMulHLd>;
defm : X86WriteResPairUnsupported<WriteMULX32>;
diff --git a/llvm/test/tools/llvm-mca/X86/SLM/resources-x86_64.s b/llvm/test/tools/llvm-mca/X86/SLM/resources-x86_64.s
index 28b67638269a9..8a8ff766bcb4b 100644
--- a/llvm/test/tools/llvm-mca/X86/SLM/resources-x86_64.s
+++ b/llvm/test/tools/llvm-mca/X86/SLM/resources-x86_64.s
@@ -1334,32 +1334,32 @@ xorq (%rax), %rdi
# CHECK-NEXT: 1 29 25.00 * U idivl (%rax)
# CHECK-NEXT: 1 25 25.00 U idivq %rcx
# CHECK-NEXT: 1 29 25.00 * U idivq (%rax)
-# CHECK-NEXT: 1 3 1.00 imulb %dil
-# CHECK-NEXT: 1 6 1.00 * imulb (%rax)
-# CHECK-NEXT: 1 3 1.00 imulw %di
-# CHECK-NEXT: 1 6 1.00 * imulw (%rax)
-# CHECK-NEXT: 1 3 1.00 imulw %si, %di
-# CHECK-NEXT: 1 6 1.00 * imulw (%rax), %di
-# CHECK-NEXT: 1 3 1.00 imulw $511, %si, %di
-# CHECK-NEXT: 1 6 1.00 * imulw $511, (%rax), %di
-# CHECK-NEXT: 1 3 1.00 imulw $7, %si, %di
-# CHECK-NEXT: 1 6 1.00 * imulw $7, (%rax), %di
-# CHECK-NEXT: 1 3 1.00 imull %edi
-# CHECK-NEXT: 1 6 1.00 * imull (%rax)
+# CHECK-NEXT: 3 5 5.00 imulb %dil
+# CHECK-NEXT: 3 8 5.00 * imulb (%rax)
+# CHECK-NEXT: 4 5 5.00 imulw %di
+# CHECK-NEXT: 5 8 5.00 * imulw (%rax)
+# CHECK-NEXT: 2 4 4.00 imulw %si, %di
+# CHECK-NEXT: 3 7 4.00 * imulw (%rax), %di
+# CHECK-NEXT: 2 4 4.00 imulw $511, %si, %di
+# CHECK-NEXT: 3 7 4.00 * imulw $511, (%rax), %di
+# CHECK-NEXT: 2 4 4.00 imulw $7, %si, %di
+# CHECK-NEXT: 3 7 4.00 * imulw $7, (%rax), %di
+# CHECK-NEXT: 3 5 5.00 imull %edi
+# CHECK-NEXT: 4 8 5.00 * imull (%rax)
# CHECK-NEXT: 1 3 1.00 imull %esi, %edi
# CHECK-NEXT: 1 6 1.00 * imull (%rax), %edi
# CHECK-NEXT: 1 3 1.00 imull $665536, %esi, %edi
# CHECK-NEXT: 1 6 1.00 * imull $665536, (%rax), %edi
# CHECK-NEXT: 1 3 1.00 imull $7, %esi, %edi
# CHECK-NEXT: 1 6 1.00 * imull $7, (%rax), %edi
-# CHECK-NEXT: 1 3 1.00 imulq %rdi
-# CHECK-NEXT: 1 6 1.00 * imulq (%rax)
-# CHECK-NEXT: 1 3 1.00 imulq %rsi, %rdi
-# CHECK-NEXT: 1 6 1.00 * imulq (%rax), %rdi
-# CHECK-NEXT: 1 3 1.00 imulq $665536, %rsi, %rdi
-# CHECK-NEXT: 1 6 1.00 * imulq $665536, (%rax), %rdi
-# CHECK-NEXT: 1 3 1.00 imulq $7, %rsi, %rdi
-# CHECK-NEXT: 1 6 1.00 * imulq $7, (%rax), %rdi
+# CHECK-NEXT: 3 7 7.00 imulq %rdi
+# CHECK-NEXT: 3 10 7.00 * imulq (%rax)
+# CHECK-NEXT: 1 5 2.00 imulq %rsi, %rdi
+# CHECK-NEXT: 1 8 2.00 * imulq (%rax), %rdi
+# CHECK-NEXT: 1 5 2.00 imulq $665536, %rsi, %rdi
+# CHECK-NEXT: 1 8 2.00 * imulq $665536, (%rax), %rdi
+# CHECK-NEXT: 1 5 2.00 imulq $7, %rsi, %rdi
+# CHECK-NEXT: 1 8 2.00 * imulq $7, (%rax), %rdi
# CHECK-NEXT: 1 100 1.00 U inb $7, %al
# CHECK-NEXT: 1 100 1.00 U inb %dx, %al
# CHECK-NEXT: 1 100 1.00 U inw $7, %ax
@@ -1416,14 +1416,14 @@ xorq (%rax), %rdi
# CHECK-NEXT: 1 4 1.00 * movzwq (%rax), %rdi
# CHECK-NEXT: 1 1 0.50 movslq %eax, %rdi
# CHECK-NEXT: 1 4 1.00 * movslq (%rax), %rdi
-# CHECK-NEXT: 1 3 1.00 mulb %dil
-# CHECK-NEXT: 1 6 1.00 * mulb (%rax)
-# CHECK-NEXT: 1 3 1.00 mulw %si
-# CHECK-NEXT: 1 6 1.00 * mulw (%rax)
-# CHECK-NEXT: 1 3 1.00 mull %edx
-# CHECK-NEXT: 1 6 1.00 * mull (%rax)
-# CHECK-NEXT: 1 3 1.00 mulq %rcx
-# CHECK-NEXT: 1 6 1.00 * mulq (%rax)
+# CHECK-NEXT: 3 5 5.00 mulb %dil
+# CHECK-NEXT: 3 8 5.00 * mulb (%rax)
+# CHECK-NEXT: 4 5 5.00 mulw %si
+# CHECK-NEXT: 5 8 5.00 * mulw (%rax)
+# CHECK-NEXT: 3 5 5.00 mull %edx
+# CHECK-NEXT: 4 8 5.00 * mull (%rax)
+# CHECK-NEXT: 3 7 7.00 mulq %rcx
+# CHECK-NEXT: 3 10 7.00 * mulq (%rax)
# CHECK-NEXT: 1 1 0.50 negb %dil
# CHECK-NEXT: 2 5 2.00 * * negb (%r8)
# CHECK-NEXT: 2 5 2.00 * * lock negb (%r8)
@@ -1953,7 +1953,7 @@ xorq (%rax), %rdi
# CHECK: Resource pressure per iteration:
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7]
-# CHECK-NEXT: 400.00 - - 49.00 - 627.00 469.00 835.00
+# CHECK-NEXT: 400.00 - - 49.00 - 627.00 565.00 835.00
# CHECK: Resource pressure by instruction:
# CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] Instructions:
@@ -2261,32 +2261,32 @@ xorq (%rax), %rdi
# CHECK-NEXT: 25.00 - - - - 0.50 0.50 1.00 idivl (%rax)
# CHECK-NEXT: 25.00 - - - - 0.50 0.50 - idivq %rcx
# CHECK-NEXT: 25.00 - - - - 0.50 0.50 1.00 idivq (%rax)
-# CHECK-NEXT: - - - - - - 1.00 - imulb %dil
-# CHECK-NEXT: - - - - - - 1.00 1.00 imulb (%rax)
-# CHECK-NEXT: - - - - - - 1.00 - imulw %di
-# CHECK-NEXT: - - - - - - 1.00 1.00 imulw (%rax)
-# CHECK-NEXT: - - - - - - 1.00 - imulw %si, %di
-# CHECK-NEXT: - - - - - - 1.00 1.00 imulw (%rax), %di
-# CHECK-NEXT: - - - - - - 1.00 - imulw $511, %si, %di
-# CHECK-NEXT: - - - - - - 1.00 1.00 imulw $511, (%rax), %di
-# CHECK-NEXT: - - - - - - 1.00 - imulw $7, %si, %di
-# CHECK-NEXT: - - - - - - 1.00 1.00 imulw $7, (%rax), %di
-# CHECK-NEXT: - - - - - - 1.00 - imull %edi
-# CHECK-NEXT: - - - - - - 1.00 1.00 imull (%rax)
+# CHECK-NEXT: - - - - - - 5.00 - imulb %dil
+# CHECK-NEXT: - - - - - - 5.00 1.00 imulb (%rax)
+# CHECK-NEXT: - - - - - - 5.00 - imulw %di
+# CHECK-NEXT: - - - - - - 5.00 1.00 imulw (%rax)
+# CHECK-NEXT: - - - - - - 4.00 - imulw %si, %di
+# CHECK-NEXT: - - - - - - 4.00 1.00 imulw (%rax), %di
+# CHECK-NEXT: - - - - - - 4.00 - imulw $511, %si, %di
+# CHECK-NEXT: - - - - - - 4.00 1.00 imulw $511, (%rax), %di
+# CHECK-NEXT: - - - - - - 4.00 - imulw $7, %si, %di
+# CHECK-NEXT: - - - - - - 4.00 1.00 imulw $7, (%rax), %di
+# CHECK-NEXT: - - - - - - 5.00 - imull %edi
+# CHECK-NEXT: - - - - - - 5.00 1.00 imull (%rax)
# CHECK-NEXT: - - - - - - 1.00 - imull %esi, %edi
# CHECK-NEXT: - - - - - - 1.00 1.00 imull (%rax), %edi
# CHECK-NEXT: - - - - - - 1.00 - imull $665536, %esi, %edi
# CHECK-NEXT: - - - - - - 1.00 1.00 imull $665536, (%rax), %edi
# CHECK-NEXT: - - - - - - 1.00 - imull $7, %esi, %edi
# CHECK-NEXT: - - - - - - 1.00 1.00 imull $7, (%rax), %edi
-# CHECK-NEXT: - - - - - - 1.00 - imulq %rdi
-# CHECK-NEXT: - - - - - - 1.00 1.00 imulq (%rax)
-# CHECK-NEXT: - - - - - - 1.00 - imulq %rsi, %rdi
-# CHECK-NEXT: - - - - - - 1.00 1.00 imulq (%rax), %rdi
-# CHECK-NEXT: - - - - - - 1.00 - imulq $665536, %rsi, %rdi
-# CHECK-NEXT: - - - - - - 1.00 1.00 imulq $665536, (%rax), %rdi
-# CHECK-NEXT: - - - - - - 1.00 - imulq $7, %rsi, %rdi
-# CHECK-NEXT: - - - - - - 1.00 1.00 imulq $7, (%rax), %rdi
+# CHECK-NEXT: - - - - - - 7.00 - imulq %rdi
+# CHECK-NEXT: - - - - - - 7.00 1.00 imulq (%rax)
+# CHECK-NEXT: - - - - - - 2.00 - imulq %rsi, %rdi
+# CHECK-NEXT: - - - - - - 2.00 1.00 imulq (%rax), %rdi
+# CHECK-NEXT: - - - - - - 2.00 - imulq $665536, %rsi, %rdi
+# CHECK-NEXT: - - - - - - 2.00 1.00 imulq $665536, (%rax), %rdi
+# CHECK-NEXT: - - - - - - 2.00 - imulq $7, %rsi, %rdi
+# CHECK-NEXT: - - - - - - 2.00 1.00 imulq $7, (%rax), %rdi
# CHECK-NEXT: - - - 1.00 - - - - inb $7, %al
# CHECK-NEXT: - - - 1.00 - - - - inb %dx, %al
# CHECK-NEXT: - - - 1.00 - - - - inw $7, %ax
@@ -2343,14 +2343,14 @@ xorq (%rax), %rdi
# CHECK-NEXT: - - - - - 0.50 0.50 1.00 movzwq (%rax), %rdi
# CHECK-NEXT: - - - - - 0.50 0.50 - movslq %eax, %rdi
# CHECK-NEXT: - - - - - 0.50 0.50 1.00 movslq (%rax), %rdi
-# CHECK-NEXT: - - - - - - 1.00 - mulb %dil
-# CHECK-NEXT: - - - - - - 1.00 1.00 mulb (%rax)
-# CHECK-NEXT: - - - - - - 1.00 - mulw %si
-# CHECK-NEXT: - - - - - - 1.00 1.00 mulw (%rax)
-# CHECK-NEXT: - - - - - - 1.00 - mull %edx
-# CHECK-NEXT: - - - - - - 1.00 1.00 mull (%rax)
-# CHECK-NEXT: - - - - - - 1.00 - mulq %rcx
-# CHECK-NEXT: - - - - - - 1.00 1.00 mulq (%rax)
+# CHECK-NEXT: - - - - - - 5.00 - mulb %dil
+# CHECK-NEXT: - - - - - - 5.00 1.00 mulb (%rax)
+# CHECK-NEXT: - - - - - - 5.00 - mulw %si
+# CHECK-NEXT: - - - - - - 5.00 1.00 mulw (%rax)
+# CHECK-NEXT: - - - - - - 5.00 - mull %edx
+# CHECK-NEXT: - - - - - - 5.00 1.00 mull (%rax)
+# CHECK-NEXT: - - - - - - 7.00 - mulq %rcx
+# CHECK-NEXT: - - - - - - 7.00 1.00 mulq (%rax)
# CHECK-NEXT: - - - - - 0.50 0.50 - negb %dil
# CHECK-NEXT: - - - - - 0.50 0.50 2.00 negb (%r8)
# CHECK-NEXT: - - - - - 0.50 0.50 2.00 lock negb (%r8)
More information about the llvm-commits
mailing list