[AArch64] Missed optimization: Make full use of smaddl instruction in AArch64

[hstk30-hw]

https://godbolt.org/z/x6z11hPj7

Below code(from OpenSSL) in GCC can make full use of smaddl instruction, in LLVM it generate madd in llvm 15, and the performance is dropped significantly.

#include<stdint.h>

typedef int32_t fe[10];
static const int64_t kBottom21Bits = 0x1fffffLL;
static const int64_t kBottom25Bits = 0x1ffffffLL;
static const int64_t kBottom26Bits = 0x3ffffffLL;
static const int64_t kTop39Bits = 0xfffffffffe000000LL;
static const int64_t kTop38Bits = 0xfffffffffc000000LL;
 void fe_mul(fe h, const fe f, const fe g)
{
    int32_t f0 = f[0];
    int32_t f1 = f[1];
    int32_t f2 = f[2];
    int32_t f3 = f[3];
    int32_t f4 = f[4];
    int32_t f5 = f[5];
    int32_t f6 = f[6];
    int32_t f7 = f[7];
    int32_t f8 = f[8];
    int32_t f9 = f[9];
    int32_t g0 = g[0];
    int32_t g1 = g[1];
    int32_t g2 = g[2];
    int32_t g3 = g[3];
    int32_t g4 = g[4];
    int32_t g5 = g[5];
    int32_t g6 = g[6];
    int32_t g7 = g[7];
    int32_t g8 = g[8];
    int32_t g9 = g[9];
    int32_t g1_19 = 19 * g1;
    int32_t g2_19 = 19 * g2;
    int32_t g3_19 = 19 * g3;
    int32_t g4_19 = 19 * g4;
    int32_t g5_19 = 19 * g5;
    int32_t g6_19 = 19 * g6;
    int32_t g7_19 = 19 * g7;
    int32_t g8_19 = 19 * g8;
    int32_t g9_19 = 19 * g9;
    int32_t f1_2 = 2 * f1;
    int32_t f3_2 = 2 * f3;
    int32_t f5_2 = 2 * f5;
    int32_t f7_2 = 2 * f7;
    int32_t f9_2 = 2 * f9;
    int64_t f0g0 = f0 * (int64_t) g0;
    int64_t f0g1 = f0 * (int64_t) g1;
    int64_t f0g2 = f0 * (int64_t) g2;
    int64_t f0g3 = f0 * (int64_t) g3;
    int64_t f0g4 = f0 * (int64_t) g4;
    int64_t f0g5 = f0 * (int64_t) g5;
    int64_t f0g6 = f0 * (int64_t) g6;
    int64_t f0g7 = f0 * (int64_t) g7;
    int64_t f0g8 = f0 * (int64_t) g8;
    int64_t f0g9 = f0 * (int64_t) g9;
    int64_t f1g0 = f1 * (int64_t) g0;
    int64_t f1g1_2 = f1_2 * (int64_t) g1;
    int64_t f1g2 = f1 * (int64_t) g2;
    int64_t f1g3_2 = f1_2 * (int64_t) g3;
    int64_t f1g4 = f1 * (int64_t) g4;
    int64_t f1g5_2 = f1_2 * (int64_t) g5;
    int64_t f1g6 = f1 * (int64_t) g6;
    int64_t f1g7_2 = f1_2 * (int64_t) g7;
    int64_t f1g8 = f1 * (int64_t) g8;
    int64_t f1g9_38 = f1_2 * (int64_t) g9_19;
    int64_t f2g0 = f2 * (int64_t) g0;
    int64_t f2g1 = f2 * (int64_t) g1;
    int64_t f2g2 = f2 * (int64_t) g2;
    int64_t f2g3 = f2 * (int64_t) g3;
    int64_t f2g4 = f2 * (int64_t) g4;
    int64_t f2g5 = f2 * (int64_t) g5;
    int64_t f2g6 = f2 * (int64_t) g6;
    int64_t f2g7 = f2 * (int64_t) g7;
    int64_t f2g8_19 = f2 * (int64_t) g8_19;
    int64_t f2g9_19 = f2 * (int64_t) g9_19;
    int64_t f3g0 = f3 * (int64_t) g0;
    int64_t f3g1_2 = f3_2 * (int64_t) g1;
    int64_t f3g2 = f3 * (int64_t) g2;
    int64_t f3g3_2 = f3_2 * (int64_t) g3;
    int64_t f3g4 = f3 * (int64_t) g4;
    int64_t f3g5_2 = f3_2 * (int64_t) g5;
    int64_t f3g6 = f3 * (int64_t) g6;
    int64_t f3g7_38 = f3_2 * (int64_t) g7_19;
    int64_t f3g8_19 = f3 * (int64_t) g8_19;
    int64_t f3g9_38 = f3_2 * (int64_t) g9_19;
    int64_t f4g0 = f4 * (int64_t) g0;
    int64_t f4g1 = f4 * (int64_t) g1;
    int64_t f4g2 = f4 * (int64_t) g2;
    int64_t f4g3 = f4 * (int64_t) g3;
    int64_t f4g4 = f4 * (int64_t) g4;
    int64_t f4g5 = f4 * (int64_t) g5;
    int64_t f4g6_19 = f4 * (int64_t) g6_19;
    int64_t f4g7_19 = f4 * (int64_t) g7_19;
    int64_t f4g8_19 = f4 * (int64_t) g8_19;
    int64_t f4g9_19 = f4 * (int64_t) g9_19;
    int64_t f5g0 = f5 * (int64_t) g0;
    int64_t f5g1_2 = f5_2 * (int64_t) g1;
    int64_t f5g2 = f5 * (int64_t) g2;
    int64_t f5g3_2 = f5_2 * (int64_t) g3;
    int64_t f5g4 = f5 * (int64_t) g4;
    int64_t f5g5_38 = f5_2 * (int64_t) g5_19;
    int64_t f5g6_19 = f5 * (int64_t) g6_19;
    int64_t f5g7_38 = f5_2 * (int64_t) g7_19;
    int64_t f5g8_19 = f5 * (int64_t) g8_19;
    int64_t f5g9_38 = f5_2 * (int64_t) g9_19;
    int64_t f6g0 = f6 * (int64_t) g0;
    int64_t f6g1 = f6 * (int64_t) g1;
    int64_t f6g2 = f6 * (int64_t) g2;
    int64_t f6g3 = f6 * (int64_t) g3;
    int64_t f6g4_19 = f6 * (int64_t) g4_19;
    int64_t f6g5_19 = f6 * (int64_t) g5_19;
    int64_t f6g6_19 = f6 * (int64_t) g6_19;
    int64_t f6g7_19 = f6 * (int64_t) g7_19;
    int64_t f6g8_19 = f6 * (int64_t) g8_19;
    int64_t f6g9_19 = f6 * (int64_t) g9_19;
    int64_t f7g0 = f7 * (int64_t) g0;
    int64_t f7g1_2 = f7_2 * (int64_t) g1;
    int64_t f7g2 = f7 * (int64_t) g2;
    int64_t f7g3_38 = f7_2 * (int64_t) g3_19;
    int64_t f7g4_19 = f7 * (int64_t) g4_19;
    int64_t f7g5_38 = f7_2 * (int64_t) g5_19;
    int64_t f7g6_19 = f7 * (int64_t) g6_19;
    int64_t f7g7_38 = f7_2 * (int64_t) g7_19;
    int64_t f7g8_19 = f7 * (int64_t) g8_19;
    int64_t f7g9_38 = f7_2 * (int64_t) g9_19;
    int64_t f8g0 = f8 * (int64_t) g0;
    int64_t f8g1 = f8 * (int64_t) g1;
    int64_t f8g2_19 = f8 * (int64_t) g2_19;
    int64_t f8g3_19 = f8 * (int64_t) g3_19;
    int64_t f8g4_19 = f8 * (int64_t) g4_19;
    int64_t f8g5_19 = f8 * (int64_t) g5_19;
    int64_t f8g6_19 = f8 * (int64_t) g6_19;
    int64_t f8g7_19 = f8 * (int64_t) g7_19;
    int64_t f8g8_19 = f8 * (int64_t) g8_19;
    int64_t f8g9_19 = f8 * (int64_t) g9_19;
    int64_t f9g0 = f9 * (int64_t) g0;
    int64_t f9g1_38 = f9_2 * (int64_t) g1_19;
    int64_t f9g2_19 = f9 * (int64_t) g2_19;
    int64_t f9g3_38 = f9_2 * (int64_t) g3_19;
    int64_t f9g4_19 = f9 * (int64_t) g4_19;
    int64_t f9g5_38 = f9_2 * (int64_t) g5_19;
    int64_t f9g6_19 = f9 * (int64_t) g6_19;
    int64_t f9g7_38 = f9_2 * (int64_t) g7_19;
    int64_t f9g8_19 = f9 * (int64_t) g8_19;
    int64_t f9g9_38 = f9_2 * (int64_t) g9_19;
    int64_t h0 = f0g0 + f1g9_38 + f2g8_19 + f3g7_38 + f4g6_19 + f5g5_38 + f6g4_19 + f7g3_38 + f8g2_19 + f9g1_38;
    int64_t h1 = f0g1 + f1g0 + f2g9_19 + f3g8_19 + f4g7_19 + f5g6_19 + f6g5_19 + f7g4_19 + f8g3_19 + f9g2_19;
    int64_t h2 = f0g2 + f1g1_2 + f2g0 + f3g9_38 + f4g8_19 + f5g7_38 + f6g6_19 + f7g5_38 + f8g4_19 + f9g3_38;
    int64_t h3 = f0g3 + f1g2 + f2g1 + f3g0 + f4g9_19 + f5g8_19 + f6g7_19 + f7g6_19 + f8g5_19 + f9g4_19;
    int64_t h4 = f0g4 + f1g3_2 + f2g2 + f3g1_2 + f4g0 + f5g9_38 + f6g8_19 + f7g7_38 + f8g6_19 + f9g5_38;
    int64_t h5 = f0g5 + f1g4 + f2g3 + f3g2 + f4g1 + f5g0 + f6g9_19 + f7g8_19 + f8g7_19 + f9g6_19;
    int64_t h6 = f0g6 + f1g5_2 + f2g4 + f3g3_2 + f4g2 + f5g1_2 + f6g0 + f7g9_38 + f8g8_19 + f9g7_38;
    int64_t h7 = f0g7 + f1g6 + f2g5 + f3g4 + f4g3 + f5g2 + f6g1 + f7g0 + f8g9_19 + f9g8_19;
    int64_t h8 = f0g8 + f1g7_2 + f2g6 + f3g5_2 + f4g4 + f5g3_2 + f6g2 + f7g1_2 + f8g0 + f9g9_38;
    int64_t h9 = f0g9 + f1g8 + f2g7 + f3g6 + f4g5 + f5g4 + f6g3 + f7g2 + f8g1 + f9g0 ;
    int64_t carry0;
    int64_t carry1;
    int64_t carry2;
    int64_t carry3;
    int64_t carry4;
    int64_t carry5;
    int64_t carry6;
    int64_t carry7;
    int64_t carry8;
    int64_t carry9;
    carry0 = h0 + (1 << 25); h1 += carry0 >> 26; h0 -= carry0 & kTop38Bits;
    carry4 = h4 + (1 << 25); h5 += carry4 >> 26; h4 -= carry4 & kTop38Bits;
    carry1 = h1 + (1 << 24); h2 += carry1 >> 25; h1 -= carry1 & kTop39Bits;
    carry5 = h5 + (1 << 24); h6 += carry5 >> 25; h5 -= carry5 & kTop39Bits;

    carry2 = h2 + (1 << 25); h3 += carry2 >> 26; h2 -= carry2 & kTop38Bits;
    carry6 = h6 + (1 << 25); h7 += carry6 >> 26; h6 -= carry6 & kTop38Bits;

    carry3 = h3 + (1 << 24); h4 += carry3 >> 25; h3 -= carry3 & kTop39Bits;
    carry7 = h7 + (1 << 24); h8 += carry7 >> 25; h7 -= carry7 & kTop39Bits;
    carry4 = h4 + (1 << 25); h5 += carry4 >> 26; h4 -= carry4 & kTop38Bits;
    carry8 = h8 + (1 << 25); h9 += carry8 >> 26; h8 -= carry8 & kTop38Bits;
    carry9 = h9 + (1 << 24); h0 += (carry9 >> 25) * 19; h9 -= carry9 & kTop39Bits;
    carry0 = h0 + (1 << 25); h1 += carry0 >> 26; h0 -= carry0 & kTop38Bits;
    h[0] = (int32_t)h0;
    h[1] = (int32_t)h1;
    h[2] = (int32_t)h2;
    h[3] = (int32_t)h3;
    h[4] = (int32_t)h4;
    h[5] = (int32_t)h5;
    h[6] = (int32_t)h6;
    h[7] = (int32_t)h7;
    h[8] = (int32_t)h8;
    h[9] = (int32_t)h9;
}

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: None (hstk30-hw)

https://godbolt.org/z/x6z11hPj7

Below code(from OpenSSL) in GCC can make full use of smaddl instruction, in LLVM it generate madd in llvm 15, and the performance is dropped significantly.

#include<stdint.h>

typedef int32_t fe[10];
static const int64_t kBottom21Bits = 0x1fffffLL;
static const int64_t kBottom25Bits = 0x1ffffffLL;
static const int64_t kBottom26Bits = 0x3ffffffLL;
static const int64_t kTop39Bits = 0xfffffffffe000000LL;
static const int64_t kTop38Bits = 0xfffffffffc000000LL;
 void fe_mul(fe h, const fe f, const fe g)
{
    int32_t f0 = f[0];
    int32_t f1 = f[1];
    int32_t f2 = f[2];
    int32_t f3 = f[3];
    int32_t f4 = f[4];
    int32_t f5 = f[5];
    int32_t f6 = f[6];
    int32_t f7 = f[7];
    int32_t f8 = f[8];
    int32_t f9 = f[9];
    int32_t g0 = g[0];
    int32_t g1 = g[1];
    int32_t g2 = g[2];
    int32_t g3 = g[3];
    int32_t g4 = g[4];
    int32_t g5 = g[5];
    int32_t g6 = g[6];
    int32_t g7 = g[7];
    int32_t g8 = g[8];
    int32_t g9 = g[9];
    int32_t g1_19 = 19 * g1;
    int32_t g2_19 = 19 * g2;
    int32_t g3_19 = 19 * g3;
    int32_t g4_19 = 19 * g4;
    int32_t g5_19 = 19 * g5;
    int32_t g6_19 = 19 * g6;
    int32_t g7_19 = 19 * g7;
    int32_t g8_19 = 19 * g8;
    int32_t g9_19 = 19 * g9;
    int32_t f1_2 = 2 * f1;
    int32_t f3_2 = 2 * f3;
    int32_t f5_2 = 2 * f5;
    int32_t f7_2 = 2 * f7;
    int32_t f9_2 = 2 * f9;
    int64_t f0g0 = f0 * (int64_t) g0;
    int64_t f0g1 = f0 * (int64_t) g1;
    int64_t f0g2 = f0 * (int64_t) g2;
    int64_t f0g3 = f0 * (int64_t) g3;
    int64_t f0g4 = f0 * (int64_t) g4;
    int64_t f0g5 = f0 * (int64_t) g5;
    int64_t f0g6 = f0 * (int64_t) g6;
    int64_t f0g7 = f0 * (int64_t) g7;
    int64_t f0g8 = f0 * (int64_t) g8;
    int64_t f0g9 = f0 * (int64_t) g9;
    int64_t f1g0 = f1 * (int64_t) g0;
    int64_t f1g1_2 = f1_2 * (int64_t) g1;
    int64_t f1g2 = f1 * (int64_t) g2;
    int64_t f1g3_2 = f1_2 * (int64_t) g3;
    int64_t f1g4 = f1 * (int64_t) g4;
    int64_t f1g5_2 = f1_2 * (int64_t) g5;
    int64_t f1g6 = f1 * (int64_t) g6;
    int64_t f1g7_2 = f1_2 * (int64_t) g7;
    int64_t f1g8 = f1 * (int64_t) g8;
    int64_t f1g9_38 = f1_2 * (int64_t) g9_19;
    int64_t f2g0 = f2 * (int64_t) g0;
    int64_t f2g1 = f2 * (int64_t) g1;
    int64_t f2g2 = f2 * (int64_t) g2;
    int64_t f2g3 = f2 * (int64_t) g3;
    int64_t f2g4 = f2 * (int64_t) g4;
    int64_t f2g5 = f2 * (int64_t) g5;
    int64_t f2g6 = f2 * (int64_t) g6;
    int64_t f2g7 = f2 * (int64_t) g7;
    int64_t f2g8_19 = f2 * (int64_t) g8_19;
    int64_t f2g9_19 = f2 * (int64_t) g9_19;
    int64_t f3g0 = f3 * (int64_t) g0;
    int64_t f3g1_2 = f3_2 * (int64_t) g1;
    int64_t f3g2 = f3 * (int64_t) g2;
    int64_t f3g3_2 = f3_2 * (int64_t) g3;
    int64_t f3g4 = f3 * (int64_t) g4;
    int64_t f3g5_2 = f3_2 * (int64_t) g5;
    int64_t f3g6 = f3 * (int64_t) g6;
    int64_t f3g7_38 = f3_2 * (int64_t) g7_19;
    int64_t f3g8_19 = f3 * (int64_t) g8_19;
    int64_t f3g9_38 = f3_2 * (int64_t) g9_19;
    int64_t f4g0 = f4 * (int64_t) g0;
    int64_t f4g1 = f4 * (int64_t) g1;
    int64_t f4g2 = f4 * (int64_t) g2;
    int64_t f4g3 = f4 * (int64_t) g3;
    int64_t f4g4 = f4 * (int64_t) g4;
    int64_t f4g5 = f4 * (int64_t) g5;
    int64_t f4g6_19 = f4 * (int64_t) g6_19;
    int64_t f4g7_19 = f4 * (int64_t) g7_19;
    int64_t f4g8_19 = f4 * (int64_t) g8_19;
    int64_t f4g9_19 = f4 * (int64_t) g9_19;
    int64_t f5g0 = f5 * (int64_t) g0;
    int64_t f5g1_2 = f5_2 * (int64_t) g1;
    int64_t f5g2 = f5 * (int64_t) g2;
    int64_t f5g3_2 = f5_2 * (int64_t) g3;
    int64_t f5g4 = f5 * (int64_t) g4;
    int64_t f5g5_38 = f5_2 * (int64_t) g5_19;
    int64_t f5g6_19 = f5 * (int64_t) g6_19;
    int64_t f5g7_38 = f5_2 * (int64_t) g7_19;
    int64_t f5g8_19 = f5 * (int64_t) g8_19;
    int64_t f5g9_38 = f5_2 * (int64_t) g9_19;
    int64_t f6g0 = f6 * (int64_t) g0;
    int64_t f6g1 = f6 * (int64_t) g1;
    int64_t f6g2 = f6 * (int64_t) g2;
    int64_t f6g3 = f6 * (int64_t) g3;
    int64_t f6g4_19 = f6 * (int64_t) g4_19;
    int64_t f6g5_19 = f6 * (int64_t) g5_19;
    int64_t f6g6_19 = f6 * (int64_t) g6_19;
    int64_t f6g7_19 = f6 * (int64_t) g7_19;
    int64_t f6g8_19 = f6 * (int64_t) g8_19;
    int64_t f6g9_19 = f6 * (int64_t) g9_19;
    int64_t f7g0 = f7 * (int64_t) g0;
    int64_t f7g1_2 = f7_2 * (int64_t) g1;
    int64_t f7g2 = f7 * (int64_t) g2;
    int64_t f7g3_38 = f7_2 * (int64_t) g3_19;
    int64_t f7g4_19 = f7 * (int64_t) g4_19;
    int64_t f7g5_38 = f7_2 * (int64_t) g5_19;
    int64_t f7g6_19 = f7 * (int64_t) g6_19;
    int64_t f7g7_38 = f7_2 * (int64_t) g7_19;
    int64_t f7g8_19 = f7 * (int64_t) g8_19;
    int64_t f7g9_38 = f7_2 * (int64_t) g9_19;
    int64_t f8g0 = f8 * (int64_t) g0;
    int64_t f8g1 = f8 * (int64_t) g1;
    int64_t f8g2_19 = f8 * (int64_t) g2_19;
    int64_t f8g3_19 = f8 * (int64_t) g3_19;
    int64_t f8g4_19 = f8 * (int64_t) g4_19;
    int64_t f8g5_19 = f8 * (int64_t) g5_19;
    int64_t f8g6_19 = f8 * (int64_t) g6_19;
    int64_t f8g7_19 = f8 * (int64_t) g7_19;
    int64_t f8g8_19 = f8 * (int64_t) g8_19;
    int64_t f8g9_19 = f8 * (int64_t) g9_19;
    int64_t f9g0 = f9 * (int64_t) g0;
    int64_t f9g1_38 = f9_2 * (int64_t) g1_19;
    int64_t f9g2_19 = f9 * (int64_t) g2_19;
    int64_t f9g3_38 = f9_2 * (int64_t) g3_19;
    int64_t f9g4_19 = f9 * (int64_t) g4_19;
    int64_t f9g5_38 = f9_2 * (int64_t) g5_19;
    int64_t f9g6_19 = f9 * (int64_t) g6_19;
    int64_t f9g7_38 = f9_2 * (int64_t) g7_19;
    int64_t f9g8_19 = f9 * (int64_t) g8_19;
    int64_t f9g9_38 = f9_2 * (int64_t) g9_19;
    int64_t h0 = f0g0 + f1g9_38 + f2g8_19 + f3g7_38 + f4g6_19 + f5g5_38 + f6g4_19 + f7g3_38 + f8g2_19 + f9g1_38;
    int64_t h1 = f0g1 + f1g0 + f2g9_19 + f3g8_19 + f4g7_19 + f5g6_19 + f6g5_19 + f7g4_19 + f8g3_19 + f9g2_19;
    int64_t h2 = f0g2 + f1g1_2 + f2g0 + f3g9_38 + f4g8_19 + f5g7_38 + f6g6_19 + f7g5_38 + f8g4_19 + f9g3_38;
    int64_t h3 = f0g3 + f1g2 + f2g1 + f3g0 + f4g9_19 + f5g8_19 + f6g7_19 + f7g6_19 + f8g5_19 + f9g4_19;
    int64_t h4 = f0g4 + f1g3_2 + f2g2 + f3g1_2 + f4g0 + f5g9_38 + f6g8_19 + f7g7_38 + f8g6_19 + f9g5_38;
    int64_t h5 = f0g5 + f1g4 + f2g3 + f3g2 + f4g1 + f5g0 + f6g9_19 + f7g8_19 + f8g7_19 + f9g6_19;
    int64_t h6 = f0g6 + f1g5_2 + f2g4 + f3g3_2 + f4g2 + f5g1_2 + f6g0 + f7g9_38 + f8g8_19 + f9g7_38;
    int64_t h7 = f0g7 + f1g6 + f2g5 + f3g4 + f4g3 + f5g2 + f6g1 + f7g0 + f8g9_19 + f9g8_19;
    int64_t h8 = f0g8 + f1g7_2 + f2g6 + f3g5_2 + f4g4 + f5g3_2 + f6g2 + f7g1_2 + f8g0 + f9g9_38;
    int64_t h9 = f0g9 + f1g8 + f2g7 + f3g6 + f4g5 + f5g4 + f6g3 + f7g2 + f8g1 + f9g0 ;
    int64_t carry0;
    int64_t carry1;
    int64_t carry2;
    int64_t carry3;
    int64_t carry4;
    int64_t carry5;
    int64_t carry6;
    int64_t carry7;
    int64_t carry8;
    int64_t carry9;
    carry0 = h0 + (1 << 25); h1 += carry0 >> 26; h0 -= carry0 & kTop38Bits;
    carry4 = h4 + (1 << 25); h5 += carry4 >> 26; h4 -= carry4 & kTop38Bits;
    carry1 = h1 + (1 << 24); h2 += carry1 >> 25; h1 -= carry1 & kTop39Bits;
    carry5 = h5 + (1 << 24); h6 += carry5 >> 25; h5 -= carry5 & kTop39Bits;

    carry2 = h2 + (1 << 25); h3 += carry2 >> 26; h2 -= carry2 & kTop38Bits;
    carry6 = h6 + (1 << 25); h7 += carry6 >> 26; h6 -= carry6 & kTop38Bits;

    carry3 = h3 + (1 << 24); h4 += carry3 >> 25; h3 -= carry3 & kTop39Bits;
    carry7 = h7 + (1 << 24); h8 += carry7 >> 25; h7 -= carry7 & kTop39Bits;
    carry4 = h4 + (1 << 25); h5 += carry4 >> 26; h4 -= carry4 & kTop38Bits;
    carry8 = h8 + (1 << 25); h9 += carry8 >> 26; h8 -= carry8 & kTop38Bits;
    carry9 = h9 + (1 << 24); h0 += (carry9 >> 25) * 19; h9 -= carry9 & kTop39Bits;
    carry0 = h0 + (1 << 25); h1 += carry0 >> 26; h0 -= carry0 & kTop38Bits;
    h[0] = (int32_t)h0;
    h[1] = (int32_t)h1;
    h[2] = (int32_t)h2;
    h[3] = (int32_t)h3;
    h[4] = (int32_t)h4;
    h[5] = (int32_t)h5;
    h[6] = (int32_t)h6;
    h[7] = (int32_t)h7;
    h[8] = (int32_t)h8;
    h[9] = (int32_t)h9;
}

[MacDue]

LLVM 15 is not supported anymore. It looks like since clang 16 LLVM makes use of smaddl: https://godbolt.org/z/onW1sE8vG

[davemgreen]

It looks like CGP is messing up some of the cases by promoting the sext to the load. -mllvm -disable-cgp-ext-ld-promotion can help generate more, but the code is still quite a bit larger than optimal due to spills and reloads. https://godbolt.org/z/4MrWWKMYh. It should be possible to get CGP to stop doing that if the mull will be more efficient.

Some CPUs nowadays have more larger multiply units, meaning the difference between smull and mul is smaller.

[c-rhodes]

LLVM 15 is not supported anymore. It looks like since clang 16 LLVM makes use of smaddl: https://godbolt.org/z/onW1sE8vG

Although LLVM is still emitting some madds whereas GCC isn't. I wonder if #135926 might help here, although it seems to have stalled.

[davemgreen]

Although LLVM is still emitting some madds whereas GCC isn't. I wonder if #135926 might help here, although it seems to have stalled.

SDAG should be able to convert to smull providing that the known-bits can be detected, and this case was a single large BB. I believe the transform in CGP looses information about them though (I may be wrong, I didn't trace what it was doing very thoroughly), which is difficult/impossible to undo.